Compare commits

...
124 Commits
Author SHA1 Message Date
minke.yu c8d06424fa gateway metrics: T25b test fix — le buckets are cumulative 2026-10-18 09:53:17 +08:00
minke.yu 97c7eabb33 gateway metrics: T25b fixup — recorder.handle() for 0.17 API 2026-10-18 09:51:58 +08:00
minke.yu e9b436dc2d gateway metrics: explicit histogram buckets for all g9i/router histograms — kill zero-quantile summaries (P2 T25b) 2026-10-18 09:50:23 +08:00
minke.yu fc63ae5392 gateway g9i: T25 test — relax timing-sensitive wait assertion 2026-10-18 08:34:56 +08:00
minke.yu 9df2d40324 gateway g9i: T25 cleanup — drop dead wrapper fns, cfg(test) the test-only ones 2026-10-18 08:31:31 +08:00
minke.yu 6a8a6e399f gateway g9i: T25 fixup — shadow-mode reservoir eligibility is unreachable, not excluded 2026-10-18 08:27:15 +08:00
minke.yu e6c0737541 gateway g9i: T25 fixup — class ttft reads ShadowResponseContext.class 2026-10-18 08:23:53 +08:00
minke.yu 9374e03f13 gateway g9i: hot-channel admission A/B — reservoir + hardlane policies (P2 T25) 2026-10-18 08:21:21 +08:00
minke.yu 86883ae89f gateway g9i: deep placeholder rings anchor classification (P2 T24)
v14 replay (c512, 32836 reqs): T23 chunking works (lookup_depth mean 41,
placeholder 6795 vs engine-hit 7183) but Hot never fired — placeholder
anchors kept suffix_start=0, so every shared-prefix request counted its
whole message list as new work -> ColdLarge. Now a placeholder ring at
depth >= MIN_PLACEHOLDER_DEPTH (2) anchors like a real hit: suffix_start
via ring_msg, matched_node filled (record writes the node on every ring),
anchor_tokens stays 0 (unknown, deliberate). depth==1 placeholders stay
misses (shared-greeting guard). ShadowDecision gains
anchor_is_placeholder; the shadow x engine agreement counter's shadow
axis widens to hit/placeholder/miss via ShadowAnchorKind. KNOWN
LIMITATION pinned in a test: a chunk-ring anchor mid-message hides an
unbounded remainder (wrong-Hot vector for >theta mid-message divergence
after a >=2048B shared prefix). No deploy.
2026-10-17 23:57:44 +08:00
minke.yu 43c2844c18 gateway g9i: intra-message 2048B chunk rings for long-prompt prefix hits (P2 T23)
A per-message chain is blind inside one long message (v13: 9034/9034
depth-0 misses on single-message long-prompt traffic). Text pieces
longer than 2048 bytes now roll into bare chunk rings
xxh3(prev || chunk) before the message-final ring, which carries only
the <=2048-byte tail; texts <=2048 hash byte-identically to pre-T23
(pinned by a verbatim legacy-chain equivalence test). Lookup itself is
unchanged, but classify's suffix_start now maps the hit ring back to
its message index via the new ring_msg sideband (the raw n-depth+1
formula overshoots the message arrays once chains outgrow messages).
No deploy.
2026-10-17 22:33:01 +08:00
minke.yu 9e288079e4 gateway g9i: fix non-stream CL truncation on shadow strip + pre-strip breakdown extraction + shadowxengine agreement metric (T22)
- Bug A (user-visible): non-stream passthrough forwarded the upstream
  Content-Length after the shadow strip had shortened the body ->
  cache-hit non-stream responses truncated (~60B, IncompleteRead).
  non_stream_passthrough_response now pins CL to the actual body length.
- Bug B (order): strip_and_extract returned the post-strip JSON to
  shadow_record, so extract_cached_breakdown always missed and
  smg_g9i_shadow_cache_hit_tokens_total never fired. The breakdown is now
  extracted pre-strip and threaded through record_response_json.
- T22 observability: extract_cached_total (standard
  usage.prompt_tokens_details.cached_tokens first, tier breakdown as
  fallback) + smg_g9i_shadow_engine_agreement_total{shadow,engine} 2x2
  counter pairing the request-time lookup hit with the engine's actual
  cached tokens; ShadowResponseContext carries shadow_hit.
2026-10-17 15:11:58 +08:00
minke.yu 738d5a50d0 gateway g9i: shadow semantic gap tests (image-in-prefix/theta boundary/deep hit/shared-prefix repoint/try_consume/reasoning_content) 2026-10-17 10:15:38 +08:00
minke.yu e80c8608dc gateway g9i: --g9i-window-fixed pins the admission window, skipping all AIMD dynamics (P2 T21) 2026-10-17 01:24:23 +08:00
minke.yu 5351d5491b gateway g9i: AIMD warmup tuning (init 128, add 64, CLI flags) + prompt_too_long 400 front validation (P2 T20 A+B) 2026-10-16 20:07:49 +08:00
minke.yu 8a2e1403a3 gateway g9i: blind-state timeout no longer blocks recovery when pipeline < 0.5W (P2 T19 P1) 2026-10-16 15:59:14 +08:00
minke.yu 1910eccb8e gateway g9i: ledger reconcile clamp on sustained ledger>>engine divergence (subtract-only, 60s fuse) (P2 T19 P0-2) 2026-10-16 15:56:56 +08:00
minke.yu 6217b9de46 gateway g9i: TTL reaper for wedged D-ledger bookings (pipeline 120s / streaming 600s) (P2 T19 P0-1) 2026-10-16 15:51:40 +08:00
minke.yu 9dd446b5c1 gateway g9i: AIMD blind-state recovery threshold prealloc < max(1, max_running/4) (P2 T18 hole-4)
The blind-state (no TTFT samples — non-streaming traffic) slow-add
condition required D prealloc_queue == 0, which is unreachable in
production (prealloc sits at 180+ under load), so the window stayed
pinned at WINDOW_MIN=16 forever — the admission window never recovered
after a cut.

New rule: slow-add when prealloc_queue < max(1, max_running_total/4)
(pipeline < 0.8W condition kept; hold-timeout still forces hold). A
missing snapshot for the node is treated as busy (no blind recovery on
zero information).

Test: aimd_blind_holds_while_prealloc_busy replaced by
aimd_blind_recovery_prealloc_threshold (0/10 recover at limit 64,
boundary == limit holds, production-shape 180 holds, max(1, mr/4) floor
at mr=8, missing snapshot holds).
2026-10-16 13:42:48 +08:00
minke.yu f76eaf0f92 gateway g9i: gates judge on last-known snapshots; blind fallback round-robins instead of cache_aware pass-through (P2 T18 hole-3)
The stale-exemption + blind-pass-through combination fed the 2026-10-16
p3 pileup: the most overloaded node answered /v1/loads slowest, its
snapshot went stale first, and staleness EXEMPTED it from every cap
judgment; with all snapshots stale, regular_plan returned
PrefillPlan::Default and the underlying cache_aware radix policy pinned
every same-prefix session onto that one node (p3: 477 queued, TTFT
~300s) while the drained node (p2) got zero traffic.

a) v1_over_cap / p_real_queue_over / least_wait / sorted_by_wait now use
   last-known values: any snapshot that ever fetched OK participates at
   any age; only a NEVER-answered node is unknown (unjudged, unpicked).
   Stale judgments are counted: smg_g9i_stale_snapshot_judged_total{gate}.
b) All-blind (no node ever answered = true cold start) no longer returns
   PrefillPlan::Default — regular_plan round-robins the pool via a new
   rr_cursor (decision path regular_blind_round_robin). Partial blind
   picks the known least-wait node as before. Prefix affinity lives in
   the shadow table (shadow_route), not the underlying policy.
c) reqs_cap=10 / token_cap=16M / hold 1s/512 / HOT-first / stateless
   per-request exclusion semantics unchanged; check_decode_pool freshness
   (TTL) semantics intentionally untouched.

Tests: v1_over_cap_snapshot_semantics + shadow_path_real_queue_gate_filters
updated to the last-known contract; new t18_least_wait_uses_last_known,
t18_all_blind_round_robin_not_default, t18_partial_blind_least_wait_override.
2026-10-16 13:41:09 +08:00
minke.yu 45fdc97082 gateway g9i: concurrent /v1/loads fetch + detached bootstrap scrape loop + last-known snapshot survival (P2 T18 hole-2)
poll_once fetched 4x /v1/loads + 3x P /metrics SERIALLY, each with a 3s
timeout — one round cost up to ~21s against a 5s snapshot TTL, so one
slow node at peak aged the whole pool's snapshots and blinded every
gate (no_fresh_snapshot on ~30% of decisions, 2026-10-16).

- /v1/loads fetches now run concurrently (JoinSet); per-node 3s timeout
  unchanged, a round costs the slowest node, not the sum. Poll interval
  1s / TTL 5s unchanged.
- The per-P /metrics kv_transfer_bootstrap_ms scrape moves to its own
  10s slow loop (concurrent scrapes); it only feeds observability gauges
  and must never delay the snapshot round.
- A failed fetch no longer overwrites the snapshot with a zeroed
  default: the last-known load survives (merge_poll_result) and keeps
  aging via its original fetched_at; a node that never answered stays
  truly unknown instead of fabricating '0 queue'.
- New gauges: smg_g9i_poll_duration_seconds (histogram),
  smg_g9i_snapshot_age_seconds{worker}.

Tests: t18_poll_failure_preserves_last_known (merge semantics),
t18_poll_fetches_concurrently (two hanging nodes cost one timeout, stay
unknown but remain routing candidates).
2026-10-16 13:33:07 +08:00
minke.yu f690861265 gateway g9i: book D pipeline ledger at decision time for deferred (miss) admissions (P2 T18 hole-1)
Root cause of the 2026-10-16 window-gate-blindness incident: for a
prefix-table MISS the D pipeline booking happened in shadow_admit at
response-headers time — and in PD mode headers arrive at ~first-token
time (the dispatch join waits on the prefill answer), so the booking
entered the ledger and was promoted to streaming within milliseconds.
The pipeline gauge pinned at 0 while 585 admitted requests awaited
their first token; the d_window_full gate (pipeline >= W) never fired.

Fix: AdmissionMark gains p_booked + a new_deferred() constructor; the
miss branch of shadow_classify_and_admit books the min-T_D D node into
the pipeline ledger at decision time (P side stays deferred — no node
known). shadow_admit consumes the mark by admitting P on the actual
landing node and rekeying the D booking; shadow_release_unconsumed
releases D (and P only when it was booked). Promote/release paths are
unchanged and conserve: admit = pipeline + streaming, complete/abort
= release exactly once.

Also adds the reconciliation gauge smg_g9i_d_engine_inflight{worker}
(running + prealloc + transfer queue, engine truth) beside the ledger
gauges, so ledger-vs-engine divergence is directly visible.

Tests: t11_miss_defers_admit_to_response rewritten as
t11_miss_books_d_pipeline_at_decision_time; new
t18_deferred_promote_release_conservation (promote idempotent,
mid-stream disconnect release, post-consume fallback no-op) and
t18_deferred_abort_pre_first_token_releases_pipeline (double-release
guard).
2026-10-16 13:28:08 +08:00
minke.yu e4bfdc949b gateway g9i: AIMD blind-state recovery on real engine signals + real-queue P gate + nonstream e2e metric (P2 T17) 2026-10-16 11:16:32 +08:00
minke.yu 59e3aee6d9 gateway g9i: pipeline-depth window (streaming excluded) + tiered hold queue + 1s hold budget (P2 T16) 2026-10-16 09:43:54 +08:00
minke.yu 406081d3e3 gateway g9i: admission hold queue (429 -> park) + AIMD window rebalance to 1x/0.25x capacity (P2 T15) 2026-10-16 08:27:52 +08:00
minke.yu 1975265315 gateway g9i: AIMD congestion window replaces slot-estimate admission gate (P2 T14) 2026-10-16 01:33:18 +08:00
minke.yu 534417b82e gateway g9i: D-side in-flight ledger + Little's-law slot release rate (P2 T13) 2026-10-15 23:21:42 +08:00
minke.yu ed0a454730 gateway g9i: T_D slot (max_running) dimension — max(memory, slot) wait with dual EWMA (P2 T12) 2026-10-15 21:41:03 +08:00
minke.yu 04ee2d03d7 gateway g9i: atomic check-and-act admission at decision time (P2 T11) 2026-10-15 18:20:01 +08:00
minke.yu 9d5db13e53 gateway g9i: P-queue hard token cap with candidate fallback (P2 T10) 2026-10-15 17:38:17 +08:00
minke.yu 7e20f0af79 gateway g9i: P2 T_D admission gates (off|shadow|enforce) + T_PREFER candidate fallback (T9) 2026-10-15 13:35:50 +08:00
minke.yu 0b74ab2ac4 gateway g9i: shadow-consistent routing switch --g9i-shadow-route (P1 shadow T7) 2026-10-14 22:22:58 +08:00
minke.yu 3c7ec33d26 gateway g9i: est_T_D shadow metrics — est_tp/est_td histograms, residual by dominant component, pool gauges (P1 shadow T5) 2026-10-14 21:52:07 +08:00
minke.yu 09ca4957ee gateway g9i: scrape P /metrics kv_transfer_bootstrap_ms as est_T_D ground truth (P1 shadow T4) 2026-10-14 21:50:11 +08:00
minke.yu 7921442a5c gateway g9i: est_TTFT max(T_P, min_D T_D) composition with C4/C13 degradation (P1 shadow T3) 2026-10-14 21:44:17 +08:00
minke.yu 05ef104208 gateway g9i: pool_free_rate completion EWMA per decode node (P1 shadow T2) 2026-10-14 21:34:56 +08:00
minke.yu d44620f285 gateway g9i: fetch disagg/queues sections in /v1/loads, extend NodeLoad for est_T_D (P1 shadow T1) 2026-10-14 21:21:11 +08:00
minke.yu 5df4f2c036 DEBUG-ONLY: use crc32 for rid fingerprint (builtin hash() is per-process randomized) 2026-10-13 18:12:56 +08:00
minke.yu db1e8cdbe3 DEBUG-ONLY: ARDBG batch fingerprint + phase + hicache sync logs, sampled device AR (1/1024) + idle spin sampling (p2 wedge trap) 2026-10-13 18:01:36 +08:00
minke.yu f3ac513d92 DEBUG-ONLY: limit ARDBG stack logging to ONE_SHOT_PUSH calls (volume control) 2026-10-13 16:01:18 +08:00
minke.yu 0054c5b8fe DEBUG-ONLY: SGLANG_AR_SEQ_DEBUG seq logging for TP device AR + host CP gloo AR (p2 wedge repro) 2026-10-13 15:52:17 +08:00
minke.yu 6da0a40b02 g9i shadow: fix temporary-borrow compile errors in new tests 2026-10-13 15:39:06 +08:00
minke.yu caf3403086 g9i shadow: fix cached_tokens_details placement to sglext container
Verified against serving_chat.py/protocol.py on this branch: the engine
emits cached_tokens_details under top-level "sglext" (SglExt), omitted
entirely when all fields are None; ids-framed streams emit a standalone
choices=[] sglext chunk with no usage. extract_cached_breakdown now reads
sglext first (usage/top-level kept as fallbacks); strip_cached_details_from
removes the field from sglext and drops the emptied key; process_sse_chunk
pre-filters on "sglext" too and drops events that only carried stripped
fields. Adds 5 placement tests.
2026-10-13 15:33:03 +08:00
minke.yu 617f19ba61 gateway tests: fix PDRouter initializer missing g9i field (pre-existing test compile break) 2026-10-13 15:18:33 +08:00
minke.yu ba3318ffbd gateway g9i: fix shadow compile errors — lookup returns owned snapshot (borrow across loop mutation), bind temp in test 2026-10-13 15:14:05 +08:00
minke.yu 8279d3d6df gateway g9i: v2 P1 shadow mode — prefix-hash chain, bookkeeping table, shadow classifier + queue model
Observability only, zero routing influence, gated by --g9i-shadow-enabled
(requires --g9i-enabled). Design: experiments/g9i/g9i-v2-design-20261013.md.

- g9i_prefix_chain.rs: rolling xxh3 chain over chat messages (seed folds
  model/template-kwargs/tools/system; tool_calls/tool_call_id/name hashed;
  byte-level image semantics; unclassifiable flag for unknown parts)
- g9i_shadow.rs: PrefixTable (LRU+TTL, deepest-hit lookup, exact tail
  anchor + placeholder middle boundaries), ShadowClassifier (hot/short/
  cold_image/cold_large/cold_unknown via byte heuristic), QueueModel
  (per-(node,class) outstanding + completion-driven drain EWMA, est_TTFT),
  usage/cached_tokens_details injection+strip pairing
- pd_router.rs: request-side shadow classify + agreement counters (shadow
  vs v1 probe verdict), response-side usage harvest on stream/non-stream,
  queue admit/complete/release conservation incl. abort paths
- main.rs: --g9i-shadow-enabled/-table-max-entries/-table-ttl-secs

Not yet compile-verified (no local toolchain); b300 container check next.
2026-10-13 15:05:56 +08:00
minke.yu ef4cccb6a0 gateway g9i: allow flow-control-only mode (no fast-P URL) — B case flat pool 2026-10-13 01:31:57 +08:00
minke.yu 98b950366d mooncake PD transfer: mix bootstrap_room into lane shard hash (fixes 1/16 lane collapse with single decode instance) 2026-10-12 23:41:23 +08:00
minke.yu 87ab6c0155 gateway g9i: probe fast-P only + L3-inclusive classification + exact-body probe cache (A案; removes accidentally reintroduced home_hit) 2026-10-12 13:03:54 +00:00
minke.yu 0308920bdb gateway g9i: retry probe once on transport errors (stale keep-alive pool)
Observed twice on 2026-10-12: reqwest reused server-closed pooled connections,
all-candidate probes failed with connect-level errors, fail-closed 429 drowned
hit-class traffic. Retry once on is_connect/is_request (ms-scale reconnect);
timeouts stay non-retried (wedged engine + 6s worst-case TTFT budget).
2026-10-12 12:02:49 +00:00
minke.yu 57db2a4ae9 gateway g9i: fail-closed 429 on probe failure / fast-P snapshot loss (p2 incident)
User directive 2026-10-12: 先严格429. When p2 hung, the fail-open fallback dumped
all classified traffic into the regular pool and drowned it (7.5M pending tokens).
All probe-failure paths now reject: prefill_probe_failed (429) when no candidate
can answer /query_cache_hit; fast_p_no_snapshot (429) when snapshots go stale
between probe and pick, with boot grace (had_any_success) mirroring the D gate.
2026-10-12 10:16:09 +00:00
minke.yu 7fd7fd58a5 gateway g9i: hit class always routes to fast-P pool (命中都在 fast-P 发生)
User directive 2026-10-12: read-your-own-node home routing sent hit-class
requests back onto the regular-P long-prefill queue (150s p50 observed),
defeating the fast path. Hit class now always goes to the fast-P pool
(least-wait aggregate); fast-P resolves the prefix from its own L2 or the
shared L3 (wait_complete), recompute-and-graft otherwise — self-healing
land-and-stay. Home override removed; fast-P overload is handled only by
the fast_p_queue 429 gate. New decision counter routed_fast_p_from_regular
distinguishes L3-fetch/recompute hits from local-L2 hits (observability).
2026-10-12 02:20:12 +00:00
minke.yuandKimi 71a72b8fea gateway g9i: keep large prefills off fast-P even without least-wait
regular_plan returned Default when regular_pool_least_wait=false, letting
the router policy pick across ALL prefill workers including the fast-P
node — ~1/3 of large prefills landed on fast-P under random policy
(observed 2026-10-12), polluting its small-prefill queue and L2. With a
fast-P configured, 'regular' now explicitly restricts to the regular
pool via the least-wait aggregate pick. Default is kept only when the
fast pool is empty (then it equals the regular pool anyway).

Co-authored-by: Kimi
2026-10-12 08:58:03 +08:00
minke.yuandKimi 6c9518fcd8 gateway g9i: move ProbeHit to module scope (fix compile)
Co-authored-by: Kimi
2026-10-12 08:20:30 +08:00
minke.yuandKimi 3a4ab451ac gateway g9i: classify on local (L1+L2) hits only, probe the whole pool
Follow-up to the home-hit routing fix. Two flaws remained:

1. The engine's /query_cache_hit new_tokens counts shared-L3 hits. Mooncake
   pages are metadata-visible (Exist) before the put finishes, so fast-P
   claimed long prefixes it could not actually serve: the gateway classified
   'small prefill, fast-P has it' while the engine prefetch found nothing
   and recomputed 76k cold — and the probe loop broke at fast-P before ever
   asking the home regular-P, which held an immediate L2 copy.

2. The probe loop stopped at the first answering node, so 'which node holds
   the prefix' was only ever 'fast-P or whoever got asked first'.

Now every fresh candidate is probed (fast pool first, early-out only on a
fully-local prompt), and the routing decision uses local_new = total - l1 -
l2. Ties keep probe order so fresh short prompts with no hits still land on
fast-P.

Co-authored-by: Kimi
2026-10-12 08:17:52 +08:00
minke.yuandKimi 5d38fc96a9 gateway g9i: route small prefills to the node that actually holds the prefix
classify_prefill remembered only new_tokens from the /query_cache_hit probe
and sent every small-prefill request to the fast-P pool. A multi-turn session
whose home P is a regular node therefore landed on fast-P with a local miss:
either a full prefix recompute or a wait for the cross-node L3 backup to
become visible (seconds for long prefixes; observed 10-20s for 76k).

Now the probe records which worker answered. When the prefix lives on a
regular-P the request is Override-routed home (immediate L2 hit), subject to
the same regular_p_queue est-wait 429 gate. Fast-P keeps serving sessions
whose prefix it already holds (its own L2) and brand-new short prefills.

Co-authored-by: Kimi
2026-10-12 08:01:10 +08:00
minke.yu 261ebbc55b G9i: regular-P queue admission gate (est wait > 10s -> 429 regular_p_queue) 2026-10-12 00:25:52 +08:00
minke.yu 53fb9be7ee G9i: TTFT p95/p99 statistics (observability only, no admission gate) 2026-10-11 22:31:30 +08:00
minke.yu e5be84f795 G9i: probe classification across pool candidates (fast-P then regular)
t15 stress through the gateway showed /query_cache_hit failing on the
sole fast-P under exactly the load where classification matters (p1
queue 62 / pending 4.7M tokens, 160 probe errors -> fallback). Probe
order is now fast-P least-wait first, regular pool as backup; first
successful answer classifies; unclassifiable (multimodal 400) still
short-circuits; total failure falls back to default routing.
2026-10-11 21:15:59 +08:00
minke.yu 39ceba8a3d G9i: router-side pool admission + fast-P routing (gateway)
Pool-aware admission control and small-prefill routing, entirely in
sgl-model-gateway (zero engine changes). Consumes G9 P0 engine surfaces:
/v1/loads (per-dp-rank snapshots, summed to per-node instance-level
views) and /query_cache_hit (new-token classification).

Decision tree per request, before PD dispatch and outside the retry
loop (own 429s are never retried):
  1. D-pool gate: sum of fresh num_running_reqs over decode pool >
     threshold -> 429 d_pool_saturated (fail-open on missing snapshots;
     engine max-running-requests is the hard backstop).
  2. Chat routes: probe least-wait fast-P /query_cache_hit; new_tokens
     < max_new_tokens (tentative 4096) -> fast-P pool.
  3. Pool selection: least estimated wait (queued_total_input_tokens /
     max(drain_rate, floor)) — doubles as the load balancer for the
     multi-fast-P / multi-D future; pool membership is config-only.
  4. Fast-P queue gate: least-wait estimate > threshold (tentative 5s)
     -> 429 fast_p_queue with x-estimated-wait-s.

CLI: --g9i-enabled/--g9i-fast-prefill-url/--g9i-max-new-tokens/
--g9i-fast-p-queue-reject-secs/--g9i-decode-max-running/... all numbers
tentative and centralized (user directive: fixed by measurement later).
Metrics: smg_g9i_decisions_total/rejects_total/no_fresh_snapshot_total,
per-node running gauge, probe latency histogram.
2026-10-11 19:14:40 +08:00
minke.yu b9d259441c deploy: l3test stack image -> e7dc175e9 (G9 P0 fixes baked in) 2026-10-11 17:47:24 +08:00
minke.yu e7dc175e94 G9 P0 fixes: v2 pool_transfers for storage hit query, disagg drain hook, query gate
build-sglang-image / build (push) Failing after 21m51s
- unified_radix_cache.query_storage_hit_length: build PREFETCH-phase
  pool_transfers (comp prepare_prefetch + build_hicache_transfers +
  sidecars) so DeepSeek V4's logical-anchor KV goes through
  batch_exists_v2 side-pool existence instead of v1 batch_exists which
  always misses (no physical KV objects are stored for V4).
- scheduler.handle_query_cache_hit: allow L3 query when hicache storage
  backend is attached (not only decode-hicache mode; SWA models forbid
  decode radix+hicache so P nodes could never query L3); construct
  RadixKey from array('q') to match tree node key types.
- disaggregation/prefill: record_prefill_drain on final chunk in the
  disagg-prefill result path (it bypasses batch_result_processor, so
  drain_rate_tokens_per_s stayed 0 on P nodes).
- deploy: 3p1d l3test compose against the real mooncake store stack.
2026-10-11 17:37:21 +08:00
minke.yu 6612af9e0e deploy: 归档 b300 实况——l3dummy 镜像升 baba18bc7d + D max-running-requests 512 2026-10-11 15:40:04 +08:00
minke.yu 71c6d8bc55 G9 P0 fix: /query_cache_hit 绕过 _tokenize_one_request 的 ReqState 依赖
_tokenize_one_request._create_tokenized_object 会读 rid_to_state[obj.rid]
(正常生成路径在 tokenize 前注册 ReqState),纯查询无 rid → KeyError: None。
改为:dsv41 chat encoding 直接渲染的 prompt_ids 直接用;否则对渲染后
text 调 _tokenize_texts(与真实路径同一分词入口)。tok8 环境实测 200。
2026-10-11 15:38:35 +08:00
minke.yu dbc661602e deploy: 默认开 SGLANG_HICACHE_PREFETCH_IO_THREADS=4(io_aux 4 lane,mc4/mc8 +23%)
- 11 个挂 L3 的 compose 模板 + k8s 3p1d RBG manifest 的 prefill-base 统一显式落 4
  (代码默认已是 4,显式化防漂移;同一实例所有 rank 必须一致——crc32(rid)%N 分发
  + 每 lane 独立 gloo 组,不一致会集合通信错位挂死)
- D 角色不挂 L3,无需加
- A/B 数据见 investigations/l3-hicache-deep-dive-20261010.md §8
2026-10-10 23:07:45 +08:00
minke.yu 054dcee896 tier compose:跨实例 P2P 的 extra_hosts 互配(A/B 互加对方 rank 主机名别名,否则副本跨端分配后 open segment DNS 失败) 2026-10-10 21:45:16 +08:00
minke.yu 73a25267ac tier realstore-x2 compose:第二实例(GPU 2,3 / 端口 30041 / localhost-b 前缀)做跨 P 实例 L3 读写验证 2026-10-10 21:28:21 +08:00
minke.yu bcad628d30 tier realstore compose: 显式 SGLANG_HICACHE_PREFETCH_IO_THREADS=4(io_aux lane 数,配合 baba18bc7d) 2026-10-10 20:05:54 +08:00
minke.yu baba18bc7d hicache L3 prefetch: io_aux 传输并行化(按 op 分 lane)
build-sglang-image / build (push) Failing after 48m34s
打点实测单条 prefetch_io_aux_func 线程串行传输是所有请求的瓶颈:
mc8 下每请求 xfer 排队 0->1.9s 线性增长(每请求 5 批 x ~68ms 固定开销,
介质无关)。本提交把传输层拆成 N 条并行 lane(默认 4,
SGLANG_HICACHE_PREFETCH_IO_THREADS 可调):

- dispatch_prefetch() 用 crc32(request_id) % N 分 lane(内置 hash 对 str
  每进程加盐,跨 rank 不一致会导致 gloo 集合通信错位,禁用)
- 同一 op 的所有 chunk 保持在同一 lane,保序 reduce 语义不变
- 每 lane 独立 io 队列 + io_aux 线程 + sync 队列 + sync 线程 +
  独立 gloo completion 组(集合通信按 lane 过滤后的分发顺序保序)
- query 阶段的 hit_reduce 仍在单 prefetch_thread,不动
- 生产点(hiradix/unified/hybrid pp-ticket)改调 dispatch_prefetch
- 模拟器与两个单测文件同步适配

注意:SGLANG_HICACHE_PREFETCH_IO_THREADS 必须在实例内所有
TP/CP/PP rank 保持一致,否则 lane 组上的集合通信会错位挂死。
2026-10-10 19:04:49 +08:00
minke.yu 0afcb334a4 realstore 回退 STORAGE_BATCH_SIZE=512:TCP 传输下大批次倒退(mc8 668k→320k),dummy shm 才受益;真实 store 保持默认 128 2026-10-10 18:16:56 +08:00
minke.yu 33e6381b7d tier dummy compose 对齐 b300 实况:wait_complete + HC_TIMING + STORAGE_BATCH_SIZE=512 2026-10-10 18:06:48 +08:00
minke.yu 9c4ff32562 tier compose:STORAGE_BATCH_SIZE=512(摊薄每批 ~60ms 固定开销,dummy 实测 mc8 +26%) 2026-10-10 18:05:16 +08:00
minke.yu 7d5759b2d9 hicache: STORAGE_BATCH_SIZE 可env化(SGLANG_HICACHE_STORAGE_BATCH_SIZE,默认128不变)
build-sglang-image / build (push) Failing after 18m40s
打点显示每批存储 IO 有 ~30-70ms 固定开销(与传输介质无关),
调大批次可摊薄。env 化便于不动代码做 128/512/1024 对照。
2026-10-10 17:55:14 +08:00
minke.yu ac11e9162d tier realstore: extra_hosts 别名让 rank 后缀主机名可解析
mooncake P2P 模式下段名/te_endpoint 内嵌 local_hostname 且需 DNS 可解析;
rank 后缀名(localhost-d0t0p0c0 等)必须配 /etc/hosts 别名指向 127.0.0.1,
否则跨段 openSegment 握手失败(TRANSFER_FAIL 全灭)。
验证:mc8 402k→668k tok/s,读路径约 40% 走 memcpy 60% 仍 TCP。
2026-10-10 17:14:18 +08:00
minke.yu 84fc632088 mooncake_store: local_hostname 加 rank 后缀,让 master 的 preferred-segment 分配落到本进程段
build-sglang-image / build (push) Failing after 1s
两个 rank 进程共用默认 localhost 时,master 按名字匹配无法区分两个段,
副本会落到对方进程段;mooncake 的 local memcpy 快路径要求同进程 endpoint
(跨进程虚拟地址无效),于是静默回退 loopback TCP(实测 ~0.78GB/s)。
唯一化后自写自读可命中 memcpy(干净集群探针实测 lo 流量归零)。
注意:memcpy 线程池 kDefaultMemcpyWorkers=1(mooncake 0.3.13 硬编码),
单流 ~1GB/s,进一步提升需改 mooncake 并重编译。
2026-10-10 16:42:00 +08:00
minke.yu 8d9bce30a7 tier realstore: 开 HC_TIMING 打点 + prefetch policy 改 wait_complete
best_effort 下低并发请求 compute 赢比赛、L3 命中不生效(mc1/2/4 实为重算);
wait_complete 后真实 L3 命中:mc1 TTFT 2989→661ms,mc8 聚合 203k→402k tok/s。
2026-10-10 15:54:11 +08:00
minke.yu 5f2ffa70b8 hicache: SGLANG_HICACHE_TIMING 打点 L3 预取全链路分阶段耗时
build-sglang-image / build (push) Failing after 1s
env 开关(默认关),开后在以下阶段输出 HC_TIMING 日志行:
query_wait/query(含 batch_exists 逐批)/hit_count all_reduce/
xfer_wait/xfer(含 batch_get 逐批)/ack_reduce/ack_drain 排队/
graft(insert_host)/prefetch_e2e/loadback。
目的:定位 L3 命中单请求 ~3s(≈重算速度)的时间分布。
真实 mooncake store 与 dummy store 实测吞吐一致,排除后端实现因素。
2026-10-10 15:29:01 +08:00
minke.yu f4ef648b16 tier L3 A/B:dummy→真实 mooncake store(线上同构 master+P2PHANDSHAKE+tcp)+ 零预热 mc 扫描脚本
dummy 后端实测单请求 L3 命中 2.96s/54k tok/s(零预热修正后),聚合 ~200k;
CPU 在 scheduler native 线程、store 守护进程与 Python 工作线程均空闲,
疑似 dummy 后端 batch_get 内部 per-key 串行。切真实 store 复核以区分
sglang L3 路径问题 vs dummy 调试后端工件。注意 bench 必须 --warmup-requests 0,
默认 warmup 会把 prompt#0 种进 L1/L2 污染 L3 测量。
2026-10-10 13:59:31 +08:00
minke.yu 6469cf4197 deploy: tier bench scripts (cp2 L1/L2/L3 matrix + R7 async aggregation) 2026-10-10 10:50:24 +08:00
minke.yu 3f8751c6ed 新增 tier-cp2 单机 compose:cp2+dspark+L2 6.4M+L3 shm dummy,测 计算/L1/L2/L3 四档 prefill 极限与异步聚合 2026-10-10 09:51:09 +08:00
minke.yu e4f1310d6f l3dummy: hicache-ratio 1.0→4.0(L2=3.2M token/P)——#38348 bench 显示收益被 L2 容量掩盖(host_capacity 拒绝 0.75M~2.97M token/P),放大 L2 验证热轮命中不被容量拒、L3 hit 嫁接后在 L2 留存 2026-10-10 09:14:07 +08:00
minke.yu a42ab6fce0 测试适配 #38348 新语义:SWA 双层缺失的 backuped 节点现在是合法匹配边界
build-sglang-image / build (push) Failing after 1s
旧语义(host_hit=0 + branching=9 + 重插建窗)被 fail-soft 取代:
FULL 段直接 host 命中(host_hit=leaf_len、branching=None),load-back
落 FULL 后由续写请求的 insert overlap 从 Full KV 重建窗口 SWA。
重写 test_swa_branching_seqlen_uses_host_full_hit 验证新链路:
match 命中 → load-back → 插入扩展 token 触发 overlap 重建 → rematch
设备全命中且洞口节点的窗口 SWA 已恢复。
2026-10-10 07:42:44 +08:00
minke.yu fa8ad18b5c 手工移植上游 #38348:hicache SWA prefetch fail-soft 降级
build-sglang-image / build (push) Failing after 1s
SWA 池 L3 取回短缺时不再整批丢弃 prefetch 结果:保留已取回的 Full KV
前缀,仅丢弃失败的 SWA 池(释放其 host buffer、从 comp_xfers 摘除),
load-back 时经既有 SWARebuild/RecoverSWAWithLockedFull 路径从 Full KV
重建窗口。Mamba 等不可重建池短缺时仍走原 all-or-nothing 整批丢弃。

配套改动(同 PR):
- swa.py validator:SWA 双层缺失但节点已 backuped(Full KV 在 host)时
  放行作为匹配边界;LOAD_BACK 收集遇双层缺失空洞时 break 而非 assert。
- prefetch_from_storage:用完整匹配前缀重算 last_hash,修复 anchor 节点
  无 host_value(异步 backup 未完成/降级嫁接)时哈希链锚错位。
- 适配差异:本树有 sidecar_hash_values 重构,belief 自愈沿用
  invalidation_hashes + trimmed_pages 语义;ongoing_prefetch 以
  CacheRequestHandle 为键,comp_xfers 改为显式参数传入检查函数。
- 单测同步更新为 fail-soft 语义(TP 一致性降级、失败池释放无泄漏、
  LOAD_BACK 锚定空洞停止收集)。
2026-10-10 07:23:15 +08:00
minke.yu 7f1edbba80 cherry-pick 上游 #40429:FlashMLA sched meta 超 48KiB smem 时跳过预计算(encoder/decoder replay 的大 batch replay pass 会触发 smem 溢出崩溃)
build-sglang-image / build (push) Failing after 1s
2026-10-10 01:20:15 +08:00
minke.yu 99c317b131 cherry-pick 上游 #42716(运行时部分):encoder SWA bounded replay 下放行 host 层前缀命中——request_window 形态的 SWA 不进树,校验门不应拦截 host-only 节点;rust adapter/test 两处 hunk 与本树结构不符未合(默认 python 树芯用不到)
build-sglang-image / build (push) Failing after 1s
2026-10-10 01:02:33 +08:00
minke.yu 2b2ede8211 l3dummy 定型为 shm+best_effort(唯一正收益组合 +26.4%);wait_complete 因 discard 白等 -17.7% 弃用 2026-10-10 00:09:02 +08:00
minke.yu b4dbcebbec mooncake standalone(dummy) 模式打通 dsv41 unified cache:放宽 standalone 检查对 LogicalHostPool 锚点的误拦(实体 buffer 由旁池运行时 register_buffer);新增 l3dummy compose(单 store 服务 64GB shm 段 + fake-cargo 跳过 rustup 下载)。实测 24x160k 热/冷 +26.4%(vs TCP best_effort -4.4%)
build-sglang-image / build (push) Failing after 1h10m35s
2026-10-10 00:06:32 +08:00
minke.yu 30ba5818e7 l3test: prefetch 策略 wait_complete→best_effort(实测等待阻塞净亏 7-17%,best_effort 持平 -4.4%)
build-sglang-image / build (push) Failing after 1s
2026-10-09 20:30:46 +08:00
minke.yu c9bfb24bcc l3 模式:默认 48×160k + 冷热轮间 120s 备份排空 2026-10-09 19:31:45 +08:00
minke.yu 3d109d50be fix: l3test L2 改 ratio 1.0(0.05 会噎死 L3 写入路径) 2026-10-09 19:30:53 +08:00
minke.yu 4eabf5e228 G9 P0: loads 排队预估字段 + /query_cache_hit 命中查询端点
A. 负载预估数据源(决策②总输入长度口径):
- LoadSnapshot 新增 queued_total_input_tokens / drain_rate_tokens_per_s
- load_inquirer 复用 pending_token_queues 求和;scheduler 60s 滑动窗
  记录每请求最终 prefill chunk 完成事件(batch_result_processor 挂钩
  set_prefill_finished_time 处,恰好每请求一次)
- /v1/loads 与 PUB socket 同源自动暴露,gateway LoadMonitor 零改动消费

B. /query_cache_hit(决策④放 D):
- io_struct 新增 QueryCacheHitReqInput/Output IPC 类型
- scheduler.handle_query_cache_hit:tree_cache.match_prefix(L1/L2)
  + query_storage_hit_length(共享 L3,逻辑镜像 decode_hicache_mixin)
- tokenizer_control_mixin:fan-out 全部 DP rank 取最优命中
- serving_chat.query_cache_hit:复用 _convert_to_internal_request +
  _tokenize_one_request 保证 chat 模板/分词与真实路径一致;含图请求
  v1 拒绝(multimodal 400);http_server 注册 /query_cache_hit
2026-10-09 18:48:27 +08:00
minke.yu 7d67d57804 fix: l3test 变体改用 hicache-ratio 0.05(V4 不支持 hicache-size) 2026-10-09 17:55:02 +08:00
minke.yu 491989284f deploy: 3p1d-l3test 变体(P 缩 L1/L2 强制 L3 命中) 2026-10-09 17:26:04 +08:00
minke.yu 53ea8fb2a7 report: 增加 input tps 列 2026-10-09 17:02:23 +08:00
minke.yu efdbfd2cc3 fix: awk sprintf 括号+TPOT/TTFT 取最后一列 2026-10-09 17:01:35 +08:00
minke.yu 8342b89e85 fix: awk sprintf 括号 2026-10-09 17:01:05 +08:00
minke.yu c92938d718 deploy: bench_suite_v2 (warmup/prefill冷热双轮/decode阶梯80%cg/L3语料溢出) + 单独P/D基线compose 2026-10-09 16:50:23 +08:00
minke.yu 7578cdd218 revert: #42035 continuous prefill input polling (CP rank collective mismatch deadlock)
build-sglang-image / build (push) Failing after 39m1s
compose 3p1d regression on 1f6ca2cf6: all 3 P nodes wedged with split-brain
CP ranks (one blocked in _sync_hicache_ready_counts all_reduce, partner in
request broadcast). Root cause: #42035 wraps get_next_disagg_prefill_batch_to_run
(which contains the hicache attn_cp_group all_reduce) in rank-local guards
(len(result_queue)<2, _deferred_input_requests, skip_batch_creation); any
cross-rank divergence in batch-None decisions misaligns collective pairing
permanently. Continuous polling requires attn_cp_size==1 anyway, so cp2
deployments get zero benefit. Reverts 93d610c1fd + test adaptations
6ba34507f1/bde1fcbf51/1f6ca2cf68 + deploy bump 236686454c.
Keeps #42194 (9ef65b51c2) and checkpoint_disagg_prefill shim (c4f3e3be33).
Evidence: D:\B300\investigations\pd-deadlock-20261009.md
2026-10-09 16:39:06 +08:00
minke.yu 21f7ce37dd build: 镜像内编译 sgl-model-gateway(Rust gateway,G9 前置)
build-sglang-image / build (push) Failing after 31m52s
三个 Dockerfile(b300 全量/fast/gitea)在 COPY 源码后加 cargo 层:
cargo build --release -p sgl-model-gateway -p sgl-model-gateway-python,
RUSTUP_TOOLCHAIN=stable 绕过 rust-toolchain.toml 1.90 钉版(基底自带 1.98.1),
BuildKit cache mount 持久化 registry/target(冷 ~15min,增量秒~分钟级),
python 扩展产物重命名 sglang_router_rs.abi3.so 覆盖 python 包内同文件。
2026-10-09 b300 实测:crates.io 直连无需镜像,冷编译成功。
2026-10-09 16:07:36 +08:00
minke.yu 236686454c deploy: bump bench compose templates to 1f6ca2cf6 (#42035+#42194) 2026-10-09 15:06:41 +08:00
minke.yu 1f6ca2cf68 test: adopt upstream #42035 auxiliary-output asserts now that checkpoint_disagg_prefill shim exists
build-sglang-image / build (push) Successful in 25m55s
2026-10-09 11:32:33 +08:00
minke.yu bde1fcbf51 test: tolerate chunked kwarg in #42035 polling test override (branch shim keeps chunked flag) 2026-10-09 11:25:05 +08:00
minke.yu c4f3e3be33 refactor: route disagg prefill checkpoints through checkpoint_disagg_prefill shim
Upstream #42035's tests and call structure expect Scheduler.checkpoint_disagg_prefill
(introduced upstream with #41520's checkpoint_kv_cache / insert_req cache API).
This branch still uses maybe_cache_unfinished_req (old cache_unfinished_req API),
so add a thin shim with the upstream name and route the three call sites through it.
No behavior change.
2026-10-09 11:22:42 +08:00
Liangsheng Yin 9ef65b51c2 [mem_cache] Skip the release-time insert on optimistic prefill requeue; make refresh_fill_ids public (#42194)
build-sglang-image / build (push) Successful in 25m53s
(cherry picked from commit 07e1a4f082)
2026-10-09 11:18:17 +08:00
minke.yu 6ba34507f1 test: adapt #42035 polling test to pre-#41520 name maybe_cache_unfinished_req
build-sglang-image / build (push) Successful in 26m20s
2026-10-09 11:10:39 +08:00
metamergebot 93d610c1fd [PD] Keep ingesting requests while a prefill forward result is pending (#42035)
build-sglang-image / build (push) Successful in 26m6s
Co-authored-by: merrymercy <merrymercy@users.noreply.github.com>
Co-authored-by: Jialin <Jialin@users.noreply.github.com>
(cherry picked from commit 6acf2f4736)
2026-10-09 10:56:14 +08:00
minke.yu 4d0efcc921 deploy: bring k8s 3p1d RBG cw manifest into git; bump to 1068926cd (merged upstream batch)
Source of truth moves from D:\B300\build-kmctl (untracked) to
deploy/b300-ds41/ds41-3p1d-rbg-cw.yaml. Image carries A-batch (7) + B-batch (4)
upstream cherry-picks + is_hip_runtime import fix; bench-validated on compose
(3p1d 8-conc 160k 98903 tok/s, +50% vs 0928 baseline).
2026-10-08 17:41:06 +08:00
minke.yu f71fafa684 deploy: bump to 1068926cd (is_hip_runtime import fix) 2026-10-08 16:39:04 +08:00
minke.yu 1068926cd6 fix: import is_hip_runtime in dsv4 low_ratio_compress (cherry-pick #41660 lost it)
build-sglang-image / build (push) Successful in 32m18s
Upstream's file gets is_hip_runtime from sglang.kernels.jit.utils via an
intermediate refactor we don't have; our cherry-pick applied cleanly but
the resulting file references is_hip_runtime() without importing it,
crashing every prefill (NameError in c2_prefill_norm_rope_store).

Also unify Dockerfile.fast with the v3.1 full path (check_deps + --no-deps
delta installs) so fast builds after a kernel-wheel bump don't silently
revert to the base image's older kernels.
2026-10-08 16:31:34 +08:00
minke.yu 48fb49baef deploy: bump cpdspark-nr/b4-dp/3p1d-otlp to dsv41-pd-visioncp-2372cdc06-local-20261008-0808 (merged upstream fixes batch) 2026-10-08 16:21:30 +08:00
minke.yu 2372cdc06e build: light delta also --no-deps — any dep-graph walk over torch re-downloads pinned nccl==2.29.7 over base's 2.30.7 2026-10-08 16:07:50 +08:00
minke.yu 7b815cc98b build: COPY check_deps.py from sglang/ tree (context .dockerignore excludes build/) 2026-10-08 16:01:12 +08:00
minke.yu d06c5dde42 build: v3.1 three-class dep delta — kernel wheels (sglang-kernel/flashinfer/sgl-deep-gemm) install via pip --no-deps when pins move
Base ships sglang-kernel 0.4.6.post1 but pyproject pins ==0.4.7; kernel
wheels must track the sglang source, so they are an allowed delta (~370MB
wheel), unlike the torch/nvidia family which must always reuse the base.
Also confirmed 2026-09-24 full build silently downgraded NCCL 2.30.7->2.29.7;
v3 keeps base NCCL untouched.
2026-10-08 15:59:36 +08:00
minke.yu 8d0ee5e51e build: v3 full path reuses base torch/nvidia via local delta check
build-sglang-image / build (push) Successful in 28m50s
Full-resolution pip walks installed torch's metadata and would re-download
the pinned nvidia family (~6GB, throttled to 2h40m on 2026-09-23) and
downgrade the deliberately newer nvidia-nccl-cu13 2.30.7 to ==2.29.7.
v3 replaces pip full-resolution with check_deps.py: compare pyproject
requirements against base-installed versions locally, pip-install only the
lightweight delta via -r, and fail loudly if a heavy pin (torch*/nvidia-*/
sglang-kernel/flashinfer/cuda-*) becomes unsatisfied. Editable install keeps
--no-deps --no-build-isolation.
2026-10-08 15:54:58 +08:00
metamergebot 75758cb0a0 [HiCache] Fix cgroup page-cache accounting and the sizing fallback when cgroup discovery fails (#42420)
Co-authored-by: merrymercy <merrymercy@users.noreply.github.com>
Co-authored-by: xiezhq-hermann <xiezhq-hermann@users.noreply.github.com>
Co-authored-by: nvjullin <nvjullin@users.noreply.github.com>
(cherry picked from commit 4ab720e655)
2026-10-08 15:41:51 +08:00
metamergebot 60e80ac7d1 [HiCache] Fix write-back SWA insert backups tripping the write-through pending-ack assert (#42264)
Co-authored-by: merrymercy <merrymercy@users.noreply.github.com>
Co-authored-by: 842974287 <842974287@users.noreply.github.com>
(cherry picked from commit 2bae12b9b3)
2026-10-08 15:37:43 +08:00
Sasha Sidorov 4ef799aaa6 [PD] Wait for prefill completion before Mooncake early KV transfer (#41395)
Co-authored-by: Shangming Cai <csmthu@gmail.com>
Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>
(cherry picked from commit c683190b40)
2026-10-08 15:36:21 +08:00
Aurick Qiao 1be762368b [Disagg] Validate state strides before Mooncake transfers (#41607)
Co-authored-by: Aurick Qiao <6137920+aurickq@users.noreply.github.com>
(cherry picked from commit 1a0250c5b6)
2026-10-08 15:35:05 +08:00
DarkSharpness f577b61b52 [DSv4.1] Trim prefill model-entry overhead (#41704)
Co-authored-by: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Co-authored-by: BBuf <1182563586@qq.com>
(cherry picked from commit 2f6e0ec45d)
2026-10-08 15:32:48 +08:00
DarkSharpness 4792b61332 [DSv4.1] Fused c1/c2 compress for eager extend, faster c2 decode (#41660)
Co-authored-by: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Co-authored-by: BBuf <1182563586@qq.com>
(cherry picked from commit 5916999afb)
2026-10-08 15:32:48 +08:00
DarkSharpness 62f1ed099e [DSv4.1] Faster fp4 index-K gather and combine_topk_swa_indices (#41658)
Co-authored-by: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Co-authored-by: BBuf <1182563586@qq.com>
(cherry picked from commit d75d5b33f7)
2026-10-08 15:32:47 +08:00
AMD-yanfeiwang bfd9bb2e03 fix(cuda-graph): remove unnecessary CP batch alignment (#40543)
(cherry picked from commit 3dc7f4b816)
2026-10-08 15:32:46 +08:00
metamergebot 6deac260bf [PD] Make the prebuilt last-token H2D copy non-blocking under overlap (#42113)
Co-authored-by: merrymercy <merrymercy@users.noreply.github.com>
Co-authored-by: bbus <bbus@users.noreply.github.com>
(cherry picked from commit 7dd93b01a5)
2026-10-08 15:32:46 +08:00
Yuwei An 842ad7c4d4 [Fix][DSV4.1] SWA page size with bounded replay (#42128)
(cherry picked from commit ef867fa40d)
2026-10-08 15:32:45 +08:00
Mohammad Miadh Angkad b605d2d370 [DSA] Stop forcing the per-step CPU seq_lens sync for the k-pool indexer (#41987)
Co-authored-by: Mohammad Angkad <mohammad.angkad@radixark.ai>
(cherry picked from commit 42f2e4fcb7)
2026-10-08 15:32:44 +08:00
100 changed files with 16229 additions and 400 deletions
+13
View File
@@ -8,4 +8,17 @@ FROM lmsysorg/sglang:dev-dsv41
# checkout 含 .git,setuptools-scm 可打戳(main 系分支 version 显示 dev 属正常)
COPY . /sgl-workspace/sglang
# Rust gateway(sgl-model-gateway)编译,与 b300 离线版 deploy/b300-build/Dockerfile 同构。
# 基底自带 rustup stable,RUSTUP_TOOLCHAIN=stable 绕过 rust-toolchain.toml 的 1.90 钉版;
# python 扩展必须显式 -p,产物重命名为 sglang_router_rs.abi3.so 覆盖 python 包内同文件。
# 海外 runner 直连 crates.io 无网络问题;cache mount 海外环境同样无害。
RUN --mount=type=cache,target=/cargo-home \
--mount=type=cache,target=/cargo-target \
cd /sgl-workspace/sglang/sgl-model-gateway \
&& CARGO_HOME=/cargo-home CARGO_TARGET_DIR=/cargo-target RUSTUP_TOOLCHAIN=stable \
cargo build --release -p sgl-model-gateway -p sgl-model-gateway-python \
&& cp /cargo-target/release/sgl-model-gateway /opt/sglang/bin/sgl-model-gateway \
&& cp /cargo-target/release/libsglang_router_rs.so \
/opt/sglang/lib/python3.12/site-packages/sglang_router/sglang_router_rs.abi3.so
RUN pip install --no-cache-dir -e /sgl-workspace/sglang/python
+61 -5
View File
@@ -1,9 +1,19 @@
# 优化版(2026-09-24):解决清华源限流下每次构建重下 torch 的问题。
# 优化版(2026-10-08 v3):torch/nvidia 全家桶永远复用基底,全量路径零重下。
# 要点:
# 1. pyproject build-system.requires 含 torch==2.13.0,pip 隔离构建环境每次都重下 ~6GB。
# 基底镜像已装 torch 2.13 → 用 --no-build-isolation 复用,构建环境零下载。
# 2. 基底缺 setuptools-scm(版本打戳要用),单层预装 + pip cache mount,只下载一次。
# 3. RUN 的 pip cache mount 持久化在宿主机,依赖解析命中的包不再重复下载。
# 2. 老全量路径让 pip 解析全部依赖:bare `torchvision` 等会让解析器弃用基底已装的
# 本地版本(torch 2.13.0+cu130),改用 PyPI 的 torch==2.13.0,连带重下整套 nvidia
# 依赖(9/23 实测 nvidia-nccl-cu13 等 6GB+,限流下 2h40m);且基底 NCCL 是
# 刻意装高的 2.30.7(torch 元数据钉 ==2.29.7),任何全量解析都会把它降级回去。
# → v3 不再让 pip 碰重型依赖:check_deps.py 在构建期对照 pyproject 与基底
# 已装版本,把增量分三类——重型(torch*/nvidia-*/cuda-*)永不进解析、
# 不满足即让构建失败并提示决策(显式失败好过静默重装);
# kernel wheel(sglang-kernel/flashinfer*/sgl-deep-gemm)要跟随源码版本,
# 钉版变动时以 pip --no-deps 单独装(防止顺藤摸瓜解析 torch);
# 其余轻量依赖只装未满足的增量。
# 3. 基底缺 setuptools-scm(版本打戳要用),单层预装 + pip cache mount,只下载一次。
# 4. RUN 的 pip cache mount 持久化在宿主机,轻量依赖增量不再重复下载。
FROM uhub.service.ucloud.cn/umirror/sglang:dev-dsv41
ARG PIP_INDEX=https://pypi.tuna.tsinghua.edu.cn/simple
@@ -13,9 +23,55 @@ ENV PIP_INDEX_URL=${PIP_INDEX}
RUN --mount=type=cache,target=/root/.cache/pip \
pip install "setuptools>=61" "setuptools-scm>=8" "setuptools-rust>=1.11" wheel
# 依赖增量计算脚本(对照 pyproject 与基底已装版本,产出轻量/kernel 增量清单)
# 注意 build context 是 /data/ymk 且 .dockerignore 排除了 build/,只能从 sglang/ 树内 COPY
COPY sglang/deploy/b300-build/check_deps.py /tmp/check_deps.py
# 整个源码树(含 .git,用于 setuptools-scm 版本打戳)
COPY sglang/ /sgl-workspace/sglang/
# editable 安装,全量依赖解析(pyproject.toml 变化时走这个)
# Rust gateway(sgl-model-gateway)编译。基底自带 rustup stable(实测 cargo 1.98.1),
# RUSTUP_TOOLCHAIN=stable 绕过仓库 rust-toolchain.toml 钉的 1.90(避免重复下载工具链);
# b300 直连 crates.io(稀疏索引 + static.crates.io 均通),无需任何镜像源。
# cache mount 持久化 registry/target:冷编译 ~15min(二进制 ~11min + python 扩展 ~4min),
# 增量重建秒~分钟级。python 扩展必须显式 -p(workspace 根包默认不编译扩展),
# 产物 libsglang_router_rs.so 需重命名为 sglang_router_rs.abi3.so 覆盖 python 包内同文件
# (与 H20 线上部署方式一致)。2026-10-09 b300 实测验证。
RUN --mount=type=cache,target=/cargo-home \
--mount=type=cache,target=/cargo-target \
cd /sgl-workspace/sglang/sgl-model-gateway \
&& CARGO_HOME=/cargo-home CARGO_TARGET_DIR=/cargo-target RUSTUP_TOOLCHAIN=stable \
cargo build --release -p sgl-model-gateway -p sgl-model-gateway-python \
&& cp /cargo-target/release/sgl-model-gateway /opt/sglang/bin/sgl-model-gateway \
&& cp /cargo-target/release/libsglang_router_rs.so \
/opt/sglang/lib/python3.12/site-packages/sglang_router/sglang_router_rs.abi3.so
# 计算依赖增量(重型 torch/nvidia/cuda 永不进入清单;不满足则 check_deps 非零、构建失败)
RUN python /tmp/check_deps.py /sgl-workspace/sglang/python/pyproject.toml \
/tmp/missing_light.txt /tmp/missing_kernel.txt \
&& echo "== kernel wheel delta (pip --no-deps) ==" && cat /tmp/missing_kernel.txt \
&& echo "== lightweight dep delta ==" && cat /tmp/missing_light.txt
# kernel wheel 增量(如 sglang-kernel 升版):--no-deps 防止 pip 顺藤摸瓜解析 torch
RUN --mount=type=cache,target=/root/.cache/pip \
pip install --no-build-isolation -e /sgl-workspace/sglang/python
if grep -q '[^[:space:]]' /tmp/missing_kernel.txt; then \
pip install --no-deps -r /tmp/missing_kernel.txt; \
else \
echo "== no kernel wheel delta"; \
fi
# 轻量依赖增量(通常为空或几个小包),重型全家桶原样复用基底。
# ⚠️ 必须 --no-deps:即使目标只是 xgrammar 这种小包,只要依赖图里出现 torch,
# pip 就会校验 torch+cu130 元数据钉死的 nvidia-nccl-cu13==2.29.7,发现基底装的是
# 2.30.7 便"好心"下载 206MB 降级(10-08 v3 首跑实测踩中)。基底的传递依赖本来就
# 齐全,增量小包的二级依赖也一样;若有漏网,b300 冒烟会在 import 阶段显式暴露。
RUN --mount=type=cache,target=/root/.cache/pip \
if grep -q '[^[:space:]]' /tmp/missing_light.txt; then \
pip install --no-deps -r /tmp/missing_light.txt; \
else \
echo "== no lightweight delta, skipping dependency install"; \
fi
# editable 安装 sglang 本体,--no-deps 保证 pip 完全不碰依赖图
RUN --mount=type=cache,target=/root/.cache/pip \
pip install --no-deps --no-build-isolation -e /sgl-workspace/sglang/python
+53 -3
View File
@@ -1,15 +1,65 @@
# fast 路径:python/pyproject.toml 未变化时使用(build.sh 自动判断)。
# --no-deps 跳过全部依赖解析(连 index 元数据请求都省掉),纯源码迭代秒级完成。
# 依赖有变化时必须走 Dockerfile 全量路径(build.sh 按 pyproject sha256 自动切换)。
# fast 路径(2026-10-08 v3.1):python/pyproject.toml 未变化时使用(build.sh 自动判断)。
# 与 Dockerfile 全量路径同构:同样跑 check_deps.py 增量安装。
# 原因:全量构建若装过 kernel wheel 增量(如 sglang-kernel 0.4.7),随后 pyproject 未变的
# fast 构建若只 --no-deps,会静默退回基底的旧 kernel(0.4.6.post1)——镜像内容漂移。
# check_deps 本地比对 ~0.5s,增量命中 pip cache 秒级,成本可忽略。
FROM uhub.service.ucloud.cn/umirror/sglang:dev-dsv41
ARG PIP_INDEX=https://pypi.tuna.tsinghua.edu.cn/simple
ENV PIP_INDEX_URL=${PIP_INDEX}
# 构建后端(pyproject build-system.requires,torch 除外——基底已有)
RUN --mount=type=cache,target=/root/.cache/pip \
pip install "setuptools>=61" "setuptools-scm>=8" "setuptools-rust>=1.11" wheel
# 依赖增量计算脚本(对照 pyproject 与基底已装版本,产出轻量/kernel 增量清单)
# 注意 build context 是 /data/ymk 且 .dockerignore 排除了 build/,只能从 sglang/ 树内 COPY
COPY sglang/deploy/b300-build/check_deps.py /tmp/check_deps.py
# 整个源码树(含 .git,用于 setuptools-scm 版本打戳)
COPY sglang/ /sgl-workspace/sglang/
# Rust gateway(sgl-model-gateway)编译。基底自带 rustup stable(实测 cargo 1.98.1),
# RUSTUP_TOOLCHAIN=stable 绕过仓库 rust-toolchain.toml 钉的 1.90(避免重复下载工具链);
# b300 直连 crates.io(稀疏索引 + static.crates.io 均通),无需任何镜像源。
# cache mount 持久化 registry/target:冷编译 ~15min,增量重建秒~分钟级。
# python 扩展必须显式 -p(workspace 根包默认不编译扩展),产物 libsglang_router_rs.so
# 需重命名为 sglang_router_rs.abi3.so 覆盖 python 包内同文件(与 H20 线上部署方式一致)。
RUN --mount=type=cache,target=/cargo-home \
--mount=type=cache,target=/cargo-target \
cd /sgl-workspace/sglang/sgl-model-gateway \
&& CARGO_HOME=/cargo-home CARGO_TARGET_DIR=/cargo-target RUSTUP_TOOLCHAIN=stable \
cargo build --release -p sgl-model-gateway -p sgl-model-gateway-python \
&& cp /cargo-target/release/sgl-model-gateway /opt/sglang/bin/sgl-model-gateway \
&& cp /cargo-target/release/libsglang_router_rs.so \
/opt/sglang/lib/python3.12/site-packages/sglang_router/sglang_router_rs.abi3.so
# 计算依赖增量(重型 torch/nvidia/cuda 永不进入清单;不满足则 check_deps 非零、构建失败)
RUN python /tmp/check_deps.py /sgl-workspace/sglang/python/pyproject.toml \
/tmp/missing_light.txt /tmp/missing_kernel.txt \
&& echo "== kernel wheel delta (pip --no-deps) ==" && cat /tmp/missing_kernel.txt \
&& echo "== lightweight dep delta ==" && cat /tmp/missing_light.txt
# kernel wheel 增量(如 sglang-kernel 升版):--no-deps 防止 pip 顺藤摸瓜解析 torch
RUN --mount=type=cache,target=/root/.cache/pip \
if grep -q '[^[:space:]]' /tmp/missing_kernel.txt; then \
pip install --no-deps -r /tmp/missing_kernel.txt; \
else \
echo "== no kernel wheel delta"; \
fi
# 轻量依赖增量(通常为空或几个小包),重型全家桶原样复用基底。
# ⚠️ 必须 --no-deps:即使目标只是 xgrammar 这种小包,只要依赖图里出现 torch,
# pip 就会校验 torch+cu130 元数据钉死的 nvidia-nccl-cu13==2.29.7,发现基底装的是
# 2.30.7 便"好心"下载 206MB 降级(10-08 v3 首跑实测踩中)。基底的传递依赖本来就
# 齐全,增量小包的二级依赖也一样;若有漏网,b300 冒烟会在 import 阶段显式暴露。
RUN --mount=type=cache,target=/root/.cache/pip \
if grep -q '[^[:space:]]' /tmp/missing_light.txt; then \
pip install --no-deps -r /tmp/missing_light.txt; \
else \
echo "== no lightweight delta, skipping dependency install"; \
fi
# editable 安装 sglang 本体,--no-deps 保证 pip 完全不碰依赖图
RUN --mount=type=cache,target=/root/.cache/pip \
pip install --no-deps --no-build-isolation -e /sgl-workspace/sglang/python
+90
View File
@@ -0,0 +1,90 @@
#!/usr/bin/env python3
"""Compute dependency deltas for a b300 sglang image build, in three classes.
Reads pyproject.toml, evaluates every requirement against the *installed*
distributions in the base image, and writes two requirements files:
argv[2] (light): unsatisfied lightweight requirements → `pip install -r`
argv[3] (kernel): unsatisfied kernel wheels (sglang-kernel, flashinfer*,
sgl-deep-gemm) → `pip install --no-deps -r`
These track the sglang source being built and are meant
to be upgraded from the index when pyproject bumps them;
--no-deps keeps pip from walking torch's metadata.
The PyTorch stack (torch/torchvision/torchaudio/torchcodec/
torch_memory_saver), the nvidia-* CUDA family and cuda-* packages are NEVER
emitted. If one of them is missing or its installed version no longer
satisfies the pinned specifier (e.g. upstream bumps nvidia-cutlass-dsl), the
script fails the build loudly instead of letting pip download ~6GB or
"fix" the base image's deliberately newer nvidia-nccl-cu13 2.30.7 back to
torch's metadata pin ==2.29.7 (any pip full-resolution would do both —
2026-09-23/24 builds did, throttled to 2h40m).
"""
import sys
import tomllib
from importlib import metadata as md
from packaging.requirements import Requirement
# Never resolved by pip: reused from the base image or the build fails.
HEAVY_PREFIXES = ("torch", "nvidia-", "cuda-")
# Kernel wheels: upgraded via pip --no-deps when pyproject pins move.
KERNEL_PREFIXES = ("sglang-kernel", "flashinfer", "sgl-deep-gemm")
def norm(name: str) -> str:
return name.lower().replace("_", "-")
def main() -> int:
pyproject, light_out, kernel_out = sys.argv[1], sys.argv[2], sys.argv[3]
with open(pyproject, "rb") as f:
deps = tomllib.load(f)["project"]["dependencies"]
light: list[str] = []
kernel: list[str] = []
heavy_unsat: list[str] = []
for dep in deps:
req = Requirement(dep)
if req.marker is not None and not req.marker.evaluate():
continue
try:
installed = md.version(req.name)
except md.PackageNotFoundError:
installed = None
satisfied = installed is not None and req.specifier.contains(
installed, prereleases=True
)
if satisfied:
continue
name = norm(req.name)
if name.startswith(HEAVY_PREFIXES):
heavy_unsat.append(f"{dep} (installed: {installed})")
elif name.startswith(KERNEL_PREFIXES):
kernel.append(dep)
else:
light.append(dep)
if heavy_unsat:
print("=" * 70, file=sys.stderr)
print("FATAL: heavy base-image deps no longer satisfy pyproject:", file=sys.stderr)
for item in heavy_unsat:
print(f" {item}", file=sys.stderr)
print(
"Refusing to let pip resolve these (would re-download ~6GB and/or\n"
"downgrade NCCL). Decide explicitly: bump the base image, or adjust\n"
"the pin in pyproject.toml.",
file=sys.stderr,
)
return 1
with open(light_out, "w") as f:
f.write("\n".join(light) + ("\n" if light else ""))
with open(kernel_out, "w") as f:
f.write("\n".join(kernel) + ("\n" if kernel else ""))
return 0
if __name__ == "__main__":
sys.exit(main())
+27
View File
@@ -0,0 +1,27 @@
cd /data/ymk/ds41
HDR="Authorization=Bearer deepseek-v4-flash-0731-key"
PORT=30040
MODEL=/data/models/DeepSeek-V4.1-Flash
# 一律 --warmup-requests 0:默认 warmup 会把 prompt#0 种进 L1/L2 污染 L3 测量(已踩过)
run() {
echo "===== [$(date +%H:%M:%S)] RUN: $1 (np=$2) ====="
python3 -m sglang.bench_serving --backend sglang --host 127.0.0.1 --port $PORT --model $MODEL \
--dataset-name random-ids --random-input-len 160000 --random-output-len 1 \
--warmup-requests 0 --num-prompts "$2" --max-concurrency "$2" --seed "$3" --tokenize-prompt \
--random-range-ratio 1.0 --header "$HDR" 2>&1 | grep -E 'Input token throughput|Mean TTFT|P99 TTFT|Successful'
sleep 2
}
flush() { curl -sf -m 10 -X POST http://127.0.0.1:$PORT/flush_cache > /dev/null && echo "[$(date +%H:%M:%S)] flushed"; sleep 3; }
echo "===== PRIME COLD mc8(冷基线 + 写入 L3) ====="
run PRIME_COLD 8 1541391
flush
run MC1_L3 1 1541391
flush
run MC2_L3 2 1541391
flush
run MC4_L3 4 1541391
flush
run MC8_L3 8 1541391
echo "===== [$(date +%H:%M:%S)] ALL DONE ====="
+204
View File
@@ -0,0 +1,204 @@
#!/bin/bash
# bench_suite_v2.sh — DeepSeek-V4.1-Flash on b300-01,2026-10-09 新版基线方法
#
# 相对 v1 的方法变化:
# 1. 服务刚起数据不准 → 每个测试模式自带 warmup 轮(结果丢弃,标记 WARM)
# 2. prefill 冷/热双轮:冷轮写入 hicache(L1/L3),同 seed 热轮测纯 cache-hit TPS
# 3. decode 阶梯并发:步长 32,上限 = cuda-graph-max-bs-decode 的 80%;
# 每档先 warm(灌 L1)再测,prefill 全 cache hit;
# 报告 TPOT + 整实例输出 TPS + 推算实际并发(= 输出TPS × TPOT / 1000)
# 4. l3 模式(3p1d 经 router):大语料溢出 L1 后重放,验证 L3 命中
#
# 在宿主机执行,bench 客户端通过 docker exec 进容器跑(需要 /data 挂载)。
#
# 用法:
# bench_suite_v2.sh warmup <container> <port>
# bench_suite_v2.sh prefill <tag> <container> <port>
# bench_suite_v2.sh decode <tag> <container> <port> [cg_max_bs=256]
# bench_suite_v2.sh l3 <tag> <container> <port> [num_prompts=24]
# bench_suite_v2.sh report <log>
# 产物: /data/ymk/ds41/bench_v2_<tag>.log
set -u
BASE=/data/ymk/ds41
MODEL=${MODEL:-/data/models/DeepSeek-V4.1-Flash}
HDR="Authorization=Bearer deepseek-v4-flash-0731-key"
usage() { sed -n 2,20p "$0" >&2; exit 1; }
# $1=container $2=port $3=tag -> 生成可执行脚本路径并 echo
emit_common() {
cat <<EOF
cd $BASE
HDR="$HDR"
PORT=$2
run() {
local name="\$1"; shift
echo "===== [\$(date +%H:%M:%S)] START: \$name ====="
python3 -m sglang.bench_serving "\$@"
echo "===== [\$(date +%H:%M:%S)] END(exit=\$?): \$name ====="
sleep 5
}
EOF
}
# ---------------- warmup ----------------
cmd_warmup() {
local c=${1:?container} port=${2:?port}
local gen=$BASE/.bench_v2_warmup.sh
{ emit_common "$c" "$port"; cat <<'EOF'
run WARMUP_a --backend sglang --host 127.0.0.1 --port $PORT --model MODELPH \
--dataset-name random-ids --random-input-len 4096 --random-output-len 16 \
--num-prompts 8 --max-concurrency 8 --seed 777001 --tokenize-prompt \
--random-range-ratio 1.0 --header "$HDR"
run WARMUP_b --backend sglang --host 127.0.0.1 --port $PORT --model MODELPH \
--dataset-name random-ids --random-input-len 16000 --random-output-len 1 \
--num-prompts 2 --max-concurrency 1 --seed 777002 --tokenize-prompt \
--random-range-ratio 1.0 --header "$HDR"
echo "===== [$(date +%H:%M:%S)] ALL DONE ====="
EOF
} | sed "s|MODELPH|$MODEL|g" > "$gen"
chmod +x "$gen"
docker exec "$c" bash "$gen"
}
# ---------------- prefill(冷/热双轮) ----------------
cmd_prefill() {
local tag=${1:?tag} c=${2:?container} port=${3:?port}
local p=$(date +%H%M%S) gen=$BASE/.bench_v2_gen_${tag}.sh
{ emit_common "$c" "$port"
# warmup(丢弃)
cat <<'EOF'
run WARMUP --backend sglang --host 127.0.0.1 --port $PORT --model MODELPH \
--dataset-name random-ids --random-input-len 4096 --random-output-len 16 \
--num-prompts 8 --max-concurrency 8 --seed 777000 --tokenize-prompt \
--random-range-ratio 1.0 --header "$HDR"
EOF
# 冷轮
for stage in "bs1_16000 16000 1 4 1" "bs1_160000 160000 1 4 2" "bs16_4096 4096 16 32 3"; do
set -- $stage
cat <<EOF
run COLD_$1 --backend sglang --host 127.0.0.1 --port \$PORT --model $MODEL \\
--dataset-name random-ids --random-input-len $2 --random-output-len 1 \\
--num-prompts $4 --max-concurrency $3 --seed ${p}$5 --tokenize-prompt \\
--random-range-ratio 1.0 --header "\$HDR"
EOF
done
# 热轮(同 seed → cache hit)
for stage in "bs1_16000 16000 1 4 1" "bs1_160000 160000 1 4 2" "bs16_4096 4096 16 32 3"; do
set -- $stage
cat <<EOF
run HIT_$1 --backend sglang --host 127.0.0.1 --port \$PORT --model $MODEL \\
--dataset-name random-ids --random-input-len $2 --random-output-len 1 \\
--num-prompts $4 --max-concurrency $3 --seed ${p}$5 --tokenize-prompt \\
--random-range-ratio 1.0 --header "\$HDR"
EOF
done
echo 'echo "===== [$(date +%H:%M:%S)] ALL DONE ====="'
} | sed "s|MODELPH|$MODEL|g" > "$gen"
chmod +x "$gen"
nohup docker exec "$c" bash "$gen" > "$BASE/bench_v2_${tag}.log" 2>&1 &
echo "launched pid=$! seeds=${p}1..3 log=$BASE/bench_v2_${tag}.log"
}
# ---------------- decode 阶梯 ----------------
cmd_decode() {
local tag=${1:?tag} c=${2:?container} port=${3:?port} cg=${4:-256}
local maxc=$(( cg * 80 / 100 / 32 * 32 )) # 80% 向下取整到 32 的倍数
local p=$(date +%H%M%S) gen=$BASE/.bench_v2_gen_${tag}.sh
echo "ladder: 32..$maxc step 32 (cg_max_bs=$cg)"
{ emit_common "$c" "$port"
cat <<'EOF'
run WARMUP --backend sglang --host 127.0.0.1 --port $PORT --model MODELPH \
--dataset-name random-ids --random-input-len 2048 --random-output-len 128 \
--num-prompts 16 --max-concurrency 16 --seed 777000 --tokenize-prompt \
--random-range-ratio 1.0 --header "$HDR"
EOF
# bs1 延迟地板(warm + measure)
cat <<EOF
run WARM_bs1 --backend sglang --host 127.0.0.1 --port \$PORT --model $MODEL \\
--dataset-name random-ids --random-input-len 2048 --random-output-len 512 \\
--num-prompts 4 --max-concurrency 1 --seed ${p}01 --tokenize-prompt \\
--random-range-ratio 1.0 --header "\$HDR"
run DEC_bs1 --backend sglang --host 127.0.0.1 --port \$PORT --model $MODEL \\
--dataset-name random-ids --random-input-len 2048 --random-output-len 512 \\
--num-prompts 4 --max-concurrency 1 --seed ${p}01 --tokenize-prompt \\
--random-range-ratio 1.0 --header "\$HDR"
EOF
local cc
for (( cc=32; cc<=maxc; cc+=32 )); do
local np=$(( cc * 2 ))
cat <<EOF
run WARM_c${cc} --backend sglang --host 127.0.0.1 --port \$PORT --model $MODEL \\
--dataset-name random-ids --random-input-len 2048 --random-output-len 512 \\
--num-prompts $np --max-concurrency $cc --seed ${p}1${cc} --tokenize-prompt \\
--random-range-ratio 1.0 --header "\$HDR"
run DEC_c${cc} --backend sglang --host 127.0.0.1 --port \$PORT --model $MODEL \\
--dataset-name random-ids --random-input-len 2048 --random-output-len 512 \\
--num-prompts $np --max-concurrency $cc --seed ${p}1${cc} --tokenize-prompt \\
--random-range-ratio 1.0 --header "\$HDR"
EOF
done
echo 'echo "===== [$(date +%H:%M:%S)] ALL DONE ====="'
} | sed "s|MODELPH|$MODEL|g" > "$gen"
chmod +x "$gen"
nohup docker exec "$c" bash "$gen" > "$BASE/bench_v2_${tag}.log" 2>&1 &
echo "launched pid=$! log=$BASE/bench_v2_${tag}.log"
}
# ---------------- L3 语料溢出(3p1d 经 router) ----------------
cmd_l3() {
local tag=${1:?tag} c=${2:?container} port=${3:?port} np=${4:-48}
local p=$(date +%H%M%S) gen=$BASE/.bench_v2_gen_${tag}.sh
{ emit_common "$c" "$port"
cat <<EOF
run WARMUP --backend sglang --host 127.0.0.1 --port \$PORT --model $MODEL \\
--dataset-name random-ids --random-input-len 4096 --random-output-len 16 \\
--num-prompts 8 --max-concurrency 8 --seed 777000 --tokenize-prompt \\
--random-range-ratio 1.0 --header "\$HDR"
run L3_COLD_160k --backend sglang --host 127.0.0.1 --port \$PORT --model $MODEL \\
--dataset-name random-ids --random-input-len 160000 --random-output-len 1 \\
--num-prompts $np --max-concurrency 8 --seed ${p}1 --tokenize-prompt \\
--random-range-ratio 1.0 --header "\$HDR"
echo "===== [\$(date +%H:%M:%S)] DRAIN: sleep 120 等 L2→L3 备份排空 ====="
sleep 120
run L3_HIT_160k --backend sglang --host 127.0.0.1 --port \$PORT --model $MODEL \\
--dataset-name random-ids --random-input-len 160000 --random-output-len 1 \\
--num-prompts $np --max-concurrency 8 --seed ${p}1 --tokenize-prompt \\
--random-range-ratio 1.0 --header "\$HDR"
EOF
echo 'echo "===== [$(date +%H:%M:%S)] ALL DONE ====="'
} > "$gen"
chmod +x "$gen"
nohup docker exec "$c" bash "$gen" > "$BASE/bench_v2_${tag}.log" 2>&1 &
echo "launched pid=$! corpus=${np}x160k log=$BASE/bench_v2_${tag}.log"
}
# ---------------- report ----------------
cmd_report() {
local log=${1:?log required}
awk '
/START:/ { name=$4; sub(/:$/, "", name); tps=""; itps=""; tpot=""; ttft=""; acc="" }
/Input token throughput/ { itps=$NF }
/Output token throughput/ { tps=$NF }
/Mean TPOT/ { tpot=$NF }
/Mean TTFT/ { ttft=$NF }
/[Aa]ccept/ { acc=$0; gsub(/^ +/, "", acc) }
/END\(exit/ {
line=sprintf("%-22s in_tps=%-10s out_tps=%-10s tpot_ms=%-8s ttft_ms=%-8s", name, itps, tps, tpot, ttft)
if (tps != "" && tpot != "" && tpot+0 > 0)
line=line sprintf(" eff_concurrency=%.1f", tps * tpot / 1000)
print line
if (acc != "") print " " acc
}
' "$log"
}
case "${1:-}" in
warmup) shift; cmd_warmup "$@" ;;
prefill) shift; cmd_prefill "$@" ;;
decode) shift; cmd_decode "$@" ;;
l3) shift; cmd_l3 "$@" ;;
report) shift; cmd_report "$@" ;;
*) usage ;;
esac
+61
View File
@@ -0,0 +1,61 @@
cd /data/ymk/ds41
HDR="Authorization=Bearer deepseek-v4-flash-0731-key"
PORT=30040
MODEL=/data/models/DeepSeek-V4.1-Flash
run() {
local name="$1"; shift
echo "===== [$(date +%H:%M:%S)] START: $name ====="
python3 -m sglang.bench_serving "$@" 2>&1 | grep -vE '^Namespace|warn'
echo "===== [$(date +%H:%M:%S)] END(exit=$?): $name ====="
sleep 3
}
r160() { # $1=name $2=seed $3=num_prompts $4=mc
run "$1" --backend sglang --host 127.0.0.1 --port $PORT --model $MODEL \
--dataset-name random-ids --random-input-len 160000 --random-output-len 1 \
--num-prompts "$3" --max-concurrency "$4" --seed "$2" --tokenize-prompt \
--random-range-ratio 1.0 --header "$HDR"
}
# WARMUP
run WARMUP --backend sglang --host 127.0.0.1 --port $PORT --model $MODEL \
--dataset-name random-ids --random-input-len 4096 --random-output-len 16 \
--num-prompts 8 --max-concurrency 8 --seed 777000 --tokenize-prompt \
--random-range-ratio 1.0 --header "$HDR"
# R0 纯计算极限(集合 A 首次出现,L1/L2/L3 全空)
r160 R0_COMPUTE_COLD 1541391 8 8
# R1 L1 极限(A 全量在 L1 设备池:1.28M < 1.6M)
r160 R1_L1_HIT 1541391 8 8
# R2 填充集 B(12×160k=1.92M > L1 1.6M),把 A 挤出 L1 到 L2
r160 R2_FILLER_EVICT 999000 12 8
# R3 L2 极限(A 已被挤出 L1,从 L2 host 池 load-back)
r160 R3_L2_HIT 1541391 8 8
# R4 flush L1+L2(L3 保留),走 L3 prefetch
echo "===== [$(date +%H:%M:%S)] FLUSH L1+L2 ====="
curl -sf -m 10 -X POST http://127.0.0.1:$PORT/flush_cache && echo "flush ok"
sleep 5
r160 R4_L3_HIT 1541391 8 8
# R5 嫁接后 L2 复核
r160 R5_L2_AFTER_GRAFT 1541391 8 8
# R6 混合负载:新集合 C(纯计算,4×160k mc4)+ 集合 A(命中,8×160k mc8)并发打
echo "===== [$(date +%H:%M:%S)] START: R6_MIX_PARALLEL ====="
python3 -m sglang.bench_serving --backend sglang --host 127.0.0.1 --port $PORT --model $MODEL \
--dataset-name random-ids --random-input-len 160000 --random-output-len 1 \
--num-prompts 4 --max-concurrency 4 --seed 555000 --tokenize-prompt \
--random-range-ratio 1.0 --header "$HDR" > /data/ymk/ds41/.tier_r6_coldclient.log 2>&1 &
COLD_PID=$!
python3 -m sglang.bench_serving --backend sglang --host 127.0.0.1 --port $PORT --model $MODEL \
--dataset-name random-ids --random-input-len 160000 --random-output-len 1 \
--num-prompts 8 --max-concurrency 8 --seed 1541391 --tokenize-prompt \
--random-range-ratio 1.0 --header "$HDR" > /data/ymk/ds41/.tier_r6_hitclient.log 2>&1 &
HIT_PID=$!
wait $COLD_PID $HIT_PID
echo "===== [$(date +%H:%M:%S)] END: R6_MIX_PARALLEL ====="
echo "--- R6 cold client (compute) ---"
grep -E 'Input token throughput|Mean TTFT|Request throughput' /data/ymk/ds41/.tier_r6_coldclient.log
echo "--- R6 hit client (cache) ---"
grep -E 'Input token throughput|Mean TTFT|Request throughput' /data/ymk/ds41/.tier_r6_hitclient.log
echo "===== [$(date +%H:%M:%S)] ALL DONE ====="
+27
View File
@@ -0,0 +1,27 @@
cd /data/ymk/ds41
HDR="Authorization=Bearer deepseek-v4-flash-0731-key"
PORT=30040
MODEL=/data/models/DeepSeek-V4.1-Flash
echo "===== [$(date +%H:%M:%S)] FLUSH L1+L2(A 只剩 L3) ====="
curl -sf -m 10 -X POST http://127.0.0.1:$PORT/flush_cache && echo "flush ok"
sleep 5
echo "===== [$(date +%H:%M:%S)] START: R7_MIX_L3_PLUS_COMPUTE ====="
python3 -m sglang.bench_serving --backend sglang --host 127.0.0.1 --port $PORT --model $MODEL \
--dataset-name random-ids --random-input-len 160000 --random-output-len 1 \
--num-prompts 4 --max-concurrency 4 --seed 666000 --tokenize-prompt \
--random-range-ratio 1.0 --header "$HDR" > /data/ymk/ds41/.tier_r7_coldclient.log 2>&1 &
COLD_PID=$!
python3 -m sglang.bench_serving --backend sglang --host 127.0.0.1 --port $PORT --model $MODEL \
--dataset-name random-ids --random-input-len 160000 --random-output-len 1 \
--num-prompts 8 --max-concurrency 8 --seed 1541391 --tokenize-prompt \
--random-range-ratio 1.0 --header "$HDR" > /data/ymk/ds41/.tier_r7_hitclient.log 2>&1 &
HIT_PID=$!
wait $COLD_PID $HIT_PID
echo "===== [$(date +%H:%M:%S)] END: R7 ====="
echo "--- R7 cold client (compute, 4x160k mc4) ---"
grep -E 'Input token throughput|Mean TTFT|Request throughput' /data/ymk/ds41/.tier_r7_coldclient.log
echo "--- R7 hit client (L3, 8x160k mc8) ---"
grep -E 'Input token throughput|Mean TTFT|Request throughput' /data/ymk/ds41/.tier_r7_hitclient.log
echo "===== [$(date +%H:%M:%S)] ALL DONE ====="
+1
View File
@@ -28,6 +28,7 @@ x-p-environment: &p-environment
MC_INTRA_NVLINK: "true"
SGLANG_MOONCAKE_SEND_AUX_TCP: "1"
SGLANG_HICACHE_MOONCAKE_CONFIG_PATH: /data/ymk/ds41/mooncake-store.json
SGLANG_HICACHE_PREFETCH_IO_THREADS: "4"
SGLANG_DISAGGREGATION_QUEUE_SIZE: "16"
SGLANG_DISAGGREGATION_THREAD_POOL_SIZE: "32"
@@ -0,0 +1,225 @@
# 【L3 dummy client 变体 2026-10-09 深夜,定型=shm+best_effort】sglang→standalone_storage,数据面经本机 mooncake_client(50052 IPC+进程映射),不碰 TCP 传输引擎
# 实测 24x160k 热/冷:shm+best_effort +26.4%(唯一正收益);wait_complete 因 all-or-nothing discard 等待白等反而 -17.7%。TCP 传输引擎两组皆负。
# 【L3 cache-hit 测试变体 2026-10-09】基于 dockerserve-3p1d-otlp.yml,仅改 P:
# --max-total-tokens 800000(L1 缩到 ~0.8M token)+ --hicache-ratio 1.0(L2=L1=0.8M token。⚠️ 教训:ratio 0.05 时 L2 太小,build_storage_backup_spec 拿不到 host 页会静默跳过 L3 写入,热轮无数据可中;V4 hicache 不支持 --hicache-size)
# 使 ~2.6M token/P 的语料(48×160k)溢出 L1+L2(1.6M),热轮强制走 L3 mooncake 命中。
# 【2026-10-10 L2 放大】hicache-ratio 1.0→4.0(L2=3.2M token/P,宿主可用内存 ~1.9TB 充裕):
# #38348 fail-soft 移植后 bench 显示收益被容量掩盖——24x160k 下 host_capacity 拒绝
# 0.75M~2.97M token/P(L2 装不下,prefetch 未发起即被拒退回重算)。放大后验证:
# ①热轮请求 L3/L2 命中不被容量拒绝;②L3 hit 嫁接进 L2 树后在 L2 长期留存。
# 与主模板同容器名/端口,用前必须先 down 主栈。D 与 router 不变。
# 原注释: 3x cp2-P + 1x dp2-D(cg256) PD 分离 + L3 mooncake + engram + OTLP trace(2026-09-27)
# 基于 dockerserve-3p1d-soak.yml 增加: --enable-trace --trace-modules request,mooncake
# --otlp-traces-endpoint 127.0.0.1:4317 --otlp-service-name ds41-3p1d-<实例名>
# 依赖: b300-01 k8s cw jaeger-trace(jaegertracing/all-in-one:1.76.0,badger 6h TTL,hostNetwork 4317)
# 安全性: BatchSpanProcessor 导出失败只丢 span;Jaeger 宕机不影响推理服务
# 调优旋钮(默认未开): SGLANG_TRACE_ASYNC=1(异步 span 创建), SGLANG_TRACE_LEVEL(默认 3)
# router(ds41-router-3p1d,rust)无 OTLP 支持,不产生 span,追踪从各 sglang HTTP 层开始
# 2026-09-23 实验;基线见 archive-20260923/ 与 D:\B300\experiments\3cp2-pd\baseline-20260923.md
# GPU: P1=0,1 P2=2,3 P3=4,5 D=6,7
# 端口: P1=30000 P2=30010 P3=30020 D=30001 router=30002
# bootstrap: P1=8998 P2=8999 P3=9000 D=9001
x-sglang-common: &sglang-common
image: ymkymx/sglang:dsv41-pd-visioncp-baba18bc7d-local-20261010-2310
shm_size: "32gb"
ipc: host
pid: host
privileged: true
network_mode: host
volumes:
- /data:/data
- /data/ymk/cache/sglang:/root/.cache/sglang
- /data/ymk/sglang/python:/sgl-workspace/sglang/python
- /data/ymk/ds41/fake-cargo:/root/.cargo/bin/cargo:ro
deploy:
resources:
reservations:
devices:
- driver: nvidia
count: all
capabilities: [gpu]
x-p-environment: &p-environment
SGLANG_TORCH_PROFILER_DIR: /data/ymk/prof/torch
RUSTUP_TOOLCHAIN: "1.92-x86_64-unknown-linux-gnu"
SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE: "1"
SGLANG_RAGGED_VERIFY_MODE: static
MC_INTRANODE_NVLINK: "true"
MC_INTRA_NVLINK: "true"
SGLANG_MOONCAKE_SEND_AUX_TCP: "1"
SGLANG_HICACHE_MOONCAKE_CONFIG_PATH: /data/ymk/ds41/mooncake-store-dummy.json
SGLANG_HICACHE_PREFETCH_IO_THREADS: "4"
SGLANG_DISAGGREGATION_QUEUE_SIZE: "16"
SGLANG_DISAGGREGATION_THREAD_POOL_SIZE: "32"
services:
p1:
<<: *sglang-common
container_name: ds41-3p1d-p1
environment:
<<: *p-environment
CUDA_VISIBLE_DEVICES: "0,1"
command: >
sglang serve
--trust-remote-code
--model-path /data/models/DeepSeek-V4.1-Flash
--tp 2 --ep-size 2
--enable-prefill-cp --cp-strategy interleave
--mem-fraction-static 0.75
--attention-backend dsv4
--moe-runner-backend flashinfer_mxfp4
--cuda-graph-max-bs-decode 32
--reasoning-parser auto --tool-call-parser auto
--default-chat-template-kwargs '{"thinking": true, "reasoning_effort": "high"}'
--max-running-requests 256
--tokenizer-worker-num 8
--host 0.0.0.0 --port 30000
--enable-cache-report --enable-metrics
--speculative-algorithm DSPARK --speculative-dspark-block-size 5
--enable-hierarchical-cache
--hicache-ratio 4.0
--hicache-mem-layout page_first_direct
--hicache-io-backend direct
--hicache-write-policy write_through
--hicache-storage-backend mooncake
--hicache-storage-prefetch-policy best_effort
--hicache-size 0
--max-total-tokens 800000
--disaggregation-mode prefill
--disaggregation-transfer-backend mooncake
--disaggregation-bootstrap-port 8998
--enable-trace
--trace-modules request,mooncake
--otlp-traces-endpoint 127.0.0.1:4317
--otlp-service-name ds41-3p1d-p1
p2:
<<: *sglang-common
container_name: ds41-3p1d-p2
environment:
<<: *p-environment
CUDA_VISIBLE_DEVICES: "2,3"
command: >
sglang serve
--trust-remote-code
--model-path /data/models/DeepSeek-V4.1-Flash
--tp 2 --ep-size 2
--enable-prefill-cp --cp-strategy interleave
--mem-fraction-static 0.75
--attention-backend dsv4
--moe-runner-backend flashinfer_mxfp4
--cuda-graph-max-bs-decode 32
--reasoning-parser auto --tool-call-parser auto
--default-chat-template-kwargs '{"thinking": true, "reasoning_effort": "high"}'
--max-running-requests 256
--tokenizer-worker-num 8
--host 0.0.0.0 --port 30010
--enable-cache-report --enable-metrics
--speculative-algorithm DSPARK --speculative-dspark-block-size 5
--enable-hierarchical-cache
--hicache-ratio 4.0
--hicache-mem-layout page_first_direct
--hicache-io-backend direct
--hicache-write-policy write_through
--hicache-storage-backend mooncake
--hicache-storage-prefetch-policy best_effort
--hicache-size 0
--max-total-tokens 800000
--disaggregation-mode prefill
--disaggregation-transfer-backend mooncake
--disaggregation-bootstrap-port 8999
--enable-trace
--trace-modules request,mooncake
--otlp-traces-endpoint 127.0.0.1:4317
--otlp-service-name ds41-3p1d-p2
p3:
<<: *sglang-common
container_name: ds41-3p1d-p3
environment:
<<: *p-environment
CUDA_VISIBLE_DEVICES: "4,5"
command: >
sglang serve
--trust-remote-code
--model-path /data/models/DeepSeek-V4.1-Flash
--tp 2 --ep-size 2
--enable-prefill-cp --cp-strategy interleave
--mem-fraction-static 0.75
--attention-backend dsv4
--moe-runner-backend flashinfer_mxfp4
--cuda-graph-max-bs-decode 32
--reasoning-parser auto --tool-call-parser auto
--default-chat-template-kwargs '{"thinking": true, "reasoning_effort": "high"}'
--max-running-requests 256
--tokenizer-worker-num 8
--host 0.0.0.0 --port 30020
--enable-cache-report --enable-metrics
--speculative-algorithm DSPARK --speculative-dspark-block-size 5
--enable-hierarchical-cache
--hicache-ratio 4.0
--hicache-mem-layout page_first_direct
--hicache-io-backend direct
--hicache-write-policy write_through
--hicache-storage-backend mooncake
--hicache-storage-prefetch-policy best_effort
--hicache-size 0
--max-total-tokens 800000
--disaggregation-mode prefill
--disaggregation-transfer-backend mooncake
--disaggregation-bootstrap-port 9000
--enable-trace
--trace-modules request,mooncake
--otlp-traces-endpoint 127.0.0.1:4317
--otlp-service-name ds41-3p1d-p3
d:
<<: *sglang-common
# G9 P0 验证(2026-10-11):D 单独升镜像拿 /query_cache_hit 修复(71c6d8bc5),
# 并开 decode radix + hicache L3 客户端(复用 dummy store,只读查询)。
image: ymkymx/sglang:dsv41-pd-visioncp-6612af9e0-local-20261011-0740
container_name: ds41-3p1d-d1
environment:
SGLANG_TORCH_PROFILER_DIR: /data/ymk/prof/torch
RUSTUP_TOOLCHAIN: "1.92-x86_64-unknown-linux-gnu"
SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE: "1"
SGLANG_RAGGED_VERIFY_MODE: static
MC_INTRANODE_NVLINK: "true"
MC_INTRA_NVLINK: "true"
SGLANG_MOONCAKE_SEND_AUX_TCP: "1"
SGLANG_DISAGGREGATION_QUEUE_SIZE: "16"
SGLANG_DISAGGREGATION_THREAD_POOL_SIZE: "32"
SGLANG_HICACHE_MOONCAKE_CONFIG_PATH: /data/ymk/ds41/mooncake-store-dummy.json
SGLANG_HICACHE_PREFETCH_IO_THREADS: "4"
CUDA_VISIBLE_DEVICES: "6,7"
command: >
sglang serve
--trust-remote-code
--model-path /data/models/DeepSeek-V4.1-Flash
--tp 2 --ep-size 2 --dp-size 2
--enable-dp-attention --enable-dp-lm-head
--mem-fraction-static 0.80
--attention-backend dsv4
--moe-runner-backend flashinfer_mxfp4
--cuda-graph-max-bs-decode 256
--reasoning-parser auto --tool-call-parser auto
--default-chat-template-kwargs '{"thinking": true, "reasoning_effort": "high"}'
--max-running-requests 256
--tokenizer-worker-num 8
--host 0.0.0.0 --port 30001
--enable-cache-report --enable-metrics
--speculative-algorithm DSPARK --speculative-dspark-block-size 5
--disaggregation-mode decode
--disaggregation-transfer-backend mooncake
--disaggregation-bootstrap-port 9001
--disaggregation-decode-enable-radix-cache
--enable-hierarchical-cache
--hicache-ratio 1.0
--hicache-mem-layout page_first_direct
--hicache-io-backend direct
--hicache-storage-backend mooncake
--hicache-storage-prefetch-policy best_effort
--enable-trace
--trace-modules request,mooncake
--otlp-traces-endpoint 127.0.0.1:4317
--otlp-service-name ds41-3p1d-d1
@@ -0,0 +1,239 @@
# 【G9 P0 验证栈:真实 store 3p1d 2026-10-11】基于 dockerserve-3p1d-otlp.yml。
# 用途:验证 P0 改动(/v1/loads 排队 token + drain rate 字段、P 侧 /query_cache_hit)。
# 与主模板同容器名/端口,用前必须先 down 主栈;router(ds41-router-3p1d)独立不在本文件。
# 2026-10-11 改造要点(vs 原 l3test):
# - 镜像升 6612af9e0(含 G9 P0 + /query_cache_hit KeyError 修复)+ scheduler.py
# 热补丁(L3 查询门限从 enable_decode_hicache 放宽到 enable_hicache_storage,
# 否则 P 节点永远查不到 L3)。
# - 每实例独立 MOONCAKE_LOCAL_HOSTNAME(a/b/c/d),extra_hosts 互配全部 rank 别名
# (真实 store 跨实例副本写要求对端 rank 主机名可解析,见 realstore-x2 注释;
# rank 命名规则 {hostname}-d{dp}t{tp}p{pp}c{attn_cp},commit 84fc632088;
# D 为 dp2 dp-attention(attn_tp=1),后缀推断为 -d0t0p0c0/-d1t1p0c0,未实证)。
# - ⚠️ D 侧 cache 方案实证否决(2026-10-11):SWA 模型下
# --disaggregation-decode-enable-radix-cache 先要 SGLANG_ENABLE_UNIFIED_RADIX_TREE=1,
# 开了又报 "supports only device-resident cache and is incompatible with
# --enable-hierarchical-cache"(kv_cache_builder.py 硬 ValueError)——D 挂不了 L3。
# 结论:/query_cache_hit 放 P(L3 全局统一,P 有完整 hicache),D 保持生产配置。
# - D mem-fraction 0.80:曾因崩溃 d1 泄漏 ~61GiB 临时降 0.75,全栈 down 掉旧容器后
# 驱动即回收(GPU 全零),已恢复 0.80。D max-running 保持 256(512 时 SWA 池挤爆 KV 池)。
# P 侧保持 l3test 原状:--max-total-tokens 800000 + ratio 1.0(L1/L2 各 ~0.8M token),
# write_through 保证 prefill 后 L3 有数据,供 /query_cache_hit 对拍。
# 原注释: 3x cp2-P + 1x dp2-D(cg256) PD 分离 + L3 mooncake + engram + OTLP trace(2026-09-27)
# 依赖: mooncake-master 容器(50051,起栈前 restart 清旧 segment 注册);
# b300-01 k8s cw jaeger-trace(4317,导出失败只丢 span 不影响服务)
# GPU: P1=0,1 P2=2,3 P3=4,5 D=6,7
# 端口: P1=30000 P2=30010 P3=30020 D=30001 router=30002
# bootstrap: P1=8998 P2=8999 P3=9000 D=9001
x-sglang-common: &sglang-common
image: ymkymx/sglang:dsv41-pd-visioncp-e7dc175e9-local-20261011-0940
shm_size: "32gb"
ipc: host
pid: host
privileged: true
network_mode: host
volumes:
- /data:/data
- /data/ymk/cache/sglang:/root/.cache/sglang
deploy:
resources:
reservations:
devices:
- driver: nvidia
count: all
capabilities: [gpu]
x-rank-hosts: &rank-hosts
# mooncake rank 主机名别名:P(tp2/cp2 interleave)-d0t0p0c0/-d0t1p0c1;
# D(dp2 dp-attention)-d0t0p0c0/-d1t1p0c0(推断,起栈后核对 master 注册名)
- "localhost-a-d0t0p0c0:127.0.0.1"
- "localhost-a-d0t1p0c1:127.0.0.1"
- "localhost-b-d0t0p0c0:127.0.0.1"
- "localhost-b-d0t1p0c1:127.0.0.1"
- "localhost-c-d0t0p0c0:127.0.0.1"
- "localhost-c-d0t1p0c1:127.0.0.1"
- "localhost-d-d0t0p0c0:127.0.0.1"
- "localhost-d-d1t1p0c0:127.0.0.1"
x-p-environment: &p-environment
SGLANG_TORCH_PROFILER_DIR: /data/ymk/prof/torch
SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE: "1"
SGLANG_RAGGED_VERIFY_MODE: static
MC_INTRANODE_NVLINK: "true"
MC_INTRA_NVLINK: "true"
SGLANG_MOONCAKE_SEND_AUX_TCP: "1"
SGLANG_HICACHE_MOONCAKE_CONFIG_PATH: /data/ymk/ds41/mooncake-store.json
SGLANG_HICACHE_PREFETCH_IO_THREADS: "4"
SGLANG_DISAGGREGATION_QUEUE_SIZE: "16"
SGLANG_DISAGGREGATION_THREAD_POOL_SIZE: "32"
services:
p1:
<<: *sglang-common
container_name: ds41-3p1d-p1
extra_hosts: *rank-hosts
environment:
<<: *p-environment
MOONCAKE_LOCAL_HOSTNAME: localhost-a
CUDA_VISIBLE_DEVICES: "0,1"
command: >
sglang serve
--trust-remote-code
--model-path /data/models/DeepSeek-V4.1-Flash
--tp 2 --ep-size 2
--enable-prefill-cp --cp-strategy interleave
--mem-fraction-static 0.75
--attention-backend dsv4
--moe-runner-backend flashinfer_mxfp4
--cuda-graph-max-bs-decode 32
--reasoning-parser auto --tool-call-parser auto
--default-chat-template-kwargs '{"thinking": true, "reasoning_effort": "high"}'
--max-running-requests 256
--tokenizer-worker-num 8
--host 0.0.0.0 --port 30000
--enable-cache-report --enable-metrics
--speculative-algorithm DSPARK --speculative-dspark-block-size 5
--enable-hierarchical-cache
--hicache-ratio 1.0
--hicache-mem-layout page_first_direct
--hicache-io-backend direct
--hicache-write-policy write_through
--hicache-storage-backend mooncake
--hicache-storage-prefetch-policy best_effort
--hicache-size 0
--max-total-tokens 800000
--disaggregation-mode prefill
--disaggregation-transfer-backend mooncake
--disaggregation-bootstrap-port 8998
--enable-trace
--trace-modules request,mooncake
--otlp-traces-endpoint 127.0.0.1:4317
--otlp-service-name ds41-3p1d-p1
p2:
<<: *sglang-common
container_name: ds41-3p1d-p2
extra_hosts: *rank-hosts
environment:
<<: *p-environment
MOONCAKE_LOCAL_HOSTNAME: localhost-b
CUDA_VISIBLE_DEVICES: "2,3"
command: >
sglang serve
--trust-remote-code
--model-path /data/models/DeepSeek-V4.1-Flash
--tp 2 --ep-size 2
--enable-prefill-cp --cp-strategy interleave
--mem-fraction-static 0.75
--attention-backend dsv4
--moe-runner-backend flashinfer_mxfp4
--cuda-graph-max-bs-decode 32
--reasoning-parser auto --tool-call-parser auto
--default-chat-template-kwargs '{"thinking": true, "reasoning_effort": "high"}'
--max-running-requests 256
--tokenizer-worker-num 8
--host 0.0.0.0 --port 30010
--enable-cache-report --enable-metrics
--speculative-algorithm DSPARK --speculative-dspark-block-size 5
--enable-hierarchical-cache
--hicache-ratio 1.0
--hicache-mem-layout page_first_direct
--hicache-io-backend direct
--hicache-write-policy write_through
--hicache-storage-backend mooncake
--hicache-storage-prefetch-policy best_effort
--hicache-size 0
--max-total-tokens 800000
--disaggregation-mode prefill
--disaggregation-transfer-backend mooncake
--disaggregation-bootstrap-port 8999
--enable-trace
--trace-modules request,mooncake
--otlp-traces-endpoint 127.0.0.1:4317
--otlp-service-name ds41-3p1d-p2
p3:
<<: *sglang-common
container_name: ds41-3p1d-p3
extra_hosts: *rank-hosts
environment:
<<: *p-environment
MOONCAKE_LOCAL_HOSTNAME: localhost-c
CUDA_VISIBLE_DEVICES: "4,5"
command: >
sglang serve
--trust-remote-code
--model-path /data/models/DeepSeek-V4.1-Flash
--tp 2 --ep-size 2
--enable-prefill-cp --cp-strategy interleave
--mem-fraction-static 0.75
--attention-backend dsv4
--moe-runner-backend flashinfer_mxfp4
--cuda-graph-max-bs-decode 32
--reasoning-parser auto --tool-call-parser auto
--default-chat-template-kwargs '{"thinking": true, "reasoning_effort": "high"}'
--max-running-requests 256
--tokenizer-worker-num 8
--host 0.0.0.0 --port 30020
--enable-cache-report --enable-metrics
--speculative-algorithm DSPARK --speculative-dspark-block-size 5
--enable-hierarchical-cache
--hicache-ratio 1.0
--hicache-mem-layout page_first_direct
--hicache-io-backend direct
--hicache-write-policy write_through
--hicache-storage-backend mooncake
--hicache-storage-prefetch-policy best_effort
--hicache-size 0
--max-total-tokens 800000
--disaggregation-mode prefill
--disaggregation-transfer-backend mooncake
--disaggregation-bootstrap-port 9000
--enable-trace
--trace-modules request,mooncake
--otlp-traces-endpoint 127.0.0.1:4317
--otlp-service-name ds41-3p1d-p3
d:
<<: *sglang-common
container_name: ds41-3p1d-d1
extra_hosts: *rank-hosts
# ⚠️ D 保持生产配置(2026-10-11 实证):SWA 模型下
# --disaggregation-decode-enable-radix-cache 与 --enable-hierarchical-cache 互斥
# (kv_cache_builder.py 硬 ValueError,"supports only device-resident cache"),
# D 无法挂 L3 —— /query_cache_hit 改由 P 节点回答(L3 全局统一,P 有完整 hicache)。
# MOONCAKE_LOCAL_HOSTNAME 保留备用(无害)。
environment:
SGLANG_TORCH_PROFILER_DIR: /data/ymk/prof/torch
SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE: "1"
SGLANG_RAGGED_VERIFY_MODE: static
MC_INTRANODE_NVLINK: "true"
MC_INTRA_NVLINK: "true"
SGLANG_MOONCAKE_SEND_AUX_TCP: "1"
SGLANG_DISAGGREGATION_QUEUE_SIZE: "16"
SGLANG_DISAGGREGATION_THREAD_POOL_SIZE: "32"
MOONCAKE_LOCAL_HOSTNAME: localhost-d
CUDA_VISIBLE_DEVICES: "6,7"
command: >
sglang serve
--trust-remote-code
--model-path /data/models/DeepSeek-V4.1-Flash
--tp 2 --ep-size 2 --dp-size 2
--enable-dp-attention --enable-dp-lm-head
--mem-fraction-static 0.80
--attention-backend dsv4
--moe-runner-backend flashinfer_mxfp4
--cuda-graph-max-bs-decode 256
--reasoning-parser auto --tool-call-parser auto
--default-chat-template-kwargs '{"thinking": true, "reasoning_effort": "high"}'
--max-running-requests 256
--tokenizer-worker-num 8
--host 0.0.0.0 --port 30001
--enable-cache-report --enable-metrics
--speculative-algorithm DSPARK --speculative-dspark-block-size 5
--disaggregation-mode decode
--disaggregation-transfer-backend mooncake
--disaggregation-bootstrap-port 9001
--enable-trace
--trace-modules request,mooncake
--otlp-traces-endpoint 127.0.0.1:4317
--otlp-service-name ds41-3p1d-d1
@@ -35,6 +35,7 @@ x-p-environment: &p-environment
MC_INTRA_NVLINK: "true"
SGLANG_MOONCAKE_SEND_AUX_TCP: "1"
SGLANG_HICACHE_MOONCAKE_CONFIG_PATH: /data/ymk/ds41/mooncake-store.json
SGLANG_HICACHE_PREFETCH_IO_THREADS: "4"
SGLANG_DISAGGREGATION_QUEUE_SIZE: "16"
SGLANG_DISAGGREGATION_THREAD_POOL_SIZE: "32"
+2 -1
View File
@@ -10,7 +10,7 @@
# 端口: P1=30000 P2=30010 P3=30020 D=30001 router=30002
# bootstrap: P1=8998 P2=8999 P3=9000 D=9001
x-sglang-common: &sglang-common
image: ymkymx/sglang:dsv41-pd-visioncp-db7d2cb7d-fix
image: ymkymx/sglang:dsv41-pd-visioncp-1068926cd-local-20261008-0831
shm_size: "32gb"
ipc: host
pid: host
@@ -35,6 +35,7 @@ x-p-environment: &p-environment
MC_INTRA_NVLINK: "true"
SGLANG_MOONCAKE_SEND_AUX_TCP: "1"
SGLANG_HICACHE_MOONCAKE_CONFIG_PATH: /data/ymk/ds41/mooncake-store.json
SGLANG_HICACHE_PREFETCH_IO_THREADS: "4"
SGLANG_DISAGGREGATION_QUEUE_SIZE: "16"
SGLANG_DISAGGREGATION_THREAD_POOL_SIZE: "32"
@@ -29,6 +29,7 @@ x-p-environment: &p-environment
MC_INTRA_NVLINK: "true"
SGLANG_MOONCAKE_SEND_AUX_TCP: "1"
SGLANG_HICACHE_MOONCAKE_CONFIG_PATH: /data/ymk/ds41/mooncake-store.json
SGLANG_HICACHE_PREFETCH_IO_THREADS: "4"
SGLANG_DISAGGREGATION_QUEUE_SIZE: "16"
SGLANG_DISAGGREGATION_THREAD_POOL_SIZE: "32"
@@ -29,6 +29,7 @@ x-p-environment: &p-environment
MC_INTRA_NVLINK: "true"
SGLANG_MOONCAKE_SEND_AUX_TCP: "1"
SGLANG_HICACHE_MOONCAKE_CONFIG_PATH: /data/ymk/ds41/mooncake-store.json
SGLANG_HICACHE_PREFETCH_IO_THREADS: "4"
SGLANG_DISAGGREGATION_QUEUE_SIZE: "16"
SGLANG_DISAGGREGATION_THREAD_POOL_SIZE: "32"
+1 -1
View File
@@ -1,6 +1,6 @@
services:
sglang:
image: ymkymx/sglang:main-ee5fcdf0d-20260920-0142
image: ymkymx/sglang:dsv41-pd-visioncp-1068926cd-local-20261008-0831
container_name: ds41-sglang-b4
shm_size: "32gb"
ipc: host
+1 -1
View File
@@ -1,6 +1,6 @@
services:
sglang:
image: ymkymx/sglang:dsv41-pd-92632a60b-20260922-2100
image: ymkymx/sglang:dsv41-pd-visioncp-1068926cd-local-20261008-0831
container_name: ds41-cpdspark-nr
shm_size: "32gb"
ipc: host
@@ -0,0 +1,49 @@
# 单独 D 节点基线(非 PD role):3p1d 的 d1 配置去掉 disaggregation,直接对外服务
# 用途:测单 D 节点 decode 基线(阶梯并发 32..80%*cuda-graph-max-bs-decode,
# 报 TPOT / 实际并发 / 整实例输出 TPS;prefill 走 L1 radix cache-hit)
# GPU 6,7;端口 30001;无 hicache 依赖(L1 即可支撑重复 prompt 命中)
# 2026-10-09 建:配合 bench_suite_v2.sh decode
x-sglang-common: &sglang-common
image: ymkymx/sglang:dsv41-pd-visioncp-1068926cd-local-20261008-0831
shm_size: "32gb"
ipc: host
pid: host
privileged: true
network_mode: host
volumes:
- /data:/data
- /data/ymk/cache/sglang:/root/.cache/sglang
deploy:
resources:
reservations:
devices:
- driver: nvidia
count: all
capabilities: [gpu]
services:
d1s:
<<: *sglang-common
container_name: ds41-d1-standalone
environment:
SGLANG_TORCH_PROFILER_DIR: /data/ymk/prof/torch
SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE: "1"
SGLANG_RAGGED_VERIFY_MODE: static
CUDA_VISIBLE_DEVICES: "6,7"
command: >
sglang serve
--trust-remote-code
--model-path /data/models/DeepSeek-V4.1-Flash
--tp 2 --ep-size 2 --dp-size 2
--enable-dp-attention --enable-dp-lm-head
--mem-fraction-static 0.80
--attention-backend dsv4
--moe-runner-backend flashinfer_mxfp4
--cuda-graph-max-bs-decode 256
--reasoning-parser auto --tool-call-parser auto
--default-chat-template-kwargs '{"thinking": true, "reasoning_effort": "high"}'
--max-running-requests 256
--tokenizer-worker-num 8
--host 0.0.0.0 --port 30001
--enable-cache-report --enable-metrics
--speculative-algorithm DSPARK --speculative-dspark-block-size 5
@@ -0,0 +1,61 @@
# 单独 P 节点基线(非 PD role):3p1d 的 p1 配置去掉 disaggregation,直接对外服务
# 用途:测单 P 节点 prefill 基线吞吐(冷)+ hicache cache-hit TPS(热)
# GPU 0,1;端口 30000;依赖 mooncake-master(hicache L3,127.0.0.1:50051)
# 2026-10-09 建:配合 bench_suite_v2.sh prefill
x-sglang-common: &sglang-common
image: ymkymx/sglang:dsv41-pd-visioncp-1068926cd-local-20261008-0831
shm_size: "32gb"
ipc: host
pid: host
privileged: true
network_mode: host
volumes:
- /data:/data
- /data/ymk/cache/sglang:/root/.cache/sglang
deploy:
resources:
reservations:
devices:
- driver: nvidia
count: all
capabilities: [gpu]
services:
p1s:
<<: *sglang-common
container_name: ds41-p1-standalone
environment:
SGLANG_TORCH_PROFILER_DIR: /data/ymk/prof/torch
SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE: "1"
SGLANG_RAGGED_VERIFY_MODE: static
MC_INTRANODE_NVLINK: "true"
MC_INTRA_NVLINK: "true"
SGLANG_MOONCAKE_SEND_AUX_TCP: "1"
SGLANG_HICACHE_MOONCAKE_CONFIG_PATH: /data/ymk/ds41/mooncake-store.json
SGLANG_HICACHE_PREFETCH_IO_THREADS: "4"
CUDA_VISIBLE_DEVICES: "0,1"
command: >
sglang serve
--trust-remote-code
--model-path /data/models/DeepSeek-V4.1-Flash
--tp 2 --ep-size 2
--enable-prefill-cp --cp-strategy interleave
--mem-fraction-static 0.75
--attention-backend dsv4
--moe-runner-backend flashinfer_mxfp4
--cuda-graph-max-bs-decode 32
--reasoning-parser auto --tool-call-parser auto
--default-chat-template-kwargs '{"thinking": true, "reasoning_effort": "high"}'
--max-running-requests 256
--tokenizer-worker-num 8
--host 0.0.0.0 --port 30000
--enable-cache-report --enable-metrics
--speculative-algorithm DSPARK --speculative-dspark-block-size 5
--enable-hierarchical-cache
--hicache-ratio 2
--hicache-mem-layout page_first_direct
--hicache-io-backend direct
--hicache-write-policy write_through
--hicache-storage-backend mooncake
--hicache-storage-prefetch-policy wait_complete
--hicache-size 0
@@ -22,6 +22,7 @@ services:
- SGLANG_MOONCAKE_SEND_AUX_TCP=1
- SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE=1
- SGLANG_HICACHE_MOONCAKE_CONFIG_PATH=/data/ymk/ds41/mooncake-store.json
- SGLANG_HICACHE_PREFETCH_IO_THREADS=4
volumes:
- /data:/data
- /data/ymk/cache/sglang:/root/.cache/sglang
@@ -18,6 +18,7 @@ services:
- SGLANG_MOONCAKE_SEND_AUX_TCP=1
- SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE=1
- SGLANG_HICACHE_MOONCAKE_CONFIG_PATH=/data/ymk/ds41/mooncake-store.json
- SGLANG_HICACHE_PREFETCH_IO_THREADS=4
- SGLANG_DISAGGREGATION_QUEUE_SIZE=16
- SGLANG_DISAGGREGATION_THREAD_POOL_SIZE=32
volumes:
@@ -0,0 +1,76 @@
# 【跨 P 实例 L3 读写验证 2026-10-10】dockerserve-tier-cp2-realstore.yml 的第二实例:
# 与 ds41-tier-t2(A:GPU 0,1,端口 30040)同时跑,共用 mooncake master 50051。
# 关键差异:
# - GPU 2,3,端口 30041,容器名 ds41-tier-t3
# - MOONCAKE_LOCAL_HOSTNAME=localhost-b:mooncake 客户端 hostname 前缀按实例区分
# (rank 后缀 -d0t0p0c0/-d0t1p0c1 由 84fc632088 拼在后面的),否则两实例 rank 同名
# → master preferred-segment 分不出段归属 → 副本落错进程段 → 跨进程回退 TCP(实例内)
# 且跨实例段地址冲突。extra_hosts 别名同步加 -b 前缀(P2P handshake 要求可解析,
# ac11e9162d)。
# 验证路径:A prime 写 L3 → flush B 的 L1/L2 → B 读同 seed prompts = 跨实例 TCP P2P 命中。
x-common: &common
image: ymkymx/sglang:dsv41-pd-visioncp-1068926cd-local-20261008-0831
shm_size: 32gb
ipc: host
pid: host
privileged: true
network_mode: host
volumes:
- /data:/data
- /data/ymk/cache/sglang:/root/.cache/sglang
- /data/ymk/sglang/python:/sgl-workspace/sglang/python
- /data/ymk/ds41/fake-cargo:/root/.cargo/bin/cargo:ro
deploy:
resources:
reservations:
devices:
- driver: nvidia
count: all
capabilities: [gpu]
services:
t3:
<<: *common
container_name: ds41-tier-t3
extra_hosts:
- "localhost-b-d0t0p0c0:127.0.0.1"
- "localhost-b-d0t1p0c1:127.0.0.1"
# 跨实例 P2P:对端实例(realstore compose 的 t2)rank 主机名也要能解析
- "localhost-d0t0p0c0:127.0.0.1"
- "localhost-d0t1p0c1:127.0.0.1"
environment:
SGLANG_TORCH_PROFILER_DIR: /data/ymk/prof/torch
SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE: 1
SGLANG_RAGGED_VERIFY_MODE: static
SGLANG_HICACHE_MOONCAKE_CONFIG_PATH: /data/ymk/ds41/mooncake-store.json
SGLANG_HICACHE_TIMING: 1
MOONCAKE_LOCAL_HOSTNAME: localhost-b
SGLANG_HICACHE_PREFETCH_IO_THREADS: 4
CUDA_VISIBLE_DEVICES: 2,3
command: >
sglang serve
--trust-remote-code
--model-path /data/models/DeepSeek-V4.1-Flash
--tp 2 --ep-size 2
--enable-prefill-cp --cp-strategy interleave
--mem-fraction-static 0.75
--attention-backend dsv4
--moe-runner-backend flashinfer_mxfp4
--cuda-graph-max-bs-decode 32
--reasoning-parser auto --tool-call-parser auto
--default-chat-template-kwargs '{"thinking": true, "reasoning_effort": "high"}'
--max-running-requests 256
--chunked-prefill-size 32768
--tokenizer-worker-num 8
--host 0.0.0.0 --port 30041
--enable-cache-report --enable-metrics
--speculative-algorithm DSPARK --speculative-dspark-block-size 5
--enable-hierarchical-cache
--hicache-ratio 4.0
--hicache-mem-layout page_first_direct
--hicache-io-backend direct
--hicache-write-policy write_through
--hicache-storage-backend mooncake
--hicache-storage-prefetch-policy wait_complete
--hicache-size 0
--max-total-tokens 1600000
@@ -0,0 +1,73 @@
# 【L3 真实 store A/B 2026-10-10】同 dockerserve-tier-cp2.yml,唯一差异:
# L3 从 dummy standalone(mooncake-store-dummy.json,50052 守护进程)
# 切到真实 mooncake store(mooncake-store.json:master 50051 + P2PHANDSHAKE + tcp,
# 客户端注册 300GB 自身内存段)——与 k8s 线上 ds41-3p1d 的 L3 模式一致。
# 目的:区分 "sglang L3 路径本身慢" 还是 "dummy 后端 per-key 路径慢"(mc 扫描 A/B)。
x-common: &common
image: ymkymx/sglang:dsv41-pd-visioncp-1068926cd-local-20261008-0831
shm_size: 32gb
ipc: host
pid: host
privileged: true
network_mode: host
volumes:
- /data:/data
- /data/ymk/cache/sglang:/root/.cache/sglang
- /data/ymk/sglang/python:/sgl-workspace/sglang/python
- /data/ymk/ds41/fake-cargo:/root/.cargo/bin/cargo:ro
deploy:
resources:
reservations:
devices:
- driver: nvidia
count: all
capabilities: [gpu]
services:
t2:
<<: *common
container_name: ds41-tier-t2
extra_hosts:
- "localhost-d0t0p0c0:127.0.0.1"
- "localhost-d0t1p0c1:127.0.0.1"
# 跨实例 P2P:对端实例(x2 compose 的 t3)rank 主机名也要能解析,
# 否则 master 把副本分配到对端段后本端 open segment 失败(TRANSFER_FAIL)
- "localhost-b-d0t0p0c0:127.0.0.1"
- "localhost-b-d0t1p0c1:127.0.0.1"
environment:
SGLANG_TORCH_PROFILER_DIR: /data/ymk/prof/torch
SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE: 1
SGLANG_RAGGED_VERIFY_MODE: static
SGLANG_HICACHE_MOONCAKE_CONFIG_PATH: /data/ymk/ds41/mooncake-store.json
SGLANG_HICACHE_TIMING: 1
# L3 预取传输并行 lane 数(crc32(rid) 分 lane,每 lane 独立 gloo 组;
# 实例内所有 rank 必须一致)。2026-10-10 A/B 见 l3-hicache-deep-dive 文档。
SGLANG_HICACHE_PREFETCH_IO_THREADS: 4
CUDA_VISIBLE_DEVICES: 0,1
command: >
sglang serve
--trust-remote-code
--model-path /data/models/DeepSeek-V4.1-Flash
--tp 2 --ep-size 2
--enable-prefill-cp --cp-strategy interleave
--mem-fraction-static 0.75
--attention-backend dsv4
--moe-runner-backend flashinfer_mxfp4
--cuda-graph-max-bs-decode 32
--reasoning-parser auto --tool-call-parser auto
--default-chat-template-kwargs '{"thinking": true, "reasoning_effort": "high"}'
--max-running-requests 256
--chunked-prefill-size 32768
--tokenizer-worker-num 8
--host 0.0.0.0 --port 30040
--enable-cache-report --enable-metrics
--speculative-algorithm DSPARK --speculative-dspark-block-size 5
--enable-hierarchical-cache
--hicache-ratio 4.0
--hicache-mem-layout page_first_direct
--hicache-io-backend direct
--hicache-write-policy write_through
--hicache-storage-backend mooncake
--hicache-storage-prefetch-policy wait_complete
--hicache-size 0
--max-total-tokens 1600000
+65
View File
@@ -0,0 +1,65 @@
# 【prefill 分层极限测试 2026-10-10】单实例 cp2(tp2+ep2+prefill-cp interleave)+ dspark b5
# + hicache L2(ratio 4.0=6.4M token) + L3(shm dummy) + best_effort
# 目的:测 纯计算 / L1 / L2 / L3 四档 prefill 极限吞吐 + 混合负载异步聚合验证
# L1=1.6M token(装得下 8×160k=1.28M 工作集);L2 6.4M;L3 shm 64GB
# GPU 0,1(用前停 p3d1 栈);端口 30040;共用 mooncake-store(50052)/master(50051)
x-common: &common
image: ymkymx/sglang:dsv41-pd-visioncp-1068926cd-local-20261008-0831
shm_size: 32gb
ipc: host
pid: host
privileged: true
network_mode: host
volumes:
- /data:/data
- /data/ymk/cache/sglang:/root/.cache/sglang
- /data/ymk/sglang/python:/sgl-workspace/sglang/python
- /data/ymk/ds41/fake-cargo:/root/.cargo/bin/cargo:ro
deploy:
resources:
reservations:
devices:
- driver: nvidia
count: all
capabilities: [gpu]
services:
t2:
<<: *common
container_name: ds41-tier-t2
environment:
SGLANG_TORCH_PROFILER_DIR: /data/ymk/prof/torch
SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE: 1
SGLANG_RAGGED_VERIFY_MODE: static
SGLANG_HICACHE_MOONCAKE_CONFIG_PATH: /data/ymk/ds41/mooncake-store-dummy.json
SGLANG_HICACHE_PREFETCH_IO_THREADS: 4
SGLANG_HICACHE_TIMING: 1
SGLANG_HICACHE_STORAGE_BATCH_SIZE: 512
CUDA_VISIBLE_DEVICES: 0,1
command: >
sglang serve
--trust-remote-code
--model-path /data/models/DeepSeek-V4.1-Flash
--tp 2 --ep-size 2
--enable-prefill-cp --cp-strategy interleave
--mem-fraction-static 0.75
--attention-backend dsv4
--moe-runner-backend flashinfer_mxfp4
--cuda-graph-max-bs-decode 32
--reasoning-parser auto --tool-call-parser auto
--default-chat-template-kwargs '{"thinking": true, "reasoning_effort": "high"}'
--max-running-requests 256
--chunked-prefill-size 32768
--tokenizer-worker-num 8
--host 0.0.0.0 --port 30040
--enable-cache-report --enable-metrics
--speculative-algorithm DSPARK --speculative-dspark-block-size 5
--enable-hierarchical-cache
--hicache-ratio 4.0
--hicache-mem-layout page_first_direct
--hicache-io-backend direct
--hicache-write-policy write_through
--hicache-storage-backend mooncake
--hicache-storage-prefetch-policy wait_complete
--hicache-size 0
--max-total-tokens 1600000
+454
View File
@@ -0,0 +1,454 @@
# ds41-3p1d: DeepSeek-V4.1-Flash 3P1D PD 分离(cs32k + OTLP trace),RBG 版
# 拓扑: 3×prefill(tp2+ep2+cp2, cs32k, dspark, hicache L3) + 1×decode(tp2+ep2+dp2, cg256 mr320, dspark,
# load-balance-method=total_requests 按请求数分流) + mooncake-master(hicache L3 store) + router(mini-lb PD)
# 全部钉 b300-01(单机 8 卡,hostNetwork + hostPID——mooncake NVLink-intra 跨 Pod CUDA IPC 必需,
# 等同 docker compose 的 pid: host;参考 AGENTS.md「部署四要素」)
# 镜像已侧载 containerd: docker.io/ymkymx/sglang:dsv41-pd-visioncp-1068926cd-local-20261008-0831(IfNotPresent)
# 参数来源: deploy/b300-ds41/dockerserve-3p1d-otlp-cs32k.yml(docker compose 版)
# 对外: 仅 ClusterIP Service ds41-3p1d:50002(router 用普通 Pod 网络,不占宿主机端口,
# 天然不对宿主机/外网暴露);new-api-gw 用集群 DNS http://ds41-3p1d:50002/v1 访问
# trace: OTLP → 127.0.0.1:4317(同节点 jaeger-trace cw,badger 6h 轮窗)
apiVersion: workloads.x-k8s.io/v1alpha2
kind: RoleBasedGroup
metadata:
name: ds41-3p1d
labels:
app: ds41-3p1d
sglang-model: ds-v41-flash
spec:
roleTemplates:
- name: prefill-base
template:
metadata:
labels:
sglang-model: ds-v41-flash
sglang-role: prefill
spec:
hostNetwork: true
hostPID: true
dnsPolicy: ClusterFirstWithHostNet
nodeSelector:
kubernetes.io/hostname: b300-01
containers:
- name: sglang
image: ymkymx/sglang:dsv41-pd-visioncp-1068926cd-local-20261008-0831
imagePullPolicy: IfNotPresent
command:
- /bin/bash
- -c
- |
exec sglang serve \
--trust-remote-code \
--model-path /data/models/DeepSeek-V4.1-Flash \
--tp 2 --ep-size 2 \
--enable-prefill-cp --cp-strategy interleave \
--chunked-prefill-size 32768 --max-prefill-tokens 32768 \
--mem-fraction-static 0.75 \
--attention-backend dsv4 \
--moe-runner-backend flashinfer_mxfp4 \
--cuda-graph-max-bs-decode 32 \
--reasoning-parser auto --tool-call-parser auto \
--default-chat-template-kwargs '{"thinking": true, "reasoning_effort": "high"}' \
--max-running-requests 256 \
--prefill-max-requests 8 \
--tokenizer-worker-num 8 \
--host 0.0.0.0 --port "$SGLANG_PORT" \
--enable-cache-report --enable-metrics \
--speculative-algorithm DSPARK --speculative-dspark-block-size 5 \
--enable-hierarchical-cache \
--hicache-ratio 2 \
--hicache-mem-layout page_first_direct \
--hicache-io-backend direct \
--hicache-write-policy write_through \
--hicache-storage-backend mooncake \
--hicache-storage-prefetch-policy wait_complete \
--hicache-size 0 \
--disaggregation-mode prefill \
--disaggregation-transfer-backend mooncake \
--disaggregation-bootstrap-port "$SGLANG_BOOTSTRAP_PORT" \
--enable-trace \
--trace-modules request,mooncake \
--otlp-traces-endpoint 127.0.0.1:4317 \
--otlp-service-name "$OTLP_NAME"
env:
- name: SGLANG_TORCH_PROFILER_DIR
value: /data/ymk/prof/torch
- name: SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE
value: "1"
- name: SGLANG_RAGGED_VERIFY_MODE
value: static
- name: MC_INTRANODE_NVLINK
value: "true"
- name: MC_INTRA_NVLINK
value: "true"
- name: SGLANG_MOONCAKE_SEND_AUX_TCP
value: "1"
- name: SGLANG_HICACHE_MOONCAKE_CONFIG_PATH
value: /data/ymk/ds41/mooncake-store.json
- name: SGLANG_HICACHE_PREFETCH_IO_THREADS
value: "4"
- name: SGLANG_DISAGGREGATION_QUEUE_SIZE
value: "16"
- name: SGLANG_DISAGGREGATION_THREAD_POOL_SIZE
value: "32"
- name: TZ
value: Asia/Shanghai
readinessProbe:
httpGet:
path: /health
port: metrics
initialDelaySeconds: 60
periodSeconds: 10
timeoutSeconds: 5
failureThreshold: 90
resources:
requests:
cpu: "32"
memory: 220Gi
nvidia.com/gpu: "2"
limits:
cpu: "96"
memory: 950Gi
nvidia.com/gpu: "2"
securityContext:
privileged: true
volumeMounts:
- name: data
mountPath: /data
- name: sglang-cache
mountPath: /root/.cache/sglang
- name: dshm
mountPath: /dev/shm
volumes:
- name: data
hostPath:
path: /data
type: Directory
- name: sglang-cache
hostPath:
path: /data/ymk/cache/sglang
type: Directory
- name: dshm
emptyDir:
medium: Memory
sizeLimit: 32Gi
- name: decode-base
template:
metadata:
labels:
sglang-model: ds-v41-flash
sglang-role: decode
spec:
hostNetwork: true
hostPID: true
dnsPolicy: ClusterFirstWithHostNet
nodeSelector:
kubernetes.io/hostname: b300-01
containers:
- name: sglang
image: ymkymx/sglang:dsv41-pd-visioncp-1068926cd-local-20261008-0831
imagePullPolicy: IfNotPresent
command:
- /bin/bash
- -c
- |
exec sglang serve \
--trust-remote-code \
--model-path /data/models/DeepSeek-V4.1-Flash \
--tp 2 --ep-size 2 --dp-size 2 \
--enable-dp-attention --enable-dp-lm-head \
--mem-fraction-static 0.83 \
--attention-backend dsv4 \
--moe-runner-backend flashinfer_mxfp4 \
--cuda-graph-max-bs-decode 256 \
--load-balance-method total_requests \
--reasoning-parser auto --tool-call-parser auto \
--default-chat-template-kwargs '{"thinking": true, "reasoning_effort": "high"}' \
--max-running-requests 320 \
--tokenizer-worker-num 8 \
--host 0.0.0.0 --port 30030 \
--enable-cache-report --enable-metrics \
--speculative-algorithm DSPARK --speculative-dspark-block-size 5 \
--disaggregation-mode decode \
--disaggregation-transfer-backend mooncake \
--disaggregation-bootstrap-port 9001 \
--enable-trace \
--trace-modules request,mooncake \
--otlp-traces-endpoint 127.0.0.1:4317 \
--otlp-service-name ds41-3p1d-d1
env:
- name: SGLANG_TORCH_PROFILER_DIR
value: /data/ymk/prof/torch
- name: SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE
value: "1"
- name: SGLANG_RAGGED_VERIFY_MODE
value: static
- name: MC_INTRANODE_NVLINK
value: "true"
- name: MC_INTRA_NVLINK
value: "true"
- name: SGLANG_MOONCAKE_SEND_AUX_TCP
value: "1"
- name: SGLANG_DISAGGREGATION_QUEUE_SIZE
value: "16"
- name: SGLANG_DISAGGREGATION_THREAD_POOL_SIZE
value: "32"
- name: TZ
value: Asia/Shanghai
ports:
- name: metrics
containerPort: 30030
readinessProbe:
httpGet:
path: /health
port: metrics
initialDelaySeconds: 60
periodSeconds: 10
timeoutSeconds: 5
failureThreshold: 90
resources:
requests:
cpu: "32"
memory: 100Gi
nvidia.com/gpu: "2"
limits:
cpu: "96"
memory: 600Gi
nvidia.com/gpu: "2"
securityContext:
privileged: true
volumeMounts:
- name: data
mountPath: /data
- name: sglang-cache
mountPath: /root/.cache/sglang
- name: dshm
mountPath: /dev/shm
volumes:
- name: data
hostPath:
path: /data
type: Directory
- name: sglang-cache
hostPath:
path: /data/ymk/cache/sglang
type: Directory
- name: dshm
emptyDir:
medium: Memory
sizeLimit: 32Gi
roles:
- name: mooncake-master
replicas: 1
standalonePattern:
template:
metadata:
labels:
app: ds41-3p1d-mooncake-master
spec:
hostNetwork: true
dnsPolicy: ClusterFirstWithHostNet
nodeSelector:
kubernetes.io/hostname: b300-01
containers:
- name: mooncake-master
image: ymkymx/sglang:dsv41-pd-visioncp-1068926cd-local-20261008-0831
imagePullPolicy: IfNotPresent
command:
- /bin/bash
- -c
- |
exec /opt/sglang/bin/mooncake_master \
--port 50051 \
--enable_http_metadata_server=true \
--http_metadata_server_port=18080 \
--eviction_high_watermark_ratio=0.90
readinessProbe:
tcpSocket:
port: 50051
initialDelaySeconds: 5
periodSeconds: 5
failureThreshold: 60
resources:
requests:
cpu: "1"
memory: 8Gi
limits:
cpu: "4"
memory: 64Gi
- name: p1
replicas: 1
dependencies:
- mooncake-master
standalonePattern:
templateRef:
name: prefill-base
patch:
metadata:
labels:
app: ds41-3p1d-p1
spec:
containers:
- name: sglang
env:
- name: CUDA_VISIBLE_DEVICES
value: "0,1"
- name: SGLANG_PORT
value: "30000"
- name: SGLANG_BOOTSTRAP_PORT
value: "8998"
- name: OTLP_NAME
value: ds41-3p1d-p1
ports:
- name: metrics
containerPort: 30000
- name: p2
replicas: 1
dependencies:
- mooncake-master
standalonePattern:
templateRef:
name: prefill-base
patch:
metadata:
labels:
app: ds41-3p1d-p2
spec:
containers:
- name: sglang
env:
- name: CUDA_VISIBLE_DEVICES
value: "2,3"
- name: SGLANG_PORT
value: "30010"
- name: SGLANG_BOOTSTRAP_PORT
value: "8999"
- name: OTLP_NAME
value: ds41-3p1d-p2
ports:
- name: metrics
containerPort: 30010
- name: p3
replicas: 1
dependencies:
- mooncake-master
standalonePattern:
templateRef:
name: prefill-base
patch:
metadata:
labels:
app: ds41-3p1d-p3
spec:
containers:
- name: sglang
env:
- name: CUDA_VISIBLE_DEVICES
value: "4,5"
- name: SGLANG_PORT
value: "30020"
- name: SGLANG_BOOTSTRAP_PORT
value: "9000"
- name: OTLP_NAME
value: ds41-3p1d-p3
ports:
- name: metrics
containerPort: 30020
- name: d1
replicas: 1
dependencies:
- mooncake-master
standalonePattern:
templateRef:
name: decode-base
patch:
metadata:
labels:
app: ds41-3p1d-d1
spec:
containers:
- name: sglang
env:
- name: CUDA_VISIBLE_DEVICES
value: "6,7"
- name: router
replicas: 1
dependencies:
- p1
- p2
- p3
- d1
standalonePattern:
template:
metadata:
labels:
app: ds41-3p1d-router
spec:
# 入口不走 hostNetwork:普通 Pod 网络 + ClusterIP,天然不对宿主机/外网暴露。
# P/D 地址用节点内网 IP(hostNetwork 监听在 192.168.1.243 上,Calico 可达)。
nodeSelector:
kubernetes.io/hostname: b300-01
containers:
- name: router
image: ymkymx/sglang:dsv41-pd-visioncp-1068926cd-local-20261008-0831
imagePullPolicy: IfNotPresent
command:
- /bin/bash
- -c
- |
exec sglang-router launch \
--mini-lb --pd-disaggregation \
--host 0.0.0.0 --port 50002 \
--prefill http://192.168.1.243:30000 8998 \
--prefill http://192.168.1.243:30010 8999 \
--prefill http://192.168.1.243:30020 9000 \
--decode http://192.168.1.243:30030
ports:
- name: http
containerPort: 50002
readinessProbe:
httpGet:
path: /health
port: http
initialDelaySeconds: 5
periodSeconds: 10
timeoutSeconds: 5
failureThreshold: 30
resources:
requests:
cpu: "1"
memory: 4Gi
limits:
cpu: "8"
memory: 16Gi
---
apiVersion: v1
kind: Service
metadata:
name: ds41-3p1d
labels:
app: ds41-3p1d
spec:
type: ClusterIP
selector:
rbg.workloads.x-k8s.io/group-name: ds41-3p1d
rbg.workloads.x-k8s.io/role-name: router
ports:
- name: http
port: 50002
targetPort: 50002
---
apiVersion: monitoring.coreos.com/v1
kind: PodMonitor
metadata:
name: ds41-3p1d
labels:
release: kube-prometheus-stack
spec:
selector:
matchLabels:
sglang-model: ds-v41-flash
podMetricsEndpoints:
- interval: 15s
path: /metrics
port: metrics
+2
View File
@@ -0,0 +1,2 @@
#!/bin/sh
exit 1
@@ -0,0 +1,4 @@
{
"standalone_storage": true,
"client_server_address": "127.0.0.1:50052"
}
+172 -38
View File
@@ -18,6 +18,8 @@
namespace sglang {
using device::math::fast_mod_div_u32_t;
/// \brief Ratio-2 decode compressor: pair-pool, RMSNorm and the main-KV write.
///
/// `kv_input` and `kv_state` rows are `2 * kHeadDim` floats, kv then score.
@@ -34,8 +36,7 @@ struct Compress2DecodeParams {
const int64_t* __restrict__ req; // [num_tokens], req_pool_idx per token
const void* __restrict__ raw_out_loc; // [num_tokens] LocT, the FULL slot; 0 marks a padded row
uint8_t* __restrict__ kvcache; // [npages, kPageBytes] uint8
/// Positions per request slot in the pair-state ring.
uint32_t ring_size;
fast_mod_div_u32_t ring_size; // Positions per request slot in the pair-state ring.
float eps;
};
@@ -43,21 +44,32 @@ struct Compress2DecodeParams {
/// At (512, 64) it also keeps the nope/rope split warp-aligned, as the fp8 amax reduction requires.
constexpr uint32_t kC2VecSize = 2;
/// \brief grid = num_tokens, block = kHeadDim / kC2VecSize.
/// \brief Where a completing row finds its partner, and who parks a pending row.
enum class C2Mode {
/// One row per request: the partner comes from the ring, a pending row parks there.
Decode,
/// `draft_len` consecutive positions per request: `blockIdx.x` is the position inside the
/// block, `blockIdx.y` the request, and every row but the first takes its partner from the
/// previous `kv_input` row instead of the ring.
Verify,
/// Arbitrary rows, paired as `_low_ratio_pair_partners` pairs them: the partner is the previous
/// row when it is the same request's previous position, else the ring. A pending row does
/// nothing here; `c2_prefill_store_kernel` parks the ring rows after this kernel read them.
Prefill,
};
/// \brief grid = num_tokens (Decode, Prefill) or `draft_len x batch` (Verify),
/// block = kHeadDim / kC2VecSize.
///
/// An odd position completes a group with its even predecessor; an even one parks in the state.
///
/// Under target-verify, `draft_len` consecutive positions per request: `blockIdx.x` is the
/// position inside the block, `blockIdx.y` the request, and every row but the first takes its
/// partner from the previous `kv_input` row instead of the ring.
///
/// The three reductions have different widths and are not interchangeable: the RMSNorm
/// statistic spans the row, an fp8 store scale 64 elements, an fp4 block 16.
///
/// kLayout is the cache's page format: V4 and V41 store the fake-quantized value; V41_FP4
/// stores the e2m1 codes and their e4m3 scales directly, so the fp4 rounding happens once.
template <
bool kVerify,
C2Mode kMode,
int64_t kHeadDim,
int64_t kRopeDim,
int32_t kPageBits,
@@ -91,16 +103,21 @@ __global__ __launch_bounds__(kHeadDim / kC2VecSize) void flash_c2_decode_kernel(
const auto tx = threadIdx.x;
// Verify gives each request a CTA column; decode a flat grid of one row each.
const auto row = kVerify ? blockIdx.y * gridDim.x + blockIdx.x : blockIdx.x;
const auto row = kMode == C2Mode::Verify ? blockIdx.y * gridDim.x + blockIdx.x : blockIdx.x;
PDLWaitPrimary<kUsePDL>();
// Slots fit in int32 whatever width the scheduler hands them in.
const auto raw_out_loc = static_cast<int32_t>(static_cast<const LocT*>(params.raw_out_loc)[row]);
// CUDA graph padding is a completely inert row: do not read schedule, input,
// state, or RoPE data, and do not publish output or update the cache.
if (raw_out_loc == 0) return PDLTriggerSecondary<kUsePDL>();
const auto pos = static_cast<const PosT*>(params.positions)[row];
// A completing row reads the slot left by `pos - 1`;
// a pending row writes its own slot, so reads and writes stay disjoint.
const auto* loc_ptr = static_cast<const LocT*>(params.raw_out_loc);
const auto* pos_ptr = static_cast<const PosT*>(params.positions);
const auto raw_out_loc = loc_ptr[row];
// Prefill rows are many, so a CUDA graph padding row exits before any load.
// Decode and verify rows are few: every row issues its loads to keep more in flight.
if constexpr (kMode == C2Mode::Prefill) {
if (raw_out_loc == 0) return PDLTriggerSecondary<kUsePDL>();
}
const auto pos = pos_ptr[row];
// Prefill leaves a pending row's latent unwritten; its index-K slot is -1, so nothing reads it.
if constexpr (kMode == C2Mode::Prefill) {
if ((pos & 1) == 0) return PDLTriggerSecondary<kUsePDL>();
}
const auto rid = params.req[row];
fp32_vec_t kv_new, score_new;
@@ -108,21 +125,32 @@ __global__ __launch_bounds__(kHeadDim / kC2VecSize) void flash_c2_decode_kernel(
score_new.load(params.kv_input + row * kStride, tx + kCTASize);
const auto ring = static_cast<int64_t>(rid) * params.ring_size;
const auto read_row = ring + (pos - 1 + params.ring_size) % params.ring_size;
const auto write_row = ring + pos % params.ring_size;
const auto write_offset = static_cast<uint32_t>(pos) % params.ring_size;
const auto read_offset = write_offset == 0 ? params.ring_size - 1 : write_offset - 1;
const auto write_row = ring + write_offset;
const auto read_row = ring + read_offset;
fp32_vec_t kv_old, score_old;
const float* partner = params.kv_state + read_row * kStride;
if constexpr (kVerify) {
if constexpr (kMode == C2Mode::Verify) {
if (blockIdx.x != 0) partner = params.kv_input + static_cast<int64_t>(row - 1) * kStride;
}
if constexpr (kMode == C2Mode::Prefill) {
if (row > 0 && params.req[row - 1] == rid && pos_ptr[row - 1] == pos - 1 && loc_ptr[row - 1] != 0) {
partner = params.kv_input + static_cast<int64_t>(row - 1) * kStride;
}
}
kv_old.load(partner, tx);
score_old.load(partner, tx + kCTASize);
if ((pos & 1) == 0) {
kv_new.store(params.kv_state + write_row * kStride, tx);
score_new.store(params.kv_state + write_row * kStride, tx + kCTASize);
return PDLTriggerSecondary<kUsePDL>();
if constexpr (kMode != C2Mode::Prefill) {
// CUDA graph padding case
if (raw_out_loc == 0) return PDLTriggerSecondary<kUsePDL>();
if ((pos & 1) == 0) {
kv_new.store(params.kv_state + write_row * kStride, tx);
score_new.store(params.kv_state + write_row * kStride, tx + kCTASize);
return PDLTriggerSecondary<kUsePDL>();
}
}
constexpr uint32_t kNumWarps = kCTASize / kWarpThreads;
@@ -203,9 +231,9 @@ __global__ __launch_bounds__(kHeadDim / kC2VecSize) void flash_c2_decode_kernel(
}
}
const int32_t out_loc = static_cast<uint32_t>(raw_out_loc) >> 1;
if constexpr (kLayout == KVLayout::V41_FP4) {
// The fp4 cache takes the rotated bf16 value as is: its row quantizer is the fake quant, minus the dequant.
const int32_t out_loc = raw_out_loc >> 1;
const auto kv_row = Paged::row(params.kvcache, out_loc);
return deepseek_v4::v41::store_row<kLayout>(kv_row.data, kv_row.scale, tx, staged);
}
@@ -228,7 +256,6 @@ __global__ __launch_bounds__(kHeadDim / kC2VecSize) void flash_c2_decode_kernel(
}
// `raw_out_loc / ratio`; ratio 2 makes it a shift.
const int32_t out_loc = raw_out_loc >> 1;
const auto kv_row = Paged::row(params.kvcache, out_loc);
if constexpr (kLayout == KVLayout::V41) {
@@ -264,21 +291,63 @@ __global__ __launch_bounds__(kHeadDim / kC2VecSize) void flash_c2_decode_kernel(
}
}
/// \brief Prefill ring write-back, launched after `flash_c2_decode_kernel<C2Mode::Prefill>`.
///
/// Each request's last `ring_size` rows park their `| kv | score |` at `req * ring_size +
/// pos % ring_size` for a later step's first partner. It cannot run inside the pooling kernel:
/// a chunk whose first row is odd and whose last row is even reads and writes the same ring
/// slot. grid = `ring_size x num_reqs`, CTA (k, i) parks the k-th row from the end of request i,
/// one float4 per thread; those rows are consecutive positions, so their slots are distinct.
template <int64_t kHeadDim, typename PosT, typename LocT, bool kUsePDL>
__global__ __launch_bounds__(kHeadDim / 2) void c2_prefill_store_kernel(
const Compress2DecodeParams params,
const int32_t* __restrict__ extend_start_loc,
const int32_t* __restrict__ extend_seq_lens) {
using namespace device;
constexpr int64_t kStride = kHeadDim * 2;
using vec_t = AlignedVector<float, 4>;
static_assert(kStride == 4 * (kHeadDim / 2));
// The pooling kernel reads the ring rows this one overwrites.
PDLWaitPrimary<kUsePDL>();
const auto bx = static_cast<int32_t>(blockIdx.x);
const auto len = extend_seq_lens[blockIdx.y];
const auto start_loc = extend_start_loc[blockIdx.y];
if (bx >= len) return;
const auto row = start_loc + (len - 1 - bx);
const auto rid = params.req[row];
const auto out = static_cast<const LocT*>(params.raw_out_loc)[row];
const auto pos = static_cast<uint32_t>(static_cast<const PosT*>(params.positions)[row]);
const auto dst_row = rid * params.ring_size + pos % params.ring_size;
if (out == 0) return;
vec_t buf;
buf.load(params.kv_input + row * kStride, threadIdx.x);
PDLTriggerSecondary<kUsePDL>();
buf.store(params.kv_state + dst_row * kStride, threadIdx.x);
}
template <int64_t kHeadDim, int64_t kRopeDim, uint32_t kPageSize, deepseek_v4::KVLayout kLayout, bool kUsePDL>
struct FlashCompress2Kernel {
static constexpr uint32_t kBlockSize = kHeadDim / kC2VecSize;
static constexpr int32_t kPageBits = std::bit_width(kPageSize) - 1;
static constexpr int64_t kPageBytes = deepseek_v4::kv_page_bytes<kLayout>(kPageSize);
static_assert(kLayout != deepseek_v4::KVLayout::V4 || kPageBytes == host::div_ceil(584ll * kPageSize, 576) * 576);
template <bool kVerify, typename PosT, typename LocT>
template <C2Mode kMode, typename PosT, typename LocT>
static constexpr auto kernel =
flash_c2_decode_kernel<kVerify, kHeadDim, kRopeDim, kPageBits, PosT, LocT, kLayout, kUsePDL>;
flash_c2_decode_kernel<kMode, kHeadDim, kRopeDim, kPageBits, PosT, LocT, kLayout, kUsePDL>;
template <typename PosT, typename LocT>
static constexpr auto store_kernel = c2_prefill_store_kernel<kHeadDim, PosT, LocT, kUsePDL>;
/// \brief The (`positions`, `raw_out_loc`) dtype pair, resolved at run time.
template <bool kVerify>
template <C2Mode kMode>
static auto select(const bool pos_i32, const bool loc_i32) {
if (pos_i32) return loc_i32 ? kernel<kVerify, int32_t, int32_t> : kernel<kVerify, int32_t, int64_t>;
return loc_i32 ? kernel<kVerify, int64_t, int32_t> : kernel<kVerify, int64_t, int64_t>;
if (pos_i32) return loc_i32 ? kernel<kMode, int32_t, int32_t> : kernel<kMode, int32_t, int64_t>;
return loc_i32 ? kernel<kMode, int64_t, int32_t> : kernel<kMode, int64_t, int64_t>;
}
static auto select_store(const bool pos_i32, const bool loc_i32) {
if (pos_i32) return loc_i32 ? store_kernel<int32_t, int32_t> : store_kernel<int32_t, int64_t>;
return loc_i32 ? store_kernel<int64_t, int32_t> : store_kernel<int64_t, int64_t>;
}
// The sum of squares reduces through a fixed-size shared array, so the CTA must be whole warps.
@@ -302,7 +371,7 @@ struct FlashCompress2Kernel {
const tvm::ffi::TensorView kvcache,
const int64_t ring_size,
const int64_t draft_len) {
launch(
launch<false>(
kv_input,
kv_state,
kv_output,
@@ -317,7 +386,62 @@ struct FlashCompress2Kernel {
draft_len);
}
/// \brief The extend-batch pooling kernel, then the ring write-back.
///
/// Rows may mix requests, prefix offsets and padding; a pending (even) row's `kv_output`
/// row is left unwritten. `extend_start_loc` / `extend_seq_lens` are int32 `[num_reqs]`:
/// request i owns rows `[start[i], start[i] + len[i])`; zero-length requests are skipped.
static void run_prefill_fusion(
const tvm::ffi::TensorView kv_input,
const tvm::ffi::TensorView kv_state,
const tvm::ffi::TensorView kv_output,
const tvm::ffi::TensorView norm_weight,
const tvm::ffi::TensorView positions,
const tvm::ffi::TensorView req,
const tvm::ffi::TensorView raw_out_loc,
const float eps,
const tvm::ffi::TensorView freqs_cis,
const tvm::ffi::TensorView kvcache,
const int64_t ring_size,
const tvm::ffi::TensorView extend_start_loc,
const tvm::ffi::TensorView extend_seq_lens) {
using namespace host;
auto B = SymbolicSize{"num_reqs"};
auto device_ = SymbolicDevice{};
device_.set_options<kDLGPU>();
TensorMatcher({B}).with_dtype<int32_t>().with_device(device_).verify(extend_start_loc);
TensorMatcher({B}).with_dtype<int32_t>().with_device(device_).verify(extend_seq_lens);
const auto requests = StoreRequests{
.extend_start_loc = static_cast<const int32_t*>(extend_start_loc.data_ptr()),
.extend_seq_lens = static_cast<const int32_t*>(extend_seq_lens.data_ptr()),
.num_reqs = static_cast<uint32_t>(B.unwrap()),
};
launch<true>(
kv_input,
kv_state,
kv_output,
norm_weight,
positions,
req,
raw_out_loc,
eps,
ring_size,
freqs_cis,
kvcache,
1,
&requests);
}
private:
struct StoreRequests {
const int32_t* extend_start_loc;
const int32_t* extend_seq_lens;
uint32_t num_reqs;
};
/// \tparam kPrefill `C2Mode::Prefill` plus the ring write-back of `requests`; otherwise
/// Decode, or Verify when `draft_len > 1`.
template <bool kPrefill>
static void launch(
const tvm::ffi::TensorView kv_input,
const tvm::ffi::TensorView kv_state,
@@ -330,7 +454,8 @@ struct FlashCompress2Kernel {
const int64_t ring_size,
const tvm::ffi::TensorView freqs_cis,
const tvm::ffi::TensorView kvcache,
const int64_t draft_len) {
const int64_t draft_len,
const StoreRequests* requests = nullptr) {
using namespace host;
auto N = SymbolicSize{"num_tokens"};
@@ -358,7 +483,7 @@ struct FlashCompress2Kernel {
const auto num_tokens = static_cast<uint32_t>(N.unwrap());
if (num_tokens == 0) return;
const auto is_verify = draft_len > 1;
const auto is_verify = !kPrefill && draft_len > 1;
CHECK_HOST(ring_size > 0 && draft_len >= 1);
CHECK_HOST(!is_verify || num_tokens % draft_len == 0);
CHECK_HOST(!is_verify || ring_size > draft_len)
@@ -373,19 +498,28 @@ struct FlashCompress2Kernel {
.req = static_cast<const int64_t*>(req.data_ptr()),
.raw_out_loc = raw_out_loc.data_ptr(),
.kvcache = static_cast<uint8_t*>(kvcache.data_ptr()),
.ring_size = static_cast<uint32_t>(ring_size),
.ring_size = static_cast<fast_mod_div_u32_t>(ring_size),
.eps = eps,
};
// `LaunchKernel` is move-only, so each arm builds its own.
const auto pos_i32 = pos_dtype.is_type<int32_t>();
const auto loc_i32 = loc_dtype.is_type<int32_t>();
if (is_verify) {
if constexpr (kPrefill) {
const auto k = select<C2Mode::Prefill>(pos_i32, loc_i32);
LaunchKernel(num_tokens, kBlockSize, device_.unwrap()) //
.enable_pdl(kUsePDL)(k, params);
if (requests->num_reqs == 0) return;
const auto store = select_store(pos_i32, loc_i32);
const auto grid = dim3{static_cast<uint32_t>(ring_size), requests->num_reqs};
LaunchKernel(grid, kHeadDim / 2, device_.unwrap()) //
.enable_pdl(kUsePDL)(store, params, requests->extend_start_loc, requests->extend_seq_lens);
} else if (is_verify) {
const auto block = static_cast<uint32_t>(draft_len);
const auto k = select<true>(pos_i32, loc_i32);
const auto k = select<C2Mode::Verify>(pos_i32, loc_i32);
LaunchKernel(dim3{block, num_tokens / block}, kBlockSize, device_.unwrap()) //
.enable_pdl(kUsePDL)(k, params);
} else {
const auto k = select<false>(pos_i32, loc_i32);
const auto k = select<C2Mode::Decode>(pos_i32, loc_i32);
LaunchKernel(num_tokens, kBlockSize, device_.unwrap()) //
.enable_pdl(kUsePDL)(k, params);
}
@@ -2,6 +2,7 @@
#include <sgl_kernel/tensor.h>
#include <sgl_kernel/utils.h>
#include <sgl_kernel/math.cuh>
#include <sgl_kernel/runtime.cuh>
#include <sgl_kernel/type.cuh>
#include <sgl_kernel/utils.cuh>
@@ -21,31 +22,7 @@ namespace sglang {
using device::distributed::PushWorkSpace;
using device::distributed::Semaphore;
// Runtime uint32 division as a multiply-high and a shift (round-up magic,
// exact below 2^31); cuda::fast_mod_div needs a newer CCCL than CUDA 13 bundles.
struct fast_mod_div_u32_t {
uint32_t divisor;
uint32_t magic;
uint32_t shift;
__host__ explicit fast_mod_div_u32_t(uint32_t d) : divisor(d), magic(0), shift(0) {
if (d > 1) {
const uint32_t log2_ceil = 32 - std::countl_zero(d - 1);
const uint32_t p = 31 + log2_ceil;
magic = static_cast<uint32_t>(((uint64_t{1} << p) + d - 1) / d);
shift = p - 32;
}
}
__device__ friend uint32_t operator/(uint32_t n, const fast_mod_div_u32_t& fd) {
return fd.divisor == 1 ? n : __umulhi(n, fd.magic) >> fd.shift;
}
__device__ friend uint32_t operator%(uint32_t n, const fast_mod_div_u32_t& fd) {
return n - (n / fd) * fd.divisor;
}
};
using device::math::fast_mod_div_u32_t;
template <uint32_t kWorldSize>
struct NVLinkCommPushParams {
@@ -168,8 +145,7 @@ PUSH_KERNEL void nvlink_push_kernel(const __grid_constant__ NVLinkCommPushParams
ptx::st_multimem_16B(vec, dst_ptr_mc, vid);
}
} else /* reduce-scatter only */ {
const auto token_id = vid / params.vecs_per_token_div;
const auto offset = vid % params.vecs_per_token_div;
const auto [token_id, offset] = math::div(vid, params.vecs_per_token_div);
// Both by a compile-time constant, so this is a mask and a shift.
const auto dst_rank = token_id % kWorldSize;
const auto dst_token_id = token_id / kWorldSize;
@@ -112,6 +112,53 @@ SGL_DEVICE float fma_f32_bf16(bf16_t a, float b, float acc) {
return fmaf(cast<fp32_t>(a), b, acc);
}
/**
* \brief Fast uint32 division and modulo by a compile-time constant divisor.
*
* This struct is only exact when the divisor is less than `2^31`. Larger dividends
* will lead to incorrect results. Be careful when using this struct.
*/
struct fast_mod_div_u32_t {
public:
explicit constexpr fast_mod_div_u32_t(uint32_t d) : m_divisor(d) {
if (d > 1) {
const uint32_t log2_ceil = 32 - std::countl_zero(d - 1);
const uint32_t p = 31 + log2_ceil;
m_magic = static_cast<uint32_t>(((uint64_t{1} << p) + d - 1) / d);
m_shift = p - 32;
} else {
m_magic = 0;
m_shift = 0;
}
}
template <typename T>
SGL_DEVICE friend uint32_t operator/(T n, const fast_mod_div_u32_t& d) {
static_assert(std::is_unsigned_v<T> && sizeof(T) <= sizeof(uint32_t));
return d.m_divisor == 1 ? n : __umulhi(n, d.m_magic) >> d.m_shift;
}
template <typename T>
SGL_DEVICE friend uint32_t operator%(T n, const fast_mod_div_u32_t& d) {
static_assert(std::is_unsigned_v<T> && sizeof(T) <= sizeof(uint32_t));
return n - (n / d) * d.m_divisor;
}
constexpr operator uint32_t() const {
return m_divisor;
}
private:
uint32_t m_divisor;
uint32_t m_magic;
uint32_t m_shift;
};
template <typename T>
SGL_DEVICE uint2 div(T n, const fast_mod_div_u32_t& d) {
return {n / d, n % d};
}
} // namespace device::math
} // namespace sglang
@@ -273,6 +273,67 @@ def store_fp4_index_k_cache(
)
@triton.jit
def _gather_fp4_index_k_kernel(
cache,
cache_stride,
slots,
payload,
scales,
n,
page_size: tl.constexpr,
BLOCK_ROWS: tl.constexpr,
):
# int32 view of a page: page_size * 16 payload words, then page_size scale words.
rows = tl.program_id(0) * BLOCK_ROWS + tl.arange(0, BLOCK_ROWS)
valid = rows < n
slot = tl.load(slots + rows, valid, 0).to(tl.int64)
page = slot // page_size
offset = slot - page * page_size
page_base = cache + page * cache_stride
words = tl.arange(0, 16)
payload_words = tl.load(
page_base[:, None] + offset[:, None] * 16 + words[None, :], valid[:, None]
)
tl.store(
payload + rows[:, None].to(tl.int64) * 16 + words[None, :],
payload_words,
valid[:, None],
)
tl.store(scales + rows, tl.load(page_base + page_size * 16 + offset, valid), valid)
def gather_fp4_index_k(
cache: torch.Tensor, slots: torch.Tensor, *, page_size: int
) -> tuple[torch.Tensor, torch.Tensor]:
"""Packed fp4 rows at ``slots`` of the index-K ``cache`` (``[pages, page_size *
68]`` uint8, ``store_fp4_index_k_cache``'s layout): (payload int8 ``[n, 64]``,
ue8m0 scales packed int32 ``[n]``), the layout ``quantize_fp4_indexer_tensor``
returns."""
assert cache.shape[1] == page_size * INDEX_K_SLOT_BYTES and cache.stride(1) == 1
# The kernel addresses slots as `slots + rows`, so a strided view would be read
# as if its elements were packed.
assert slots.is_contiguous()
words = cache.view(torch.int32)
n = slots.shape[0]
payload = torch.empty((n, 16), device=cache.device, dtype=torch.int32)
scales = torch.empty((n,), device=cache.device, dtype=torch.int32)
if n > 0:
block_rows = 64
_gather_fp4_index_k_kernel[(triton.cdiv(n, block_rows),)](
words,
words.stride(0),
slots,
payload,
scales,
n,
page_size,
BLOCK_ROWS=block_rows,
num_warps=4,
)
return payload.view(torch.int8), scales
@triton.jit
def _index_k_rope_pack_kernel(
X,
@@ -1,5 +1,6 @@
"""Fused ratio-1 and ratio-2 decode compressors: RMSNorm, RoPE and the FlashMLA
cache write in one launch.
"""Fused ratio-1 and ratio-2 compressors: RMSNorm, RoPE and the FlashMLA cache
write in one launch. The ratio-1 kernel is per token, so decode and extend share
it; ratio 2 has a decode/verify entry point and an extend one.
Ratio 1 takes the bf16 ``wkv`` projection as is: RoPE uses the token's own
position and the compressed slot equals the FULL slot. Ratio 2 pair-pools the
@@ -17,6 +18,7 @@ import torch
from sglang.kernels.jit.utils import (
cache_once,
is_arch_support_pdl,
is_hip_runtime,
load_jit,
make_cpp_args,
)
@@ -119,11 +121,12 @@ def _jit_c2_module(
is_arch_support_pdl(),
)
return load_jit(
make_name("c2_decode"),
make_name("c2"),
*args,
cuda_files=["deepseek_v4/c2.cuh"],
cuda_wrappers=[
("decode_fusion", f"FlashCompress2Kernel<{args}>::run_decode_fusion"),
("prefill_fusion", f"FlashCompress2Kernel<{args}>::run_prefill_fusion"),
],
)
@@ -149,6 +152,12 @@ def c2_decode_norm_rope_store(
The cache store uses ``raw_out_loc // 2`` as its slot.
A padded graph row is marked by ``raw_out_loc == 0`` and writes nothing, but
decode and verify rows issue their ring load before that check, so a padded
row's ``req`` and ``positions`` must still address an in-bounds ring slot.
The decode graph buffers zero both tails (``PaddingPolicy.ZERO``); a runner
that left stale request indices there would read out of the ring pool.
:param freqs_cis: ``[max_pos, rope_dim]`` fp32, real/imag interleaved --
``torch.view_as_real(freqs).flatten(-2)``. Indexed
in-kernel at ``positions - 1``, the position the latent
@@ -181,3 +190,60 @@ def c2_decode_norm_rope_store(
draft_len,
)
return out
def c2_prefill_norm_rope_store(
kv_input: torch.Tensor,
kv_state: torch.Tensor,
norm_weight: torch.Tensor,
positions: torch.Tensor,
req: torch.Tensor,
raw_out_loc: torch.Tensor,
eps: float,
freqs_cis: torch.Tensor,
k_cache: torch.Tensor,
extend_start_loc: torch.Tensor,
extend_seq_lens: torch.Tensor,
*,
page_size: int,
ring_size: int,
layout: Union[KVLayout, str] = KVLayout.V4,
out: Optional[torch.Tensor] = None,
) -> torch.Tensor:
""":func:`c2_decode_norm_rope_store` for an extend batch, then the ring write-back.
Rows may mix requests, prefix offsets and CUDA graph padding (``raw_out_loc
== 0``). An odd position pairs with the previous row when that row is the
same request's previous position, else with its ring slot. Afterwards each
request's last ``ring_size`` rows are saved to the ring, which a later step
reads for its first partner.
:param extend_start_loc: int32 ``[num_reqs]``, each request's first row.
:param extend_seq_lens: int32 ``[num_reqs]``, each request's row count.
:return: ``out``, ``[num_tokens, head_dim]`` bf16. Only the rows of odd
positions are written; the rest are left as they were.
"""
num_tokens, fused_dim = kv_input.shape
head_dim = fused_dim // 2
if out is None:
out = kv_input.new_empty((num_tokens, head_dim), dtype=torch.bfloat16)
layout = KVLayout.parse(layout)
module = _jit_c2_module(head_dim, freqs_cis.shape[-1], page_size, layout)
module.prefill_fusion(
kv_input,
kv_state,
out,
norm_weight,
positions,
req,
raw_out_loc,
eps,
freqs_cis,
# HIP's fp8_e4m3_t is uint8_t, so the kernel matches the fp8 pool as bytes
k_cache.view(torch.uint8) if is_hip_runtime() else k_cache,
ring_size,
extend_start_loc,
extend_seq_lens,
)
return out
@@ -36,7 +36,7 @@ def _build_swa_token_ids_kernel(
tl.store(out_ptr + out_off + i, swa_id)
@triton.jit(do_not_specialize=["top_k"])
@triton.jit(do_not_specialize=["num_tokens", "num_reqs", "top_k"])
def _combine_topk_swa_indices_kernel(
combined_indices_ptr,
combined_indices_stride,
@@ -49,59 +49,76 @@ def _combine_topk_swa_indices_kernel(
gather_lens_ptr,
compressed_base_ptr,
swa_base_ptr,
num_tokens,
num_reqs,
top_k,
COMPRESS_RATIO: tl.constexpr,
WINDOW_SIZE: tl.constexpr,
PADDED_TOP_K: tl.constexpr,
WIDTH: tl.constexpr,
PADDED_WIDTH: tl.constexpr,
BLOCK_T: tl.constexpr,
SEARCH_STEPS: tl.constexpr,
):
batch_idx = tl.program_id(0)
worker_id = tl.program_id(1)
num_workers = tl.num_programs(1)
# BLOCK_T consecutive tokens per program, so the grid scales with tokens, not requests.
token = tl.program_id(0) * BLOCK_T + tl.arange(0, BLOCK_T)
valid = token < num_tokens
# query_start_loc may be a global tensor; rebase to chunk-local offsets
# by subtracting the chunk's starting value.
base = tl.load(query_start_loc_ptr)
query_start = tl.load(query_start_loc_ptr + batch_idx) - base
query_end = tl.load(query_start_loc_ptr + batch_idx + 1) - base
seq_len = tl.load(seq_lens_ptr + batch_idx)
gather_len = tl.load(gather_lens_ptr + batch_idx)
compressed_base = tl.load(compressed_base_ptr + batch_idx)
swa_base = tl.load(swa_base_ptr + batch_idx)
# Each token's request: the last r with query_start_loc[r] - base <= token.
lo = tl.zeros([BLOCK_T], dtype=tl.int32)
hi = tl.full([BLOCK_T], 0, dtype=tl.int32) + num_reqs
for _ in tl.static_range(SEARCH_STEPS):
mid = (lo + hi + 1) // 2
start = tl.load(query_start_loc_ptr + mid, mask=lo < hi, other=0) - base
go_right = (lo < hi) & (start <= token)
lo = tl.where(go_right, mid, lo)
hi = tl.where((lo < hi) & ~go_right, mid - 1, hi)
batch_idx = tl.minimum(lo, num_reqs - 1)
# Rows past the last request's queries belong to no request: -1 rows, length 0.
owned = valid & (token < tl.load(query_start_loc_ptr + num_reqs) - base)
pos = tl.load(query_pos_ptr + token, mask=owned, other=0)
seq_len = tl.load(seq_lens_ptr + batch_idx, mask=owned, other=0)
gather_len = tl.load(gather_lens_ptr + batch_idx, mask=owned, other=0)
compressed_base = tl.load(compressed_base_ptr + batch_idx, mask=owned, other=0)
swa_base = tl.load(swa_base_ptr + batch_idx, mask=owned, other=0)
# SWA portion of the gathered buffer starts from position
# (seq_len - gather_len), not 0. The +pos-gather_start formula maps a
# query's window back into the workspace's SWA region.
gather_start = seq_len - gather_len
for token_idx in range(query_start + worker_id, query_end, num_workers):
pos = tl.load(query_pos_ptr + token_idx)
# -1 entries inside the top-k span stay -1 (attention skips them).
# top_k=0 disables the compressed portion for SWA-only layers.
topk_len = tl.minimum((pos + 1) // COMPRESS_RATIO, top_k)
swa_len = tl.minimum(pos + 1, WINDOW_SIZE)
# -1 entries inside the top-k span stay -1 (attention skips them).
# top_k=0 disables the compressed portion for SWA-only layers.
topk_len = tl.where(owned, tl.minimum((pos + 1) // COMPRESS_RATIO, top_k), 0)
swa_len = tl.where(owned, tl.minimum(pos + 1, WINDOW_SIZE), 0)
combined_row = token_idx.to(tl.int64) * combined_indices_stride
topk_row = token_idx.to(tl.int64) * topk_indices_stride
combined_row = token.to(tl.int64)[:, None] * combined_indices_stride
topk_row = token.to(tl.int64)[:, None] * topk_indices_stride
offset = tl.arange(0, PADDED_TOP_K)
mask = offset < topk_len
topk_vals = tl.load(
topk_indices_ptr + topk_row + offset,
mask=mask,
)
tl.store(
combined_indices_ptr + combined_row + offset,
tl.where(topk_vals >= 0, topk_vals + compressed_base, -1),
mask=mask,
)
# Whole rows, -1 padded, so the output needs no fill pass:
# [top-k (-1 holes kept) | window | -1 up to WIDTH].
offset = tl.arange(0, PADDED_WIDTH)[None, :]
in_topk = offset < topk_len[:, None]
topk_vals = tl.load(
topk_indices_ptr + topk_row + offset, mask=valid[:, None] & in_topk, other=-1
)
# Workspace SWA index: swa_base[r] + (gather_offset_in_buffer).
# For positions [pos - swa_len + 1, pos], the buffer offsets are
# [pos - swa_len + 1 - gather_start, pos - gather_start].
swa_start = swa_base + pos - swa_len + 1 - gather_start
swa_vals = swa_start[:, None] + (offset - topk_len[:, None])
in_swa = offset < (topk_len + swa_len)[:, None]
vals = tl.where(
in_topk,
tl.where(topk_vals >= 0, topk_vals + compressed_base[:, None], -1),
tl.where(in_swa, swa_vals, -1),
)
tl.store(
combined_indices_ptr + combined_row + offset,
vals,
mask=valid[:, None] & (offset < WIDTH),
)
offset = tl.arange(0, WINDOW_SIZE)
# Workspace SWA index: swa_base[r] + (gather_offset_in_buffer).
# For positions [pos - swa_len + 1, pos], the buffer offsets are
# [pos - swa_len + 1 - gather_start, pos - gather_start].
tl.store(
combined_indices_ptr + combined_row + topk_len + offset,
swa_base + offset + pos - swa_len + 1 - gather_start,
mask=offset < swa_len,
)
tl.store(combined_lens_ptr + token_idx, topk_len + swa_len)
tl.store(combined_lens_ptr + token, topk_len + swa_len, mask=valid)
@@ -1882,6 +1882,34 @@ def hc_expand(x: torch.Tensor, n: int) -> torch.Tensor:
return x.repeat(1, n)
@triton.jit
def _hc_broadcast_kernel(
x_ptr, y_ptr, H, x_stride_m, HC: tl.constexpr, BLOCK_H: tl.constexpr
):
row = tl.program_id(0).to(tl.int64)
offs = tl.program_id(1) * BLOCK_H + tl.arange(0, BLOCK_H)
mask = offs < H
x = tl.load(x_ptr + row * x_stride_m + offs, mask=mask)
y_row = y_ptr + row * (HC * H)
for k in tl.static_range(HC):
tl.store(y_row + k * H + offs, x, mask=mask)
def hc_broadcast(x: torch.Tensor, hc_mult: int) -> torch.Tensor:
"""[T, H] -> [T, hc_mult, H], every copy equal to x: the initial mHC residual."""
if not x.is_cuda or x.stride(-1) != 1:
return x.unsqueeze(1).repeat(1, hc_mult, 1)
num_tokens, hidden = x.shape
y = x.new_empty(num_tokens, hc_mult, hidden)
if num_tokens == 0:
return y
block_h = 1024
_hc_broadcast_kernel[(num_tokens, triton.cdiv(hidden, block_h))](
x, y, hidden, x.stride(0), HC=hc_mult, BLOCK_H=block_h, num_warps=4
)
return y
def hc_contract(x: torch.Tensor, n: int) -> torch.Tensor:
return x.unflatten(-1, (n, -1)).mean(dim=-2)
@@ -32,7 +32,7 @@ class TransferKVChunk:
# Set when the staging worker first counts this chunk toward the per-room
# outstanding count; stays set across re-enqueue on a watermark defer.
staging_counted: bool = False
# Mori early-send: CUDA event to synchronize before RDMA (optional).
# Early-send: GPU event to synchronize before reading KV (optional).
wait_event: Optional[object] = None
@@ -10,6 +10,7 @@ from sglang.srt.managers.overlap_utils import RelayPayload
from sglang.srt.mem_cache.common import maybe_cache_unfinished_req
from sglang.srt.model_executor.forward_batch_info import ForwardMode
from sglang.srt.sampling.sampling_batch_info import SamplingBatchInfo
from sglang.srt.utils.common import is_pin_memory_available
logger = logging.getLogger(__name__)
@@ -140,9 +141,14 @@ class ScheduleBatchDisaggregationDecodeMixin:
error_message, HTTPStatus.INTERNAL_SERVER_ERROR
)
req.grammar.finished = req.finished()
# Non-blocking H2D: with overlap this runs on the schedule stream after
# its wait on the in-flight forward, so a blocking copy would stall the
# host until that forward ends and leave the GPU idle.
last_tokens_tensor = torch.tensor(
last_tokens, dtype=torch.int64, device=self.device
)
last_tokens,
dtype=torch.int64,
pin_memory=is_pin_memory_available(self.device),
).to(self.device, non_blocking=True)
spec_info = self.spec_algorithm.build_disagg_draft_input(
self,
@@ -960,6 +960,8 @@ class MooncakeKVManager(StagingManagerMixin, CommonKVManager):
dst_layer_ids: Optional[List[int]] = None,
dst_device_data_indices: Optional[npt.NDArray[np.int32]] = None,
dst_device_data_ptrs: Optional[set[int]] = None,
dst_item_lens: list[int] | None = None,
bootstrap_room: Optional[int] = None,
) -> int:
"""
Generic KV cache transfer supporting both MHA and MLA architectures.
@@ -1011,6 +1013,21 @@ class MooncakeKVManager(StagingManagerMixin, CommonKVManager):
len(dst_data_ptrs),
allow_positional_fallback=self.pp_size == 1,
)
if dst_item_lens is not None:
for i, j in pairs:
if item_lens[i] != dst_item_lens[j]:
assert bootstrap_room is not None
failure_reason = (
f"{state_type} item length mismatch for paired "
f"entries src[{i}]={item_lens[i]} "
f"dst[{j}]={dst_item_lens[j]}"
)
logger.error(failure_reason)
self.conclude_failure(
bootstrap_room=bootstrap_room,
failure_reason=failure_reason,
)
return -1
layers_params = [
(src_data_ptrs[i], dst_data_ptrs[j], item_lens[i]) for i, j in pairs
]
@@ -1020,6 +1037,24 @@ class MooncakeKVManager(StagingManagerMixin, CommonKVManager):
src_data_ptrs, dst_data_ptrs, state_type
)
)
if dst_item_lens is not None:
_, mapped_dst_lens, _ = self.get_mla_kv_ptrs_with_pp(
item_lens, dst_item_lens, state_type
)
for layer_id in range(layers_current_pp_stage):
if item_lens[layer_id] != mapped_dst_lens[layer_id]:
assert bootstrap_room is not None
failure_reason = (
f"{state_type} item length mismatch for positional "
f"entry {layer_id}: prefill={item_lens[layer_id]} "
f"decode={mapped_dst_lens[layer_id]}"
)
logger.error(failure_reason)
self.conclude_failure(
bootstrap_room=bootstrap_room,
failure_reason=failure_reason,
)
return -1
layers_params = [
(
src_kv_ptrs[layer_id],
@@ -2110,9 +2145,13 @@ class MooncakeKVManager(StagingManagerMixin, CommonKVManager):
),
src_layer_ids=src_state_layer_ids,
dst_layer_ids=dst_state_layer_ids,
dst_item_lens=dst_item_lens,
bootstrap_room=req.room,
)
or rc
)
if self.request_status.get(req.room, KVPoll.Failed) == KVPoll.Failed:
return rc
elif st in (StateType.MINIMAX_INDEX_K, StateType.MINIMAX_DENSE_KV):
# Compacted layer lists require equal TP and PP=1 on both peers.
if self.pp_size is not None and self.pp_size > 1:
@@ -2376,6 +2415,19 @@ class MooncakeKVManager(StagingManagerMixin, CommonKVManager):
self._staging_outstanding[kv_chunk.room] += 1
kv_chunk.staging_counted = True
# Under overlap scheduling an early cached-prefix send can
# precede completion of the previous prefill forward. Wait on
# the worker, not the scheduler, before reading source KV.
# A room that already failed skips the wait; the status check
# below then drops the chunk, including on an abort during it.
if kv_chunk.wait_event is not None:
if (
kv_chunk.room in self.request_status
and self.check_status(kv_chunk.room) != KVPoll.Failed
):
kv_chunk.wait_event.synchronize()
kv_chunk.wait_event = None
if (
kv_chunk.room not in self.request_status
or self.check_status(kv_chunk.room) == KVPoll.Failed
@@ -2588,6 +2640,13 @@ class MooncakeKVManager(StagingManagerMixin, CommonKVManager):
executor,
target_rank_registration_info,
)
if (
self.request_status.get(
kv_chunk.room, KVPoll.Failed
)
== KVPoll.Failed
):
break
if state_rc != 0:
with self.session_lock:
self.session_failures[
@@ -2919,6 +2978,7 @@ class MooncakeKVManager(StagingManagerMixin, CommonKVManager):
state_indices: Optional[List] = None,
num_kv_tokens: Optional[int] = None,
trace_ctx: Optional[Union[TraceReqContext, TraceNullContext]] = None,
wait_event: Optional[object] = None,
):
assert self.disaggregation_mode == DisaggregationMode.PREFILL
assert not is_last_chunk or (is_last_chunk and aux_index is not None)
@@ -2938,12 +2998,22 @@ class MooncakeKVManager(StagingManagerMixin, CommonKVManager):
# add further chunks into the transfer queue.
return
# NOTE(shangming): sharding according to the dst_infos to make sure
# requests with the same dst_sessions will be added into the same
# queue, which enables early abort with failed sessions.
# Shard by dst session ports AND bootstrap_room. The port term alone
# degenerates to a constant when all requests target a single decode
# instance (its session ports are fixed), collapsing every transfer
# onto 1/16 of the lanes and throttling P->D intake (measured at
# client concurrency 300: #transfer-req piled at 100+ while decode
# itself stayed healthy). Mixing in bootstrap_room keeps all chunks
# of one request on the same lane (per-request FIFO order preserved)
# while spreading distinct rooms across all lanes. Abort/cleanup is
# unaffected: chunks are dropped per-room at dequeue via
# request_status (see the worker loop) and per-request via
# failed_sessions before send — no lane-level purge exists, and the
# deferred abort ack is room-scoped (_staging_outstanding). Same
# trade-off as mori/conn.py which shards by bootstrap_room alone.
dst_infos = self.transfer_infos[bootstrap_room].keys()
session_port_sum = sum(int(session.rsplit(":", 1)[1]) for session in dst_infos)
shard_idx = session_port_sum % len(self.transfer_queues)
shard_idx = (session_port_sum + bootstrap_room) % len(self.transfer_queues)
if trace_ctx is None:
trace_ctx = TraceNullContext()
@@ -2958,6 +3028,7 @@ class MooncakeKVManager(StagingManagerMixin, CommonKVManager):
state_indices=state_indices,
num_kv_tokens=num_kv_tokens,
trace_ctx=trace_ctx,
wait_event=wait_event,
)
)
@@ -3064,6 +3135,8 @@ class MooncakeKVSender(MooncakeFailureExceptionMixin, CommonKVSender):
if should_skip:
return
wait_event = getattr(self, "_early_send_wait_event", None)
self._early_send_wait_event = None
if not is_last_chunk:
self.kv_mgr.add_transfer_request(
self.bootstrap_room,
@@ -3072,6 +3145,7 @@ class MooncakeKVSender(MooncakeFailureExceptionMixin, CommonKVSender):
False,
num_kv_tokens=num_kv_tokens,
trace_ctx=self.trace_ctx.copy_for_thread(),
wait_event=wait_event,
)
else:
self.kv_mgr.add_transfer_request(
@@ -3083,6 +3157,7 @@ class MooncakeKVSender(MooncakeFailureExceptionMixin, CommonKVSender):
state_indices=state_indices,
num_kv_tokens=num_kv_tokens,
trace_ctx=self.trace_ctx.copy_for_thread(),
wait_event=wait_event,
)
self._record_transfer_indices(kv_indices, state_indices)
+92 -7
View File
@@ -21,6 +21,8 @@ from __future__ import annotations
import hashlib
import logging
import os
import zlib
from array import array
from collections import deque
from http import HTTPStatus
@@ -105,6 +107,64 @@ logger = logging.getLogger(__name__)
_is_npu = is_npu()
# DEBUG-ONLY (p2 wedge trap 2026-10-13): per-step batch-decision fingerprint
# and event-loop phase logging. Same env as custom_all_reduce_v2 ARDBG.
# 1 = sampled/one-line, 2 = + stacks on ONE_SHOT_PUSH. Off by default.
_AR_SEQ_DEBUG = int(os.environ.get("SGLANG_AR_SEQ_DEBUG", "0"))
_plan_seq = 0
_plan_none_seq = 0
def _dbg_plan(self, batch) -> None:
global _plan_seq, _plan_none_seq
if not _AR_SEQ_DEBUG:
return
_plan_seq += 1
rank = torch.distributed.get_rank()
if batch is None:
# 空转步采样记录,避免 idle 自旋刷爆日志
_plan_none_seq += 1
if _plan_none_seq & 511:
return
logger.info("[ARDBG] plan rank=%s step=%d batch=None", rank, _plan_seq)
return
reqs = batch.reqs
fp = 0
ext_total = 0
for r in reqs:
# crc32 而非 hash():后者按进程随机化种子,跨 rank 不可比
fp = (fp * 131 + zlib.crc32(r.rid.encode())) & 0xFFFFFFFFFFFFFFFF
er = getattr(r, "extend_range", None)
el = int(er.length) if er is not None else -1
fp = (fp * 131 + el) & 0xFFFFFFFFFFFFFFFF
fp = (fp * 131 + int(getattr(r, "prefix_len", -1))) & 0xFFFFFFFFFFFFFFFF
if el > 0:
ext_total += el
logger.info(
"[ARDBG] plan rank=%s step=%d nreq=%d ext=%d fp=%016x",
rank,
_plan_seq,
len(reqs),
ext_total,
fp,
)
def _dbg_phase(label: str, batch=None) -> None:
if not _AR_SEQ_DEBUG:
return
if batch is None and label in ("inflight_enter", "inflight_exit"):
# 空转迭代的 inflight 相位价值低,采样记录
if _plan_seq & 255:
return
logger.info(
"[ARDBG] phase rank=%s step=%s %s nreq=%s",
torch.distributed.get_rank(),
_plan_seq,
label,
len(batch.reqs) if batch is not None else "-",
)
def should_force_retry(req: Req) -> bool:
"""Test hook to force a request into optimistic prefill retry."""
@@ -628,6 +688,7 @@ class SchedulerDisaggregationPrefillMixin:
if batch:
set_schedule_time_batch(batch)
_dbg_plan(self, batch)
return NextBatchPlan(batch_to_run=batch, running_batch=running_batch)
@torch.no_grad()
@@ -658,13 +719,17 @@ class SchedulerDisaggregationPrefillMixin:
if batch:
if self.enable_staging:
self.maybe_prefetch_staging_for_batch(batch)
_dbg_phase("run_batch_enter", batch)
result = self.run_batch(batch)
_dbg_phase("run_batch_exit", batch)
self.process_batch_result(batch, result)
else:
self._sched_idled = True
self.on_idle()
_dbg_phase("inflight_enter")
self.process_disagg_prefill_inflight_queue()
_dbg_phase("inflight_exit")
# Update last_batch
self.last_batch = batch
@@ -698,7 +763,9 @@ class SchedulerDisaggregationPrefillMixin:
if batch:
if self.enable_staging:
self.maybe_prefetch_staging_for_batch(batch)
_dbg_phase("run_batch_enter", batch)
batch_result = self.run_batch(batch)
_dbg_phase("run_batch_exit", batch)
self._apply_war_barrier()
self.result_queue.append((batch.copy(), batch_result))
else:
@@ -713,7 +780,9 @@ class SchedulerDisaggregationPrefillMixin:
# When the server is idle, do self-check and re-init some states
self.on_idle()
_dbg_phase("inflight_enter")
self.process_disagg_prefill_inflight_queue()
_dbg_phase("inflight_exit")
# Run sample of the current batch
# It depends on the result of the last batch (e.g., grammar), so we run it after the last batch is processed.
@@ -851,7 +920,11 @@ class SchedulerDisaggregationPrefillMixin:
advance_logprob_pt(i, req)
continue
maybe_cache_unfinished_req(req, self.tree_cache)
self.checkpoint_disagg_prefill(req)
# G9: the disagg-prefill path bypasses batch_result_processor's
# prefill loop, so record the drain event here — final chunk
# done, request heading to the inflight (KV transfer) queue.
self.record_prefill_drain(req.seqlen)
self.disagg_prefill_inflight_queue.append(req)
if self.spec_algorithm.is_eagle() and draft_input is not None:
req.output_topk_p = draft_input.topk_p[i]
@@ -1208,7 +1281,7 @@ class SchedulerDisaggregationPrefillMixin:
chunked_req_to_exclude = set()
if (req := self.chunked_req) is not None:
chunked_req_to_exclude.add(req)
maybe_cache_unfinished_req(req, self.tree_cache, chunked=True)
self.checkpoint_disagg_prefill(req, chunked=True)
if not self.check_bootstrap(req):
if is_aborted(req):
@@ -1503,10 +1576,21 @@ class SchedulerDisaggregationPrefillMixin:
else:
self.disagg_prefill_pending_chunk_rids.add(req.rid)
def checkpoint_disagg_prefill(
self: Scheduler, req: Req, chunked: bool = False
) -> None:
"""Publish the request's computed prefix to the tree cache.
Name shim for the #42035-era call structure: this branch predates
#41520's ``checkpoint_kv_cache`` / ``insert_req`` cache API, so the
equivalent here is ``maybe_cache_unfinished_req``.
"""
maybe_cache_unfinished_req(req, self.tree_cache, chunked=chunked)
def optimistic_release_and_requeue(self: Scheduler, req: Req) -> None:
"""Release KV cache and requeue an optimistic prefill request."""
max_attempts = get_disagg().optimistic_prefill_attempts
maybe_cache_unfinished_req(req, self.tree_cache)
self.checkpoint_disagg_prefill(req)
# The cached prefix is evictable once the KV is released. Its length
# (capped at what a retry can match) seeds the retry's storage baseline,
# so an evicted prefix is looked up in L3 once before it is recomputed.
@@ -1519,10 +1603,11 @@ class SchedulerDisaggregationPrefillMixin:
)
)
self._release_aborted_request(req)
# Mamba insertion donates the checkpoint and clears its sequence marker.
release_kv_cache(
req, self.tree_cache, is_insert=not self.tree_cache.supports_mamba()
)
# The checkpoint above (maybe_cache_unfinished_req) already handed the
# prefill KV to the tree; the request is not finished, so the release
# only frees the rest. (#42194 semantics; our tree predates #41520's
# mamba-donation rework, mamba models are not run on this branch.)
release_kv_cache(req, self.tree_cache, is_insert=False)
req.reset_for_retract()
req.output_ids = array("q")
req.start_send_idx = 0
+36 -1
View File
@@ -1,5 +1,7 @@
from __future__ import annotations
import logging
import os
import random
from collections import deque
from contextlib import nullcontext
@@ -43,6 +45,32 @@ if is_npu():
DSV4NPUTokenToKVPool,
)
logger = logging.getLogger(__name__)
# DEBUG-ONLY (p2 wedge repro 2026-10-13): env-gated seq logging for host-side
# gloo all-reduce sync points. Mirrors SGLANG_AR_SEQ_DEBUG in
# custom_all_reduce_v2.py. 1 = one-line, 2 = + python stack. Not for prod.
_AR_SEQ_DEBUG = int(os.environ.get("SGLANG_AR_SEQ_DEBUG", "0"))
_host_ar_seq = 0
def _dbg_log_host_ar(phase: str, polls) -> None:
global _host_ar_seq
if not _AR_SEQ_DEBUG:
return
seq = _host_ar_seq
_host_ar_seq += 1
# 空转迭代(无 poller)采样 1/256,避免 idle 自旋刷爆日志
if (not polls) and seq & 255:
return
logger.info(
"[ARDBG] host rank=%s seq=%d phase=%s polls=%s",
dist.get_rank(),
seq,
phase,
polls,
)
#########################
# Constants & Enums
#########################
@@ -213,13 +241,20 @@ def poll_and_all_reduce_attn_cp_tp_group(
attn_cp_cpu_group: dist.ProcessGroup,
attn_tp_cpu_group: dist.ProcessGroup,
):
# DEBUG-ONLY
_dbg_log_host_ar("enter", len(pollers))
# First sync across attn-tp ranks so all TP participants for a given (dp, cp)
# shard observe the same status transitions.
polls = poll_and_all_reduce(pollers, attn_tp_cpu_group)
# DEBUG-ONLY
_dbg_log_host_ar("tp-done", polls)
# Then sync across attn-cp ranks, so all TPxCP participants in one DP shard
# converge to the same global status.
return _all_reduce_polls(polls, attn_cp_cpu_group)
result = _all_reduce_polls(polls, attn_cp_cpu_group)
# DEBUG-ONLY
_dbg_log_host_ar("cp-done", result)
return result
def poll_and_all_reduce_with_staging(
@@ -25,6 +25,8 @@ the kernel captured in the graph dereferences its row at replay time.
"""
import logging
import os
import traceback
from contextlib import contextmanager
from typing import List, NamedTuple, Optional, Tuple
@@ -68,6 +70,9 @@ MB = 1024 * 1024
_ALIGN_BYTES = 1024
_SEMAPHORE_BYTES = 128
_MAX_GRAPH_INPUTS = 131072
# DEBUG-ONLY (p2 wedge repro 2026-10-13): env-gated seq logging for TP device
# all-reduce calls. 1 = one-line per call, 2 = + python stack. Not for prod.
_AR_SEQ_DEBUG = int(os.environ.get("SGLANG_AR_SEQ_DEBUG", "0"))
# resolved once at import time; explicit constructor sizes take precedence
_DEFAULT_MAX_SIZE = envs.SGLANG_CUSTOM_ALL_REDUCE_V2_MAX_SIZE_KB.get() * 1024
# forced per-direction workspace sizes; highest priority, override both the
@@ -297,6 +302,9 @@ class CustomAllReduceV2:
self.pull_enabled,
)
dist.barrier(group=self.group)
# DEBUG-ONLY: per-instance seq counter for SGLANG_AR_SEQ_DEBUG logging
self._dbg_seq = 0
self._dbg_rank = self.rank
# ------------------------------------------------------------------
# Algo selection
@@ -376,6 +384,27 @@ class CustomAllReduceV2:
assert config is not None, f"No config for {nbytes = }"
algo, use_graph, use_multicast = config
graph_params = self._allocate_graph_row(input, nbytes) if use_graph else None
if _AR_SEQ_DEBUG:
seq = self._dbg_seq
self._dbg_seq += 1
# 采样 1/1024:生产高频路径下全量日志会拖垮 P 节点
if seq & 1023 == 0:
logger.info(
"[ARDBG] dev rank=%s seq=%d nbytes=%d algo=%s graph=%s capturing=%s",
self._dbg_rank,
seq,
nbytes,
getattr(algo, "name", algo),
use_graph,
torch.cuda.is_current_stream_capturing(),
)
if _AR_SEQ_DEBUG >= 2 and getattr(algo, "name", algo) == "ONE_SHOT_PUSH":
logger.info(
"[ARDBG] stack rank=%s seq=%d\n%s",
self._dbg_rank,
seq,
"".join(traceback.format_stack(limit=15)),
)
return custom_all_reduce(
self.obj,
input,
@@ -1776,6 +1776,18 @@ async def openai_v1_chat_completions(
)
@app.post("/query_cache_hit", dependencies=[Depends(validate_json_request)])
async def query_cache_hit(request: ChatCompletionRequest, raw_request: Request):
"""G9: probe prefix-cache hit (L1/L2 + L3 storage) for a chat request.
Deployed on decode engines; the PD router calls it before dispatch to
decide conditional bypass. Accepts a normal chat body or `input_ids`.
"""
return await raw_request.app.state.openai_serving_chat.query_cache_hit(
request, raw_request
)
@app.post(
"/v1/embeddings",
response_class=ORJSONResponse,
@@ -1140,6 +1140,73 @@ class OpenAIServingChat(OpenAIServingBase):
return "text", processed_messages.prompt_ids
return "input_ids", processed_messages.prompt_ids
async def query_cache_hit(
self, request: ChatCompletionRequest, raw_request: Request = None
) -> ORJSONResponse:
"""G9 probe: report how much of this chat request's prompt is already
cached (per-rank-best L1/L2 radix + shared L3 storage) without running
any prefill. The router uses it for conditional PD-bypass decisions.
Multimodal requests are rejected (v1): their sentinel-token expansion
costs ~10-100ms per image for only ~2% of traffic; the router skips
them for bypass anyway.
"""
try:
internal_req, _ = self._convert_to_internal_request(request, raw_request)
except Exception as e:
return ORJSONResponse(
status_code=400,
content={"error": f"failed to convert chat request: {e}"},
)
if internal_req.contains_mm_input():
return ORJSONResponse(
status_code=400,
content={"error": "multimodal requests are not supported by /query_cache_hit"},
)
try:
# 只取 input_ids,不走 _tokenize_one_request——它要求 rid_to_state
# 里已有 ReqState(正常生成路径在 tokenize 前注册),纯查询没有。
# dsv41 系 chat encoding 直接渲染 prompt_ids;否则对渲染后的
# text 走 tokenizer(_tokenize_texts 与真实路径同一入口)。
if internal_req.input_ids is not None:
input_ids = list(internal_req.input_ids)
elif internal_req.text:
input_ids, _ = await self.tokenizer_manager._tokenize_texts(
internal_req.text, False
)
else:
return ORJSONResponse(
status_code=400,
content={"error": "request produced neither input_ids nor text"},
)
from sglang.srt.managers.io_struct import QueryCacheHitReqInput
result = await self.tokenizer_manager.query_cache_hit(
QueryCacheHitReqInput(
input_ids=list(input_ids),
extra_key=getattr(internal_req, "extra_key", None),
cache_salt=getattr(internal_req, "cache_salt", None),
)
)
except Exception as e:
return ORJSONResponse(
status_code=500, content={"error": f"cache-hit query failed: {e}"}
)
if result.error is not None:
return ORJSONResponse(status_code=500, content={"error": result.error})
hit = result.l1_hit_tokens + result.l2_hit_tokens + result.l3_hit_tokens
return ORJSONResponse(
content={
"total_tokens": result.total_tokens,
"l1_hit_tokens": result.l1_hit_tokens,
"l2_hit_tokens": result.l2_hit_tokens,
"l3_hit_tokens": result.l3_hit_tokens,
"hit_tokens": hit,
"new_tokens": max(0, result.total_tokens - hit),
"dp_rank": result.dp_rank,
}
)
def _convert_to_internal_request(
self,
request: ChatCompletionRequest,
+10
View File
@@ -758,6 +758,16 @@ class Envs:
SGLANG_HICACHE_HF3FS_CONFIG_PATH = EnvStr(None)
SGLANG_HICACHE_DECODE_OFFLOAD_STRIDE = EnvInt(None)
SGLANG_HICACHE_SKIP_HOST_DUPLICATE_RECLAIM = EnvBool(False)
# Per-stage timing logs for the L3 storage prefetch path (query / transfer /
# ack reduce / drain / graft / load-back). Emits "HC_TIMING ..." lines.
SGLANG_HICACHE_TIMING = EnvBool(False)
# Pages per batched storage IO call (query / get / put chunking).
SGLANG_HICACHE_STORAGE_BATCH_SIZE = EnvInt(128)
# Number of parallel prefetch IO lanes (worker threads) per controller.
# Operations are sharded by crc32(request_id) so every rank maps a request
# to the same lane; each lane owns dedicated completion sync groups.
# MUST be identical across all TP/CP/PP ranks of an instance.
SGLANG_HICACHE_PREFETCH_IO_THREADS = EnvInt(4)
SGLANG_HICACHE_FILE_BACKEND_STORAGE_DIR = EnvStr(None)
# File-backend LRU eviction (opt-in; sizes accept SI/IEC suffixes, "0" disables).
SGLANG_HICACHE_FILE_BACKEND_MAX_SIZE = EnvStr(None)
@@ -206,6 +206,7 @@ def _create_flashmla_metadata():
# only for that shape and go stale silently if FlashMLA retunes it.
_FLASHMLA_SCHED_BLOCK_SIZE_N = 64
_FLASHMLA_SCHED_FIXED_OVERHEAD = 5
_FLASHMLA_SCHED_MAX_SHARED_BYTES = 48 << 10
@functools.lru_cache(maxsize=None)
@@ -242,6 +243,10 @@ def _maybe_precompute_flashmla_sched_meta(
b, s_q = q.shape[0], q.shape[1]
num_sm_parts = max(_num_sms(q.device.index) // s_q, 1)
# Skip when shared memory exceeds the limit.
smem_bytes = 4 * (b * 5 + 1 + num_sm_parts * 8)
if smem_bytes > _FLASHMLA_SCHED_MAX_SHARED_BYTES:
return
meta = torch.empty((num_sm_parts, META_INTS), dtype=torch.int32, device=q.device)
num_splits = torch.empty((b + 1,), dtype=torch.int32, device=q.device)
flashmla_sched_meta(
@@ -2543,7 +2548,7 @@ class DeepseekV4AttnBackend(
req_to_token=self.req_to_token,
full_to_swa=self.token_to_kv_pool.full_to_swa_index_mapping,
swa_window_size=SWA_WINDOW,
swa_page_size=self.token_to_kv_pool.swa_kv_pool.page_size,
swa_page_size=self.token_to_kv_pool.get_swa_key_page_size(),
num_qo_tokens=num_qo_tokens,
max_seq_len=max(seq_lens_cpu_list),
total_swa=total_swa,
@@ -2930,6 +2935,22 @@ class DeepseekV4AttnBackend(
self._low_ratio_compress_fused(
layer, x, req, pos, draft_len=self.speculative_num_draft_tokens
)
elif (
layer.compressor.use_fused_compress
and forward_batch.forward_mode.is_extend_without_speculative()
and forward_batch.extend_start_loc is not None
):
# DP padding extends extend_seq_lens but not extend_start_loc, whose
# length is the request count; padded rows belong to no request.
start_loc = forward_batch.extend_start_loc
seq_lens = forward_batch.extend_seq_lens[: start_loc.shape[0]]
self._low_ratio_compress_fused(
layer,
x,
req,
pos,
extend_offsets=(start_loc.to(torch.int32), seq_lens.to(torch.int32)),
)
else:
self._low_ratio_compress_torch(
layer,
@@ -3001,13 +3022,18 @@ class DeepseekV4AttnBackend(
fuse_index_store=_is_sm100_or_newer(),
)
def _low_ratio_compress_fused(self, layer, x, req, pos, *, draft_len=1) -> None:
def _low_ratio_compress_fused(
self, layer, x, req, pos, *, draft_len=1, extend_offsets=None
) -> None:
"""Decode or target-verify; an extend batch passes ``extend_offsets``, the int32
(first row, row count) of each request."""
from sglang.kernels.ops.attention.dsv4.fp4_indexer_rope import (
index_k_norm_rope_pack_store,
)
from sglang.kernels.ops.attention.dsv4.low_ratio_compress import (
c1_decode_norm_rope_store,
c2_decode_norm_rope_store,
c2_prefill_norm_rope_store,
)
pool = self.token_to_kv_pool
@@ -3040,7 +3066,7 @@ class DeepseekV4AttnBackend(
# CompressStatePool stores each request's pending pairs in a position ring.
# KVAndScore rows use | kv | score |, addressed as req * ring_size + pos % ring_size.
state = pool.get_attention_compress_states(layer_id)
latent = c2_decode_norm_rope_store(
args = (
compressor.project_fused(x),
state.kv_score_buffer.kv_score,
compressor.norm.weight.data,
@@ -3050,11 +3076,23 @@ class DeepseekV4AttnBackend(
compressor.norm.eps,
freqs_cis,
kv_cache,
page_size=page_size,
ring_size=state.ring_size,
draft_len=draft_len,
layout=kv_layout,
)
if extend_offsets is None:
latent = c2_decode_norm_rope_store(
*args,
page_size=page_size,
ring_size=state.ring_size,
draft_len=draft_len,
layout=kv_layout,
)
else:
latent = c2_prefill_norm_rope_store(
*args,
*extend_offsets,
page_size=page_size,
ring_size=state.ring_size,
layout=kv_layout,
)
out_loc = core.c2_out_loc
indexer = layer.indexer
@@ -3760,7 +3798,7 @@ class DeepseekV4AttnBackend(
compress_ratio
)
swa_kv_page_size = token_to_kv_pool.swa_kv_pool.page_size
swa_kv_page_size = token_to_kv_pool.get_swa_key_page_size()
assert swa_k_cache.ndim == 2
# The kernel detects each cache's format from the last dim of this
# view: 584 (V4), 528 (V4.1 fp8) or 288 (V4.1 fp4, extra cache only).
@@ -1472,9 +1472,10 @@ class IndexerKPool(MultiPlatformOp):
if metadata is None:
return None
assert forward_batch.seq_lens_cpu is not None
# Spec-v2 decode/verify batches may carry no CPU seq_lens mirror (the
# backend opts out of the D2H); test emptiness on the device shape.
mode = forward_batch.forward_mode
if mode.is_idle() or len(forward_batch.seq_lens_cpu) == 0:
if mode.is_idle() or forward_batch.seq_lens.shape[0] == 0:
return torch.full(
(x.shape[0], self.index_topk + self.index_kpool - 1),
-1,
@@ -306,7 +306,8 @@ class DeepseekSparseAttnBackend(
extend_dummy_seqs_capped_by_req_pool: bool = True
# Decode/verify/draft graph replay rebuilds metadata from static buffers
# (page-table width) and never reads seq_lens_cpu / seq_lens_sum; opt out of
# the D2H sync. The eager fallback derives lengths from GPU seq_lens.
# the D2H sync. The eager fallback derives lengths from GPU seq_lens, and the
# KPool indexer reads the host mirror only for extend, which always has it.
needs_cpu_seq_lens: bool = False
# init_cuda_graph_state sizes this for every backend, but only the TRT-LLM
# branch of __init__ allocates one.
@@ -337,7 +338,6 @@ class DeepseekSparseAttnBackend(
)
self.dsa_index_topk = get_dsa_index_topk(hf_config)
self.dsa_index_kpool = get_dsa_index_kpool(hf_config)
self.needs_cpu_seq_lens = self.dsa_index_kpool > 1
self._init_kpool_metadata_fusion()
self.max_context_len = model_runner.model_config.context_len
self.num_q_heads = (
@@ -136,8 +136,8 @@ def combine_topk_swa_indices(
Returns int32 indices [num_tokens, padded_topk_swa] and per-token scanned-prefix
lengths, including -1 entries skipped by attention. Width is padded to 128.
Preallocated out_indices must contain -1 outside the written prefix;
reuse requires chunk-invariant scanned-prefix lengths. out_lens is overwritten.
Rows are written whole and -1 padded, so preallocated out_indices / out_lens
are overwritten entirely; rows past the last request are -1 with length 0.
"""
assert topk_indices.dtype == torch.int32
assert query_start_loc.dtype == torch.int32
@@ -155,9 +155,9 @@ def combine_topk_swa_indices(
num_reqs = seq_lens.shape[0]
combined_topk = combined_topk_width(topk, window_size)
if out_indices is None:
combined_indices = torch.full(
# The kernel writes whole -1 padded rows.
combined_indices = torch.empty(
(num_tokens, combined_topk),
-1,
dtype=torch.int32,
device=topk_indices.device,
)
@@ -166,7 +166,7 @@ def combine_topk_swa_indices(
assert out_indices.dtype == torch.int32
combined_indices = out_indices
if out_lens is None:
combined_lens = torch.zeros(
combined_lens = torch.empty(
num_tokens, dtype=torch.int32, device=topk_indices.device
)
else:
@@ -174,8 +174,10 @@ def combine_topk_swa_indices(
assert out_lens.dtype == torch.int32
combined_lens = out_lens
NUM_WORKERS = 128
_combine_topk_swa_indices_kernel[(num_reqs, NUM_WORKERS)](
if num_tokens == 0:
return combined_indices, combined_lens
block_t = 2 # Measured on B200 with num_warps=2, 16384 tokens, bs 1 to 64.
_combine_topk_swa_indices_kernel[(triton.cdiv(num_tokens, block_t),)](
combined_indices,
combined_indices.stride(0),
combined_lens,
@@ -187,10 +189,16 @@ def combine_topk_swa_indices(
gather_lens,
compressed_base,
swa_base,
num_tokens,
num_reqs,
top_k=topk,
COMPRESS_RATIO=compress_ratio,
WINDOW_SIZE=window_size,
PADDED_TOP_K=triton.next_power_of_2(topk_indices.shape[-1]),
WIDTH=combined_topk,
PADDED_WIDTH=triton.next_power_of_2(combined_topk),
BLOCK_T=block_t,
SEARCH_STEPS=max(num_reqs, 1).bit_length(),
num_warps=2,
)
return combined_indices, combined_lens
+5
View File
@@ -239,6 +239,11 @@ class EngramHasher(nn.Module):
self.image_token_id: Optional[int] = None
self.history: Optional[torch.Tensor] = None
self.pad_row = 0
# typing only
self.token_map: torch.Tensor
self.multipliers: torch.Tensor
self.primes: torch.Tensor
self.offsets: torch.Tensor
def init_history(self, num_req_slots: int, device) -> None:
"""Allocate oldest-first history with a spare row for graph padding."""
+181 -53
View File
@@ -17,6 +17,7 @@ limitations under the License.
import logging
import threading
import time
import zlib
from dataclasses import dataclass
from queue import Empty, Queue
from typing import TYPE_CHECKING, Callable, List, NamedTuple, Optional
@@ -32,6 +33,17 @@ from sglang.srt.mem_cache.hicache_storage import (
count_pool_hits,
)
from sglang.srt.environ import envs as _envs
_HC_TIMING = None
def _hc_timing_enabled() -> bool:
global _HC_TIMING
if _HC_TIMING is None:
_HC_TIMING = _envs.SGLANG_HICACHE_TIMING.get()
return _HC_TIMING
if TYPE_CHECKING:
from sglang.srt.mem_cache.allocator import BaseTokenToKVPoolAllocator
from sglang.srt.mem_cache.pool_host import HostKVCache
@@ -306,7 +318,18 @@ class HiCacheController:
self.attn_tp_group = attn_tp_group
self.pp_group = pp_group
self.prefetch_hits_sync_groups: List[torch.distributed.ProcessGroup] = []
self.prefetch_completion_sync_groups: List[torch.distributed.ProcessGroup] = []
# One set of completion sync groups per prefetch IO lane. Each lane's
# worker/sync threads form a serial pipeline, so the collective
# sequence on a lane's groups is identical across ranks as long as
# lane assignment is deterministic (see dispatch_prefetch).
self.prefetch_completion_sync_groups_by_lane: List[
List[torch.distributed.ProcessGroup]
] = []
# Read once at construction; attach/reset reuse the stored value so
# lane count never changes underneath live sync groups.
self.num_prefetch_io_lanes = max(
1, _envs.SGLANG_HICACHE_PREFETCH_IO_THREADS.get()
)
self.mem_pool_device_allocator = token_to_kv_pool_allocator
mem_pool_device = token_to_kv_pool_allocator.get_kvcache()
from sglang.srt.mem_cache.memory_pool import HybridLinearKVPool
@@ -335,8 +358,8 @@ class HiCacheController:
self.storage_stop_event = threading.Event()
# Storage control queues, (re)created whenever the storage threads start.
self.prefetch_buffer: Optional[Queue[PrefetchOperation]] = None
self.prefetch_sync_queue: Optional[Queue[PrefetchAck]] = None
self.prefetch_io_queues: List[Queue[PrefetchOperation]] = []
self.prefetch_sync_queues: List[Queue[PrefetchAck]] = []
self.prefetch_hit_queue: Optional[Queue[StorageOperation]] = None
self.ack_prefetch_queue = Queue[PrefetchAck]()
self.ack_backup_queue: Optional[Queue[StorageOperation]] = None
@@ -444,27 +467,39 @@ class HiCacheController:
self.prefetch_thread = threading.Thread(
target=self.prefetch_thread_func, daemon=True
)
self.prefetch_io_aux_thread = threading.Thread(
target=self.prefetch_io_aux_func, daemon=True
)
self.prefetch_sync_thread = threading.Thread(
target=self.prefetch_sync_thread_func, daemon=True
)
self.prefetch_io_aux_threads = [
threading.Thread(
target=self.prefetch_io_aux_func, args=(lane,), daemon=True
)
for lane in range(self.num_prefetch_io_lanes)
]
self.prefetch_sync_threads = [
threading.Thread(
target=self.prefetch_sync_thread_func, args=(lane,), daemon=True
)
for lane in range(self.num_prefetch_io_lanes)
]
self.backup_thread = threading.Thread(
target=self.backup_thread_func, daemon=True
)
self.prefetch_queue = Queue()
self.backup_queue = Queue()
self.prefetch_buffer = Queue()
self.prefetch_sync_queue = Queue()
self.prefetch_io_queues = [
Queue() for _ in range(self.num_prefetch_io_lanes)
]
self.prefetch_sync_queues = [
Queue() for _ in range(self.num_prefetch_io_lanes)
]
self.prefetch_hit_queue = Queue()
self.ack_prefetch_queue = Queue()
self.ack_backup_queue = Queue()
self.host_mem_release_queue = Queue()
self.prefetch_thread.start()
self.prefetch_io_aux_thread.start()
self.prefetch_sync_thread.start()
for thread in self.prefetch_io_aux_threads:
thread.start()
for thread in self.prefetch_sync_threads:
thread.start()
self.backup_thread.start()
def _stop_storage_threads(self):
@@ -485,10 +520,10 @@ class HiCacheController:
self.prefetch_queue.put_nowait(None)
if hasattr(self, "backup_queue"):
self.backup_queue.put_nowait(None)
if hasattr(self, "prefetch_buffer"):
self.prefetch_buffer.put_nowait(None)
if hasattr(self, "prefetch_sync_queue"):
self.prefetch_sync_queue.put_nowait(None)
for queue in getattr(self, "prefetch_io_queues", []) or []:
queue.put_nowait(None)
for queue in getattr(self, "prefetch_sync_queues", []) or []:
queue.put_nowait(None)
except Exception:
pass
@@ -498,10 +533,8 @@ class HiCacheController:
threads.append(self.prefetch_thread)
if hasattr(self, "backup_thread"):
threads.append(self.backup_thread)
if hasattr(self, "prefetch_io_aux_thread"):
threads.append(self.prefetch_io_aux_thread)
if hasattr(self, "prefetch_sync_thread"):
threads.append(self.prefetch_sync_thread)
threads.extend(getattr(self, "prefetch_io_aux_threads", []) or [])
threads.extend(getattr(self, "prefetch_sync_threads", []) or [])
for t in threads:
try:
@@ -585,7 +618,13 @@ class HiCacheController:
# Use dedicated gloo groups so storage prefetch sync is isolated
# from other collectives and consistent across CPxTP participants.
self.prefetch_hits_sync_groups = self._create_sync_groups()
self.prefetch_completion_sync_groups = self._create_sync_groups()
# Each IO lane gets its own completion groups: per-lane collectives
# stay ordered because a lane is a serial pipeline and dispatch is
# deterministic (crc32(request_id) % num_prefetch_io_lanes).
self.prefetch_completion_sync_groups_by_lane = [
self._create_sync_groups()
for _ in range(self.num_prefetch_io_lanes)
]
# Select the get and set functions
self.page_get_func = self._generic_page_get
@@ -620,9 +659,10 @@ class HiCacheController:
except Exception:
pass
self._destroy_sync_groups(self.prefetch_hits_sync_groups)
self._destroy_sync_groups(self.prefetch_completion_sync_groups)
for lane_groups in self.prefetch_completion_sync_groups_by_lane:
self._destroy_sync_groups(lane_groups)
self.prefetch_hits_sync_groups = []
self.prefetch_completion_sync_groups = []
self.prefetch_completion_sync_groups_by_lane = []
try:
if (
hasattr(self, "storage_backend")
@@ -660,11 +700,11 @@ class HiCacheController:
raise RuntimeError("Stop storage threads failed; detach aborted.") from e
# Best-effort destroy process groups created for storage ops.
self._destroy_sync_groups(
self.prefetch_hits_sync_groups + self.prefetch_completion_sync_groups
)
self._destroy_sync_groups(self.prefetch_hits_sync_groups)
for lane_groups in self.prefetch_completion_sync_groups_by_lane:
self._destroy_sync_groups(lane_groups)
self.prefetch_hits_sync_groups = []
self.prefetch_completion_sync_groups = []
self.prefetch_completion_sync_groups_by_lane = []
# Best-effort close (some backends rely on GC/destructor).
try:
@@ -758,13 +798,17 @@ class HiCacheController:
self.ack_load_queue.clear()
if self.enable_storage:
self.prefetch_thread.join()
self.prefetch_io_aux_thread.join()
self.prefetch_sync_thread.join()
for thread in self.prefetch_io_aux_threads:
thread.join()
for thread in self.prefetch_sync_threads:
thread.join()
self.backup_thread.join()
self.prefetch_queue.queue.clear()
self.backup_queue.queue.clear()
self.prefetch_buffer.queue.clear()
self.prefetch_sync_queue.queue.clear()
for queue in self.prefetch_io_queues:
queue.queue.clear()
for queue in self.prefetch_sync_queues:
queue.queue.clear()
self.prefetch_hit_queue.queue.clear()
self.ack_prefetch_queue.queue.clear()
self.ack_backup_queue.queue.clear()
@@ -777,18 +821,26 @@ class HiCacheController:
self.prefetch_thread = threading.Thread(
target=self.prefetch_thread_func, daemon=True
)
self.prefetch_io_aux_thread = threading.Thread(
target=self.prefetch_io_aux_func, daemon=True
)
self.prefetch_sync_thread = threading.Thread(
target=self.prefetch_sync_thread_func, daemon=True
)
self.prefetch_io_aux_threads = [
threading.Thread(
target=self.prefetch_io_aux_func, args=(lane,), daemon=True
)
for lane in range(self.num_prefetch_io_lanes)
]
self.prefetch_sync_threads = [
threading.Thread(
target=self.prefetch_sync_thread_func, args=(lane,), daemon=True
)
for lane in range(self.num_prefetch_io_lanes)
]
self.backup_thread = threading.Thread(
target=self.backup_thread_func, daemon=True
)
self.prefetch_thread.start()
self.prefetch_io_aux_thread.start()
self.prefetch_sync_thread.start()
for thread in self.prefetch_io_aux_threads:
thread.start()
for thread in self.prefetch_sync_threads:
thread.start()
self.backup_thread.start()
def has_inflight_device_transfers(self) -> bool:
@@ -1075,7 +1127,23 @@ class HiCacheController:
count += 1
return count
def _page_transfer(self, operation: PrefetchOperation) -> int:
def dispatch_prefetch(self, operation: PrefetchOperation) -> None:
"""Route a prefetch operation to its IO lane.
Lane assignment must be deterministic and identical on every rank:
each lane owns dedicated completion sync groups, and the collective
sequence on a lane is the filtered dispatch order. crc32 (not the
per-process salted builtin hash) keeps the mapping stable across
processes. All chunks of one operation stay in one lane, preserving
the per-operation ack order the sync thread reduces over.
"""
lane = (
zlib.crc32(str(operation.request_id).encode())
% self.num_prefetch_io_lanes
)
self.prefetch_io_queues[lane].put(operation)
def _page_transfer(self, operation: PrefetchOperation, sync_queue: Queue) -> int:
# Transfer batch by batch
prefix_keys = operation.prefix_keys
kv_derived_transfers = [
@@ -1087,8 +1155,8 @@ class HiCacheController:
completed_pages = 0
for i in range(0, len(operation.hash_value), STORAGE_BATCH_SIZE):
# When an error is occurred, we should keep looping and produce the same number of
# PrefetchAck as other ranks do, because prefetch_sync_thread (i.e. consumer of
# prefetch_sync_queue) perform reduce on the results. This is so tricky.
# PrefetchAck as other ranks do, because the prefetch sync thread (i.e. consumer of
# the lane sync queue) perform reduce on the results. This is so tricky.
if all_success and operation.is_terminated():
all_success = False
if all_success:
@@ -1100,6 +1168,7 @@ class HiCacheController:
# Get one batch token, and update the completed_tokens if succeed
extra_info = HiCacheStorageExtraInfo(prefix_keys=prefix_keys)
_hc_tb = time.monotonic() if _hc_timing_enabled() else 0.0
hit_pages = self._page_transfer_kv_batch(
operation,
batch_hashes,
@@ -1107,6 +1176,14 @@ class HiCacheController:
extra_info,
kv_derived_transfers,
)
if _hc_timing_enabled():
logger.info(
"HC_TIMING stage=xfer_batch req=%s keys=%d ms=%.1f hit=%d",
operation.request_id,
len(batch_hashes),
(time.monotonic() - _hc_tb) * 1e3,
hit_pages,
)
# Check termination
if hit_pages != len(batch_hashes):
all_success = False
@@ -1118,7 +1195,7 @@ class HiCacheController:
completed_tokens=completed_pages * self.page_size,
operation=operation,
)
self.prefetch_sync_queue.put(ack)
sync_queue.put(ack)
return completed_pages
def _page_transfer_kv_batch(
@@ -1160,18 +1237,33 @@ class HiCacheController:
# Clamp to minimal number of hits.
return min([kv_hits, *sidecar_hits.values()])
def prefetch_io_aux_func(self):
def prefetch_io_aux_func(self, lane: int = 0):
"""
Auxiliary function conducting IO operations for prefetching.
Each lane has its own worker thread and queues; an operation's whole
chunk sequence stays on one lane so the sync thread sees it in order.
"""
io_queue = self.prefetch_io_queues[lane]
sync_queue = self.prefetch_sync_queues[lane]
while not self.storage_stop_event.is_set():
try:
operation = self.prefetch_buffer.get(block=True, timeout=1)
operation = io_queue.get(block=True, timeout=1)
if operation is None:
continue
self._page_transfer(operation)
_hc_t0 = time.monotonic() if _hc_timing_enabled() else 0.0
self._page_transfer(operation, sync_queue)
if _hc_timing_enabled():
logger.info(
"HC_TIMING stage=xfer req=%s lane=%d wait_ms=%.1f xfer_ms=%.1f",
operation.request_id,
lane,
(_hc_t0 - getattr(operation, "query_done_ts", operation.start_time))
* 1e3,
(time.monotonic() - _hc_t0) * 1e3,
)
self.prefetch_sync_queue.put(
sync_queue.put(
PrefetchAck(
rid=operation.request_id,
completed_req=True,
@@ -1216,7 +1308,16 @@ class HiCacheController:
for start in range(0, len(page_hashes), STORAGE_BATCH_SIZE):
batch_hashes = page_hashes[start : start + STORAGE_BATCH_SIZE]
extra_info = HiCacheStorageExtraInfo(prefix_keys=prefix_keys)
_hc_tb = time.monotonic() if _hc_timing_enabled() else 0.0
hit_page_num = self.storage_backend.batch_exists(batch_hashes, extra_info)
if _hc_timing_enabled():
logger.info(
"HC_TIMING stage=query_batch req=%s keys=%d ms=%.1f hit=%d",
operation.request_id,
len(batch_hashes),
(time.monotonic() - _hc_tb) * 1e3,
hit_page_num,
)
hash_value.extend(batch_hashes[:hit_page_num])
storage_query_count += hit_page_num * self.page_size
if hit_page_num < len(batch_hashes):
@@ -1235,10 +1336,12 @@ class HiCacheController:
operation = self.prefetch_queue.get(block=True, timeout=1)
if operation is None:
continue
_hc_t0 = time.monotonic() if _hc_timing_enabled() else 0.0
if operation.is_terminated():
hash_value, storage_hit_count = [], 0
else:
hash_value, storage_hit_count = self._storage_hit_query(operation)
_hc_t1 = time.monotonic() if _hc_timing_enabled() else 0.0
storage_hit_count_tensor = torch.tensor(
storage_hit_count, dtype=torch.int
)
@@ -1248,6 +1351,18 @@ class HiCacheController:
self.prefetch_hits_sync_groups,
)
storage_hit_count = storage_hit_count_tensor.item()
if _hc_timing_enabled():
_hc_t2 = time.monotonic()
operation.query_done_ts = _hc_t2
logger.info(
"HC_TIMING stage=query req=%s wait_ms=%.1f query_ms=%.1f "
"hit_reduce_ms=%.1f hit_tokens=%d",
operation.request_id,
(_hc_t0 - operation.start_time) * 1e3,
(_hc_t1 - _hc_t0) * 1e3,
(_hc_t2 - _hc_t1) * 1e3,
storage_hit_count,
)
# Record the TP-synced hit count; the scheduler thread decides
# at drain time whether to revoke (below threshold) or allocate.
@@ -1329,28 +1444,41 @@ class HiCacheController:
except Empty:
continue
def prefetch_sync_thread_func(self):
def prefetch_sync_thread_func(self, lane: int = 0):
"""Synchronize prefetch results across all PP and TP ranks."""
sync_queue = self.prefetch_sync_queues[lane]
groups = self.prefetch_completion_sync_groups_by_lane[lane]
while not self.storage_stop_event.is_set():
try:
ack = self.prefetch_sync_queue.get(block=True, timeout=1)
ack = sync_queue.get(block=True, timeout=1)
if ack is None:
continue
self._reduce_prefetch_ack(ack)
self._reduce_prefetch_ack(ack, groups)
if _hc_timing_enabled() and ack.completed_req:
ack.operation.final_ack_enqueue_ts = time.monotonic()
self.ack_prefetch_queue.put(ack)
except Empty:
continue
def _reduce_prefetch_ack(self, ack: PrefetchAck) -> None:
def _reduce_prefetch_ack(
self, ack: PrefetchAck, groups: List[torch.distributed.ProcessGroup]
) -> None:
"""Synchronize all ranks to agree on a PrefetchAck."""
if ack.completed_tokens is not None:
# Determine the minimal successful prefix of tokens.
_hc_t0 = time.monotonic() if _hc_timing_enabled() else 0.0
completed_tokens_tensor = torch.tensor(
ack.completed_tokens, dtype=torch.int
)
self._all_reduce(
completed_tokens_tensor,
torch.distributed.ReduceOp.MIN,
self.prefetch_completion_sync_groups,
groups,
)
ack.completed_tokens = completed_tokens_tensor.item()
if _hc_timing_enabled():
logger.info(
"HC_TIMING stage=ack_reduce req=%s ms=%.1f",
ack.operation.request_id,
(time.monotonic() - _hc_t0) * 1e3,
)
+21
View File
@@ -2141,6 +2141,27 @@ class SetInternalStateReqOutput(BaseReq, kw_only=True):
updated: bool
class QueryCacheHitReqInput(BaseReq, kw_only=True):
"""G9: read-only probe of prefix-cache hit for a token sequence.
Sent to every DP rank of a (decode) engine; each rank answers with its own
L1/L2 radix match plus the shared L3 storage hit continuing from it.
"""
input_ids: List[int]
extra_key: Optional[str] = None
cache_salt: Optional[str] = None
class QueryCacheHitReqOutput(BaseReq, kw_only=True):
total_tokens: int = 0
l1_hit_tokens: int = 0
l2_hit_tokens: int = 0
l3_hit_tokens: int = 0
dp_rank: int = 0
error: Optional[str] = None
class ProfileReqType(Enum):
START_PROFILE = 1
STOP_PROFILE = 2
@@ -197,6 +197,12 @@ class LoadSnapshot(msgspec.Struct, omit_defaults=True):
num_running_reqs: int = 0
num_waiting_reqs: int = 0
num_waiting_uncached_tokens: int = 0
# G9 queue-time estimation (总输入长度口径, self-consistent numerator/denominator):
# sum of total input lengths (req.seqlen) over all pending queues, and the
# self-learned drain rate in the same unit — completed prefills' total
# input tokens / wall second over a sliding window.
queued_total_input_tokens: int = 0
drain_rate_tokens_per_s: float = 0.0
num_used_tokens: int = 0
num_total_tokens: int = 0
# num_total_tokens minus tokens still awaiting a KV transfer (equal to it
+6 -6
View File
@@ -1034,12 +1034,12 @@ class Req(ReqDllmMixin):
else self.origin_input_ids
) # Before image padding
# Each decode stage's output ids. Append-only by contract:
# _refresh_fill_ids infers how many output tokens are already in
# refresh_fill_ids infers how many output tokens are already in
# full_untruncated_fill_ids from lengths alone, so in-place rewrites
# that preserve length would silently corrupt fill_ids.
self.output_ids = array("q")
# Full untruncated sequence: origin + output (+ DLLM mask block).
# Kept in sync by _refresh_fill_ids; admission only updates
# Kept in sync by refresh_fill_ids; admission only updates
# extend_range, never mutates this array's length.
self.full_untruncated_fill_ids = array("q")
self.extend_range: Optional[Range] = None
@@ -1533,7 +1533,7 @@ class Req(ReqDllmMixin):
def get_fill_ids(self) -> array:
return self.full_untruncated_fill_ids[: self.extend_range.end]
def _refresh_fill_ids(self) -> None:
def refresh_fill_ids(self) -> None:
"""Keep full_untruncated_fill_ids == origin_input_ids + output_ids by
appending only the new output tokens.
@@ -1562,7 +1562,7 @@ class Req(ReqDllmMixin):
self._init_fill_ids_for_dllm()
self.determine_dllm_phase()
else:
self._refresh_fill_ids()
self.refresh_fill_ids()
input_len = len(self.full_untruncated_fill_ids)
@@ -3107,7 +3107,7 @@ class ScheduleBatch(ScheduleBatchDisaggregationDecodeMixin):
for req, prefix_len in zip(
running_batch.reqs, running_prefix_lens, strict=True
):
req._refresh_fill_ids()
req.refresh_fill_ids()
req.set_extend_range(prefix_len, prefix_len + 1)
# Decode tokens of the running portion live in future_map.output_tokens_buf.
@@ -3185,7 +3185,7 @@ class ScheduleBatch(ScheduleBatchDisaggregationDecodeMixin):
# cannot stand in for it. Rows past bs are a beam tail, not requests.
seq_lens = self.seq_lens_cpu[:bs].tolist()
for req, seq_len in zip(self.reqs, seq_lens, strict=True):
req._refresh_fill_ids()
req.refresh_fill_ids()
# end runs one past full_untruncated_fill_ids while output_ids
# trails; safe only while decoding_reqs suppresses cache_unfinished_req.
req.set_extend_range(seq_len - 1, seq_len)
+116
View File
@@ -161,6 +161,8 @@ from sglang.srt.managers.io_struct import (
PauseGenerationReqInput,
PdRoleSwitchReqInput,
ProfileReq,
QueryCacheHitReqInput,
QueryCacheHitReqOutput,
ReleaseMemoryOccupationReqInput,
RemoveExternalCorpusReqInput,
RemoveExternalCorpusReqOutput,
@@ -1806,6 +1808,7 @@ class Scheduler(
(ShutdownReq, self.handle_shutdown),
(GetInternalStateReq, self.get_internal_state),
(SetInternalStateReq, self.set_internal_state),
(QueryCacheHitReqInput, self.handle_query_cache_hit),
(RpcReqInput, self.handle_rpc_request),
(ExpertDistributionReq, self.expert_distribution_handle),
(LoadLoRAAdapterReqInput, self.load_lora_adapter),
@@ -2431,6 +2434,11 @@ class Scheduler(
self._prev_step: Optional[Tuple[int, float, bool]] = None
self._prev_prefill_end_ts: Optional[float] = None
self._sched_idled = False
# G9: sliding-window drain events (monotonic_ts, total_input_tokens) for
# the queue-time estimator. Recorded once per request when its final
# prefill chunk completes (batch_result_processor), in the same
# total-input-length unit as LoadSnapshot.queued_total_input_tokens.
self.prefill_drain_window: deque = deque()
self.load_inquirer = SchedulerLoadInquirer(
disaggregation_mode=self.disaggregation_mode,
server_args=self.server_args,
@@ -2469,8 +2477,35 @@ class Scheduler(
),
get_total_prefill_busy_us=lambda: self.total_prefill_busy_us,
get_decode_moment_totals=lambda: self.decode_moment_totals,
get_prefill_drain_rate=self.get_prefill_drain_rate,
)
PREFILL_DRAIN_WINDOW_S = 60.0
def record_prefill_drain(self, total_input_tokens: int) -> None:
"""G9: one drain event per request whose final prefill chunk completed.
Tokens are the request's total input length (seqlen), matching the
queued_total_input_tokens numerator so the router's
wait = queued_tokens / drain_rate is unit-consistent (决策②).
"""
now = time.monotonic()
self.prefill_drain_window.append((now, total_input_tokens))
cutoff = now - self.PREFILL_DRAIN_WINDOW_S
while self.prefill_drain_window and self.prefill_drain_window[0][0] < cutoff:
self.prefill_drain_window.popleft()
def get_prefill_drain_rate(self) -> float:
"""Completed prefill input tokens per wall second over the window."""
now = time.monotonic()
cutoff = now - self.PREFILL_DRAIN_WINDOW_S
while self.prefill_drain_window and self.prefill_drain_window[0][0] < cutoff:
self.prefill_drain_window.popleft()
if not self.prefill_drain_window:
return 0.0
span = max(now - self.prefill_drain_window[0][0], 1.0)
return sum(tokens for _, tokens in self.prefill_drain_window) / span
def init_output_streamer(self) -> None:
self.output_streamer = self.get_output_streamer_class()(
send_to_detokenizer=self.ipc_channels.send_to_detokenizer,
@@ -2521,6 +2556,7 @@ class Scheduler(
output_streamer=self.output_streamer,
beam_coordinator=self.beam_coordinator,
abort_request=self.abort_request,
record_prefill_drain=self.record_prefill_drain,
)
def init_req_max_new_tokens(self, req):
@@ -5215,6 +5251,86 @@ class Scheduler(
success = False
return success
def handle_query_cache_hit(self, recv_req: QueryCacheHitReqInput):
"""G9: read-only prefix-hit probe (L1/L2 radix + L3 storage).
The router calls this (via /query_cache_hit on the decode engine) to
decide conditional PD-bypass. Runs inside the scheduler event loop so
radix/storage state is consistent, and reuses the exact admission-time
match + L3 query logic (mirrors DecodeHiCachePreallocMixin
._build_decode_prefix_match) without creating a Req or mutating
request state. Node access metadata (LRU) may be touched; a hit found
here can still be evicted before the real request arrives (inherent
TOCTOU, acceptable for an estimate).
"""
from sglang.srt.mem_cache.base_prefix_cache import MatchPrefixParams
from sglang.srt.mem_cache.radix_cache import RadixKey
total = len(recv_req.input_ids)
# 树节点键是 array("q"),RadixKey.match_at 断言两侧类型一致——
# 查询侧必须同样用 array,否则树非空时匹配直接炸(2026-10-11 实测)。
from array import array
query_ids = array("q", recv_req.input_ids)
try:
match_result = self.tree_cache.match_prefix(
MatchPrefixParams(
key=RadixKey(
token_ids=query_ids,
extra_key=recv_req.extra_key,
cache_salt=recv_req.cache_salt,
)
)
)
l1_hit = len(match_result.device_indices)
l2_hit = getattr(match_result, "host_hit_length", 0) or 0
l3_hit = 0
last_host_node = getattr(match_result, "last_host_node", None)
# L3 查询条件:本调度器挂了 hicache 存储后端即可(P 节点 HiCache
# write_through/write_back 场景),不必是 decode-hicache 模式。
# SWA 模型的 decode radix 与 hicache 互斥(kv_cache_builder 硬限制),
# 若只认 enable_decode_hicache,P 节点永远查不到 L3。
can_query_l3 = getattr(self, "enable_decode_hicache", False) or getattr(
self, "enable_hicache_storage", False
)
if (
can_query_l3
and last_host_node is not None
and hasattr(self.tree_cache, "query_storage_hit_length")
and (
self.tree_cache.is_backuped(last_host_node)
or self.tree_cache.is_root(last_host_node)
)
):
matched_len = l1_hit + l2_hit
suffix_tokens = query_ids[matched_len:]
last_hash = self.tree_cache.get_last_hash_value(last_host_node)
prefix_keys = (
self.tree_cache.get_prefix_hash_values(last_host_node)
if getattr(
self.tree_cache, "hicache_storage_pass_prefix_keys", False
)
else None
)
l3_hit = self.tree_cache.query_storage_hit_length(
last_host_node,
suffix_tokens,
last_hash,
prefix_keys,
extra_key=recv_req.extra_key,
cache_salt=recv_req.cache_salt,
)
return QueryCacheHitReqOutput(
total_tokens=total,
l1_hit_tokens=l1_hit,
l2_hit_tokens=l2_hit,
l3_hit_tokens=l3_hit,
dp_rank=int(get_parallel().dp_rank or 0),
)
except Exception as e:
logger.warning(f"handle_query_cache_hit failed: {e}")
return QueryCacheHitReqOutput(total_tokens=total, error=str(e))
def get_internal_state(self, recv_req: GetInternalStateReq):
# Resolved config (pristine server_args + post-publish overrides) so a
# readback reflects values changed via /set_internal_state, not startup.
@@ -112,6 +112,8 @@ class SchedulerBatchResultProcessor:
output_streamer: SchedulerOutputStreamer
beam_coordinator: BeamCoordinator
abort_request: Callable
# G9: scheduler-side drain-event recorder for the queue-time estimator.
record_prefill_drain: Callable
def process_batch_result_prebuilt(self, batch: ScheduleBatch):
assert self.disaggregation_mode == DisaggregationMode.DECODE
@@ -343,6 +345,9 @@ class SchedulerBatchResultProcessor:
if req.inflight_middle_chunks <= 0:
req.time_stats.set_prefill_finished_time()
# G9: exactly once per request (final chunk only) — feed the
# queue-time estimator with the drained total input length.
self.record_prefill_drain(req.seqlen)
if sampling_mask_finish_reason is not None:
req.to_finish = sampling_mask_finish_reason
@@ -54,6 +54,7 @@ class SchedulerLoadInquirer:
get_total_prefill_uncached_tokens: Callable
get_total_prefill_busy_us: Callable
get_decode_moment_totals: Callable
get_prefill_drain_rate: Callable
def _get_num_pending_tokens(self, chunk_deduct: int = 0) -> int:
"""Get the total number of tokens pending prefill.
@@ -128,9 +129,10 @@ class SchedulerLoadInquirer:
num_used_tokens, kv_token_usage = (
self.pool_stats_observer.get_pool_stats().get_kv_token_stats()
)
num_total_tokens = num_used_tokens + sum(
queued_total_input_tokens = sum(
req.seqlen for queue in pending_token_queues for req in queue
)
num_total_tokens = num_used_tokens + queued_total_input_tokens
num_active_tokens = max(0, num_total_tokens - awaiting_kv_tokens)
memory = None
@@ -220,6 +222,8 @@ class SchedulerLoadInquirer:
num_running_reqs=num_running_reqs,
num_waiting_reqs=num_waiting_reqs,
num_waiting_uncached_tokens=self.get_num_waiting_uncached_tokens(),
queued_total_input_tokens=queued_total_input_tokens,
drain_rate_tokens_per_s=round(self.get_prefill_drain_rate(), 2),
num_used_tokens=num_used_tokens,
num_total_tokens=num_total_tokens,
num_active_tokens=num_active_tokens,
@@ -54,6 +54,8 @@ from sglang.srt.managers.io_struct import (
ProfileReq,
ProfileReqOutput,
ProfileReqType,
QueryCacheHitReqInput,
QueryCacheHitReqOutput,
ReleaseMemoryOccupationReqInput,
ReleaseMemoryOccupationReqOutput,
RemoveExternalCorpusReqInput,
@@ -129,6 +131,7 @@ _COMMUNICATOR_SPECS = [
("detach_hicache_storage", DetachHiCacheStorageReqOutput),
("profile", ProfileReqOutput),
("get_internal_state", GetInternalStateReqOutput),
("query_cache_hit", QueryCacheHitReqOutput),
("set_internal_state", SetInternalStateReqOutput),
("expert_distribution", ExpertDistributionReqOutput),
("update_lora_adapter", LoRAUpdateOutput),
@@ -905,6 +908,37 @@ class TokenizerControlMixin:
self.auto_create_handle_loop()
return await self.dumper_control_communicator(obj)
async def query_cache_hit(
self: TokenizerManager, obj: QueryCacheHitReqInput
) -> QueryCacheHitReqOutput:
"""G9: fan the probe out to all DP ranks and keep the best hit.
The decode node's DP controller routes new requests to any rank, so
the bypass decision uses the best per-rank L1/L2 (L3 is shared); the
real request still goes through normal dispatch afterwards.
"""
self.auto_create_handle_loop()
results: List[QueryCacheHitReqOutput] = await self.query_cache_hit_communicator(
obj
)
ok = [r for r in results if r.error is None]
if not ok:
msg = (
"; ".join(r.error or "empty response" for r in results)
or "no scheduler responded"
)
return QueryCacheHitReqOutput(total_tokens=len(obj.input_ids), error=msg)
best = max(
ok, key=lambda r: r.l1_hit_tokens + r.l2_hit_tokens + r.l3_hit_tokens
)
return QueryCacheHitReqOutput(
total_tokens=best.total_tokens,
l1_hit_tokens=best.l1_hit_tokens,
l2_hit_tokens=best.l2_hit_tokens,
l3_hit_tokens=best.l3_hit_tokens,
dp_rank=best.dp_rank,
)
async def get_loads(
self: TokenizerManager,
include: Optional[List[str]] = None,
@@ -866,3 +866,15 @@ class PureSWATokenToKVPoolAllocator(SWATokenToKVPoolAllocator):
def is_swa_req_ring(allocator) -> bool:
return isinstance(allocator, SWATokenToKVPoolAllocator) and allocator.swa_req_ring
def swa_kept_out_of_tree(allocator) -> bool:
"""Whether radix-tree nodes do not own the SWA state: a per-request SWA ring,
or the DeepSeek-V4.1 encoder-replay request window, which replay rebuilds on a
prefix hit. SWA bookkeeping must not gate prefix matching for either."""
if not isinstance(allocator, SWATokenToKVPoolAllocator):
return False
if allocator.swa_req_ring:
return True
kvcache = getattr(allocator, "_kvcache", None)
return getattr(kvcache, "request_window", None) is not None
@@ -659,6 +659,12 @@ class DeepSeekV4IndexerPool(KVCache):
from the page layout [page_size * 64 payload | page_size * 4 scale]."""
assert self.use_fp4_indexer, "packed readback only applies to the fp4 layout"
buf = self.index_k_with_scale_buffer[layer_id - self.start_layer]
if buf.is_cuda:
from sglang.kernels.ops.attention.dsv4.fp4_indexer import (
gather_fp4_index_k,
)
return gather_fp4_index_k(buf, slots, page_size=self.page_size)
slots = slots.to(torch.int64)
p = self.page_size
page, off = (slots // p).unsqueeze(-1), slots % p
@@ -1855,6 +1861,11 @@ class DeepSeekV4TokenToKVPool(BaseSWAKVPool):
# swa_kv_pool is None under the request window and unified_kv.
return self.kv_layout
def get_swa_key_page_size(self) -> int:
if self.request_window is not None:
return self.request_window.page_size
return self.swa_kv_pool.page_size
def get_swa_key_bytes_per_token(self) -> int:
"""Last dim of the ``(pages, page_size, 1, bytes)`` view the attention
kernel detects the SWA cache's format from."""
@@ -91,7 +91,7 @@ def auto_size_hicache(
return
requested = get_memory().hicache_ratio
device_bytes = _estimate_hicache_bytes(params, draft_plan)
budget = int(host_memory_budget_bytes() * fraction)
budget = int(host_memory_budget_bytes(auto_size=True) * fraction)
ratio = min(requested, budget * (1 - _ALLOCATION_SLACK_FRACTION) / device_bytes)
# One collective before any pool is built: PP stages own different pool
# counts, so a per-pool collective could deadlock.
@@ -20,7 +20,7 @@ if TYPE_CHECKING:
logger = logging.getLogger(__name__)
# Max pages per batched storage IO call.
STORAGE_BATCH_SIZE = 128
STORAGE_BATCH_SIZE = envs.SGLANG_HICACHE_STORAGE_BATCH_SIZE.get()
@dataclass
+1 -1
View File
@@ -653,7 +653,7 @@ class HiRadixCache(RadixCache):
: alloc_len // self.page_size
]
operation.host_indices = host_indices
cc.prefetch_buffer.put(operation)
cc.dispatch_prefetch(operation)
def _drain_ack_prefetch():
for ack in _drain_queue(cc.ack_prefetch_queue, n_ack_prefetch):
+34 -4
View File
@@ -15,6 +15,11 @@ def _unescape_mount_path(value: str) -> str:
return re.sub(r"\\([0-7]{3})", lambda m: chr(int(m[1], 8)), value)
def _memory_stat(directory: Path) -> dict[str, int]:
lines = (directory / "memory.stat").read_text().splitlines()
return {key: int(value) for key, value in map(str.split, lines)}
def _cgroup_memory_headroom(proc_root: Path = Path("/proc")) -> int | None:
memberships = {}
try:
@@ -65,6 +70,13 @@ def _cgroup_memory_headroom(proc_root: Path = Path("/proc")) -> int | None:
usage_name = (
"memory.current" if filesystem == "cgroup2" else "memory.usage_in_bytes"
)
# File LRU pages, as MemAvailable counts; shmem sits on the anon LRU.
# v1's unprefixed counters omit descendants, unlike its usage.
page_cache_keys = (
("active_file", "inactive_file")
if filesystem == "cgroup2"
else ("total_active_file", "total_inactive_file")
)
while True:
for name in limits:
try:
@@ -78,7 +90,10 @@ def _cgroup_memory_headroom(proc_root: Path = Path("/proc")) -> int | None:
# Do not silently ignore an unreadable usage file for a known
# limit: falling back to host RAM could overrun the container.
usage = int((directory / usage_name).read_text())
remaining = max(0, limit - usage)
stat = _memory_stat(directory)
page_cache = sum(stat[key] for key in page_cache_keys)
# Charged page cache is reclaimed under the limit, not used.
remaining = max(0, limit - max(0, usage - page_cache))
headroom = remaining if headroom is None else min(headroom, remaining)
if directory == mount:
break
@@ -90,10 +105,25 @@ def _cgroup_memory_headroom(proc_root: Path = Path("/proc")) -> int | None:
return headroom
def available_host_memory_bytes() -> int:
"""Conservative allocatable RAM; charged file cache is not assumed reclaimable."""
def available_host_memory_bytes(*, allow_cgroup_fallback: bool = False) -> int:
"""Allocatable RAM, counting reclaimable page cache as free."""
available = psutil.virtual_memory().available
cgroup_headroom = _cgroup_memory_headroom()
try:
cgroup_headroom = _cgroup_memory_headroom()
except (OSError, RuntimeError) as exc:
if not allow_cgroup_fallback:
raise RuntimeError(
"Cannot automatically size HiCache host memory; "
"set --hicache-size explicitly (GB per rank), leaving room "
"for other memory use within the container limit."
) from exc
logger.warning(
"Cannot determine cgroup memory headroom: %s. Checking explicitly "
"sized HiCache pools against host available memory; ensure all "
"local ranks' pools fit within the container memory limit.",
exc,
)
return available
if cgroup_headroom is not None:
logger.info(
"HiCache memory headroom: host %.1f GiB, cgroup %.1f GiB",
@@ -15,6 +15,7 @@ import torch
from sglang.srt.managers.cache_controller import (
CacheOperation,
_hc_timing_enabled,
)
from sglang.srt.managers.cache_controller import (
HiCacheController as BaseHiCacheController,
@@ -1005,7 +1006,7 @@ class HybridCacheController(BaseHiCacheController):
if not self._allocate_pp_prefetch_buffers(ticket, operation):
operation.host_indices = torch.empty(0, dtype=torch.int64)
operation.mark_terminate()
self.prefetch_buffer.put(operation)
self.dispatch_prefetch(operation)
except Exception:
logger.exception("PP HiCache ticket processing failed.")
if operation is None or not operation.hash_value:
@@ -1015,7 +1016,7 @@ class HybridCacheController(BaseHiCacheController):
operation.host_indices = torch.empty(0, dtype=torch.int64)
operation.mark_terminate()
# Preserve every KV/sidecar ACK even when local allocation fails.
self.prefetch_buffer.put(operation)
self.dispatch_prefetch(operation)
finally:
if is_source and objects:
self.pp_prefetch_command_queue.task_done()
@@ -1103,15 +1104,15 @@ class HybridCacheController(BaseHiCacheController):
)
return host_indices, device_indices, resolved_pool_transfers
def _page_transfer(self, operation: PrefetchOperation) -> bool:
def _page_transfer(self, operation: PrefetchOperation, sync_queue: Queue) -> bool:
# KV pools and KV-derived pools first — determines actual completed page count
kv_completed_pages = super()._page_transfer(operation)
kv_completed_pages = super()._page_transfer(operation, sync_queue)
# Read non-KV derived sidecar pool, e.g. SWA, Mamba.
self._page_transfer_sidecar(operation, kv_completed_pages)
self._page_transfer_sidecar(operation, kv_completed_pages, sync_queue)
def _page_transfer_sidecar(
self, operation: PrefetchOperation, kv_completed_pages: int
self, operation: PrefetchOperation, kv_completed_pages: int, sync_queue: Queue
) -> None:
if operation.pool_transfers is None:
return
@@ -1143,10 +1144,10 @@ class HybridCacheController(BaseHiCacheController):
transfers_nonkv, extra_info=extra_info
)
pool_hits = count_pool_hits(results)
# Emit PrefetchAck to prefetch_sync_queue, even the operation has been canceled by the
# Emit PrefetchAck to the lane sync queue, even the operation has been canceled by the
# scheduler thread. The prefetch sync thread expects the same number of PrefetchAck objects
# to perform all_reduce.
self.prefetch_sync_queue.put(
sync_queue.put(
PrefetchAck(
rid=operation.request_id,
operation=operation,
@@ -1420,9 +1421,11 @@ class HybridCacheController(BaseHiCacheController):
pool.device_indices = source.device_indices
return extra_pools
def _reduce_prefetch_ack(self, ack: PrefetchAck) -> None:
def _reduce_prefetch_ack(
self, ack: PrefetchAck, groups: List[torch.distributed.ProcessGroup]
) -> None:
# Handle KV-derived pool.
super()._reduce_prefetch_ack(ack)
super()._reduce_prefetch_ack(ack, groups)
# Handle other sidecar pools, e.g. SWA, Mamba.
if ack.pool_hits is not None:
@@ -1435,20 +1438,24 @@ class HybridCacheController(BaseHiCacheController):
self._all_reduce(
packed,
torch.distributed.ReduceOp.MIN,
self.prefetch_completion_sync_groups,
groups,
)
for i, pool in enumerate(PoolName):
ack.pool_hits[pool.value] = packed[i].item()
def prefetch_sync_thread_func(self) -> None:
def prefetch_sync_thread_func(self, lane: int = 0) -> None:
"""Synchronize progressive ACKs and publish PP ticket completion."""
sync_queue = self.prefetch_sync_queues[lane]
groups = self.prefetch_completion_sync_groups_by_lane[lane]
while not self.storage_stop_event.is_set():
try:
ack = self.prefetch_sync_queue.get(block=True, timeout=1)
ack = sync_queue.get(block=True, timeout=1)
if ack is None:
continue
self._reduce_prefetch_ack(ack)
self._reduce_prefetch_ack(ack, groups)
if not getattr(ack.operation, "is_pp_broadcast", False):
if _hc_timing_enabled() and ack.completed_req:
ack.operation.final_ack_enqueue_ts = time.monotonic()
self.ack_prefetch_queue.put(ack)
continue
@@ -61,12 +61,16 @@ def ranks_per_host() -> int:
return max(launch_world_size // get_parallel().nnodes, 1)
def host_memory_budget_bytes(requested_bytes: int = 0) -> int:
def host_memory_budget_bytes(
requested_bytes: int = 0, *, auto_size: bool = False
) -> int:
"""Host RAM this rank may claim for a HiCache pool.
Bound machine availability by the visible cgroup limits before splitting
among local ranks. Independent engines with separate container budgets
therefore size against their own remaining allowance.
Auto-sizing requires successful cgroup discovery; explicit pool sizing
can fall back to checking host availability when discovery fails.
Inside host_memory_budget_scope, requested_bytes is booked against the
snapshot when it fits; the allowance before booking is returned.
@@ -77,7 +81,10 @@ def host_memory_budget_bytes(requested_bytes: int = 0) -> int:
_host_memory_budget.set(available - requested_bytes)
return available
free = available_host_memory_bytes() - HICACHE_HOST_MEMORY_RESERVE_BYTES
free = (
available_host_memory_bytes(allow_cgroup_fallback=not auto_size)
- HICACHE_HOST_MEMORY_RESERVE_BYTES
)
return free // ranks_per_host()
@@ -460,7 +460,18 @@ class MooncakeStore(HiCacheStorage, MooncakeBaseStore):
)
device_name = ""
if self.config.standalone_storage:
if not isinstance(mem_pool.allocator, MooncakeHostTensorAllocator):
# V4 unified cache hands us a LogicalHostPool anchor that owns no
# physical buffers (kv_buffer=None, allocator=None); the real
# buffers live in side pools registered later via
# register_mem_host_pool_v2() -> register_buffer(ptr, size).
# Only enforce the MooncakeHostTensorAllocator requirement when
# the anchor actually owns memory.
_anchor_owns_buffers = (
getattr(mem_pool, "kv_buffer", None) is not None
)
if _anchor_owns_buffers and not isinstance(
getattr(mem_pool, "allocator", None), MooncakeHostTensorAllocator
):
raise RuntimeError(
"MooncakeStore with standalone_storage=True requires MooncakeHostTensorAllocator. "
"Please set standalone_storage=False "
@@ -504,6 +515,22 @@ class MooncakeStore(HiCacheStorage, MooncakeBaseStore):
else:
client_hostname = self.config.local_hostname
transfer_engine = None
# Each rank embeds its own Mooncake client process. Giving
# every rank the same local_hostname makes the master's
# preferred-segment allocation unable to tell their
# segments apart, so replicas can land on a peer rank's
# segment and the local-memcpy fast path (which requires
# same-process endpoints) silently falls back to TCP.
# Make the hostname unique per rank so own-writes land on
# the own segment and reads can use local memcpy.
if storage_config is not None:
client_hostname = (
f"{client_hostname}"
f"-d{storage_config.dp_rank}"
f"t{storage_config.tp_rank}"
f"p{storage_config.pp_rank}"
f"c{storage_config.attn_cp_rank}"
)
setup_kwargs = {}
if self.config.enable_ssd_offload:
@@ -5,7 +5,7 @@ from typing import TYPE_CHECKING, Callable, Optional, Sequence
import torch
from sglang.srt.environ import envs
from sglang.srt.mem_cache.allocator.swa import is_swa_req_ring
from sglang.srt.mem_cache.allocator.swa import swa_kept_out_of_tree
from sglang.srt.mem_cache.base_prefix_cache import (
DecLockRefParams,
EvictParams,
@@ -323,9 +323,9 @@ class SWAComponent(TreeComponent):
ct = self.component_type
state = {"len": float("inf")}
# A per-request SWA ring is not stored in tree nodes, so its bookkeeping
# must not gate prefix matching.
swa_req_ring = is_swa_req_ring(self.cache.token_to_kv_pool_allocator)
# SWA state that tree nodes do not own (a per-request ring, or a request
# window rebuilt by replay) must not gate prefix matching.
swa_out_of_tree = swa_kept_out_of_tree(self.cache.token_to_kv_pool_allocator)
def validator(node: UnifiedTreeNode) -> bool:
cd = node.component_data[ct]
@@ -333,7 +333,15 @@ class SWAComponent(TreeComponent):
# — load_back will restore SWA from host before use.
if cd.value is None and (match_device_only or cd.host_value is None):
state["len"] = 0
if swa_req_ring and (node.backuped or not node.evicted):
if swa_out_of_tree and (node.backuped or not node.evicted):
return True
# Fail-soft: a backuped node whose SWA is absent on *both*
# layers (device tombstone plus no L3 backup — the hole a
# degraded prefetch can graft) still has Full KV on host.
# Treat only that case as a valid match boundary — SWA will
# be rebuilt from Full KV at load-back via
# SWARebuild/RecoverSWAWithLockedFull.
if cd.host_value is None and node.backuped:
return True
return False
state["len"] += len(node.key)
@@ -1080,7 +1088,12 @@ class SWAComponent(TreeComponent):
cur is not self.tree_core.root_node and n_swa < self.sliding_window_size
):
cd = cur.component_data[ct]
assert cd.host_value is not None or cd.value is not None
# Fail-soft: a degraded prefetch may insert a backuped node
# whose SWA was never backed up (tombstoned before backup).
# Stop collecting — SWA will be rebuilt from Full KV at
# load-back via SWARebuild/RecoverSWAWithLockedFull.
if cd.host_value is None and cd.value is None:
break
if cd.value is not None:
# device exists, skip it
n_swa += len(cd.value)
@@ -1360,7 +1360,8 @@ class UnifiedTreeCore(UnifiedTreeCoreInterface):
)
def _insert_tail_step(self, state: _InsertWalkState) -> None:
"""Refresh the LRUs and append terminal backup actions."""
"""Refresh the LRUs and append the insert backup: a new-leaf write-through,
or an SWA window publish on an existing backed node."""
if state.target_node is not self.root_node:
for component in self.components:
if component.component_type == BASE_COMPONENT_TYPE:
@@ -1371,7 +1372,9 @@ class UnifiedTreeCore(UnifiedTreeCoreInterface):
if self._should_backup_after_insert(state):
state.pending_actions.append(
self._build_backup_kv_action(state.target_node)
self._build_backup_kv_action(
state.target_node, write_back=self.is_write_back
)
)
def _split_node(
@@ -2350,7 +2353,8 @@ class UnifiedTreeCore(UnifiedTreeCoreInterface):
def _build_backup_kv_action(
self, node: UnifiedTreeNode, write_back: bool = False
) -> BackupKV:
"""Build the backup action for a node and its not-yet-persisted ancestors."""
"""Build the backup action for a node; write-through also chains its
unbacked ancestors."""
chain = [node]
if not write_back:
ancestor = node.parent
@@ -2,6 +2,7 @@ from __future__ import annotations
import atexit
import logging
import os
import threading
import time
from dataclasses import replace
@@ -12,7 +13,7 @@ import torch
from sglang.srt.distributed.communication_tags import P2PTag
from sglang.srt.environ import envs
from sglang.srt.managers.cache_controller import CacheOperation
from sglang.srt.managers.cache_controller import CacheOperation, _hc_timing_enabled
from sglang.srt.mem_cache.allocator.page_interleave import (
page_interleave_shard_size,
)
@@ -48,6 +49,7 @@ from sglang.srt.mem_cache.memory_pool import MHATokenToKVPool
from sglang.srt.mem_cache.radix_cache import RadixKey
from sglang.srt.mem_cache.storage_prefetch import StoragePrefetchRetries
from sglang.srt.mem_cache.swa_memory_pool import SWAKVPool
from sglang.srt.mem_cache.utils import get_hash_str
from sglang.srt.mem_cache.unified_cache.cache_action import (
BackupKV,
CacheAction,
@@ -131,6 +133,11 @@ COMPONENT_REGISTRY: dict[ComponentType, type[TreeComponent]] = {
logger = logging.getLogger(__name__)
# DEBUG-ONLY (p2 wedge trap 2026-10-13): env-gated logging for the per-step
# hicache event sync. Same SGLANG_AR_SEQ_DEBUG switch as the ARDBG sites.
_AR_SEQ_DEBUG = int(os.environ.get("SGLANG_AR_SEQ_DEBUG", "0"))
_hc_step = 0
class _OngoingWriteThrough(NamedTuple):
"""Tracks an in-flight D→H write-through operation."""
@@ -1688,6 +1695,7 @@ class UnifiedRadixCache(BasePrefixCache):
for comp in self._components_tuple
}
success = False
_hc_t0 = time.monotonic() if _hc_timing_enabled() else 0.0
try:
success = self._load_back_transfers(
node_id=node_id,
@@ -1699,6 +1707,14 @@ class UnifiedRadixCache(BasePrefixCache):
)
return success
finally:
if _hc_timing_enabled():
logger.info(
"HC_TIMING stage=loadback req=%s node=%d ms=%.1f success=%s",
getattr(req, "rid", None) if req is not None else "-",
node_id,
(time.monotonic() - _hc_t0) * 1e3,
success,
)
for comp in self._components_tuple:
comp.finalize_load_back(req, preps[comp.component_type], success)
@@ -1910,11 +1926,46 @@ class UnifiedRadixCache(BasePrefixCache):
if len(prefetch_key) < self.prefetch_threshold:
return 0
# DeepSeek V4 keeps no physical KV objects in storage (logical anchor),
# so a plain batch_exists on KV keys always misses; the usable prefix is
# decided by the side pools. Mirror prefetch_from_storage's transfer
# construction so the query goes through batch_exists_v2. Both hooks are
# pure sizing/descriptor builders: prepare_prefetch only computes
# staging_tokens, and PREFETCH-phase transfers carry placeholder keys
# whose length sizes the TRAILING_PAGES window for batch_exists_v2.
comp_xfers: dict[ComponentType, list[PoolTransfer]] = {}
for ct in self.tree_components:
if ct == BASE_COMPONENT_TYPE:
continue
prep = self.components[ct].prepare_prefetch(
last_host_node_id, prefetch_tokens=len(prefetch_key)
)
if prep.staging_tokens == 0:
continue
transfers = self.tree_core.build_hicache_transfers(
ct,
last_host_node_id,
CacheTransferPhase.PREFETCH,
token_ids=prefetch_key.token_ids,
prefetch_tokens=len(prefetch_key),
staging_tokens=prep.staging_tokens,
last_hash=last_hash,
)
if transfers:
comp_xfers[ct] = transfers
kv_xfer = PoolTransfer(name=PoolName.KV, host_indices=None)
sidecar_xfers = self._build_sidecar_transfers(
CacheTransferPhase.PREFETCH, kv_xfer, comp_xfers
)
aux_xfers = [x for xfers in comp_xfers.values() for x in xfers]
aux_xfers.extend(sidecar_xfers)
operation = PrefetchOperation(
CacheRequestHandle("__storage_hit_query__", 0),
prefetch_key,
last_hash,
prefix_keys,
pool_transfers=aux_xfers or None,
)
_, storage_hit_count = self.cache_controller._storage_hit_query(operation)
storage_hit_count_tensor = torch.tensor(storage_hit_count, dtype=torch.int)
@@ -2005,6 +2056,15 @@ class UnifiedRadixCache(BasePrefixCache):
self.storage_prefetch_retries.poll_miss(req_id, storage_hit_end)
return
# Recompute last_hash from the full matched prefix so the hash chain
# is correct regardless of which node was selected as anchor. The
# anchor's hash_value may be at a different position than matched_len
# when the anchor has no host_value (async backup not yet complete).
if matched_prefix_tokens and len(matched_prefix_tokens) >= self.page_size:
prefix_hashes = get_hash_str(matched_prefix_tokens, None, self.page_size)
if prefix_hashes:
last_hash = prefix_hashes[-1]
# Buffer mode holds no tree state during the fetch: buffers are
# operation-owned, so the anchor needs no pin.
anchor_lock_params = (
@@ -2236,9 +2296,14 @@ class UnifiedRadixCache(BasePrefixCache):
last_host_node_id,
anchor_lock_params,
prefetch_key,
comp_xfers,
):
# Hybrid all-or-nothing check failed; result already discarded.
return
# The check may have degraded the result (fail-soft): re-read the
# retained KV prefix length and hash chain off the operation.
completed_tokens = operation.completed_tokens
hash_value = operation.hash_value
allocated_tokens = len(host_indices)
if completed_tokens < allocated_tokens:
@@ -2262,12 +2327,15 @@ class UnifiedRadixCache(BasePrefixCache):
)
fetched_key = prefetch_key[:completed_tokens]
_hc_tg = time.monotonic() if _hc_timing_enabled() else 0.0
insert_result = self.tree_core.insert_host(
last_host_node_id,
fetched_key,
host_indices[:completed_tokens],
hash_value[: completed_tokens // self.page_size],
)
if _hc_timing_enabled():
operation.graft_ms = (time.monotonic() - _hc_tg) * 1e3
# Apply the host-insert walk's actions before the transfer commit.
self._apply_cache_actions(insert_result.cache_actions)
@@ -2323,6 +2391,16 @@ class UnifiedRadixCache(BasePrefixCache):
loaded_from_storage,
self.cache_controller.prefetch_tokens_occupied,
)
if _hc_timing_enabled():
logger.info(
"HC_TIMING stage=prefetch_e2e req=%s e2e_ms=%.1f graft_ms=%.1f "
"completed=%d loaded=%d",
request.rid,
(time.monotonic() - operation.start_time) * 1e3,
getattr(operation, "graft_ms", -1.0),
completed_tokens,
loaded_from_storage,
)
return
def _check_hybrid_prefetch_result(
@@ -2335,6 +2413,7 @@ class UnifiedRadixCache(BasePrefixCache):
last_host_node_id: NodeId,
anchor_lock_params: DecLockRefParams,
prefetch_key: RadixKey,
comp_xfers: dict[ComponentType, list[PoolTransfer]],
) -> bool:
"""Decide the length of usable prefix.
@@ -2345,11 +2424,16 @@ class UnifiedRadixCache(BasePrefixCache):
the Full KV pool and every sidecar. A partial prefix is still usable
because the sidecar is page-aligned with KV and required for every page.
* Everything else (SWA / Mamba components, mixed DeepSeekV4 stacks):
*all-or-nothing*. Their pools only cover a window / tail and cannot be
truncated page by page, so any shortfall discards the whole prefetch.
*fail-soft degrade* when the shortfall is confined to pools with a
cheap rebuild path (SWA — the missing window is rebuilt from Full KV
at load-back): the fetched KV prefix is retained and only the failed
pool is discarded. Falls back to *all-or-nothing* when nothing is
usable or a non-rebuildable pool (Mamba) falls short, since those
pools cannot be truncated page by page.
Returns true if prefetch success, or false when an all-or-nothing prefetch
was discarded (the caller should then treat the prefetch as finished).
Returns true if prefetch success (possibly degraded), or false when an
all-or-nothing prefetch was discarded (the caller should then treat the
prefetch as finished).
"""
# Sync completed tokens and per-pool hit pages across ATTN groups, taking
# the minimum so every rank agrees on the same usable prefix length.
@@ -2375,57 +2459,118 @@ class UnifiedRadixCache(BasePrefixCache):
for transfer, count in zip(pool_transfers, pool_hit_pages)
)
if pool_transfers and not all_succeeded:
# Drop the KV beliefs from the first page any pool failed to serve;
# the next insert then re-writes that span through one FULL check,
# restoring the missing aux pages.
# Identify which non-KV components fell short. KV-derived
# sidecars (indices_from_pool == KV) are excluded — they share
# the KV hit length and are always covered when KV succeeds.
failed_components = [
ct
for ct, xfers in comp_xfers.items()
if any(x.indices_from_pool != PoolName.KV for x in xfers)
]
# Only SWA has a cheap rebuild path (_translate_full_to_swa);
# the missing window is rebuilt at load-back via
# SWARebuild/RecoverSWAWithLockedFull. Mamba SSM states require
# O(seq_len) recompute — not degradable.
rebuildable = all(ct != ComponentType.MAMBA for ct in failed_components)
if completed_tokens == 0 or not rebuildable:
# Full discard — nothing usable from any pool, or a
# non-rebuildable pool (Mamba) fell short so the whole
# result is unusable.
# Drop the KV beliefs from the first page any pool failed to serve;
# the next insert then re-writes that span through one FULL check,
# restoring the missing aux pages.
invalidation_hashes = (
operation.sidecar_hash_values
if operation.sidecar_hash_values is not None
else hash_value
)
trimmed_pages = len(invalidation_hashes) - len(hash_value)
keep_pages = trimmed_pages + completed_tokens // self.page_size
for transfer, count in zip(pool_transfers, pool_hit_pages):
if transfer.keys is None:
keep_pages = 0
elif count < len(transfer.keys):
# Aux transfers key the chain's trailing pages.
keep_pages = min(
keep_pages,
max(0, len(invalidation_hashes) - len(transfer.keys)),
)
self.storage_existence_cache.invalidate_beyond(
PoolName.KV, invalidation_hashes, keep_pages=keep_pages
)
# The controller's prefetch IO thread already releases the untransferred
# tail (host_indices[completed_tokens:])
self.cache_controller.append_host_mem_release(
host_indices=host_indices[:completed_tokens],
extra_pools=pool_transfers if operation.pool_transfers_done else None,
)
self._finish_storage_prefetch(
request, fulfilled_tokens=0, reason="storage_transfer"
)
if anchor_lock_params is not None:
self.dec_host_lock_ref(last_host_node_id, anchor_lock_params)
if self.buffer_pipeline is not None:
self.buffer_pipeline.pop_prefix_ctx(request)
self.buffer_pipeline.release_anchor_lock(request)
del self.ongoing_prefetch[request]
self.cache_controller.prefetch_tokens_occupied -= (
self._prefetch_occupied_span(prefetch_key, host_indices)
)
self.prefetch_loaded_tokens_by_reqid[request] = 0
self.prefetch_loaded_storage_start_by_reqid.pop(request, None)
logger.warning(
"HiCache hybrid prefetch discarded req=%s completed=%d requested=%d "
"kv_beliefs_kept_pages=%d",
request.rid,
completed_tokens,
expected_tokens,
keep_pages,
)
return False
# Fail-soft degrade: retain the Full KV (+ KV-derived sidecar)
# result and discard **only** the failed non-KV independent
# pools (SWA) whose storage fetch fell short. Their host
# buffers are released here and their component entries are
# removed from comp_xfers so the caller's normal commit /
# staging path never touches the freed memory. At load-back
# time the missing SWA window is rebuilt via the existing
# RecoverSWAWithLockedFull / SWARebuild path in swa.py.
# Belief self-heal is scoped to the retained KV prefix.
invalidation_hashes = (
operation.sidecar_hash_values
if operation.sidecar_hash_values is not None
else hash_value
)
trimmed_pages = len(invalidation_hashes) - len(hash_value)
keep_pages = trimmed_pages + completed_tokens // self.page_size
for transfer, count in zip(pool_transfers, pool_hit_pages):
if transfer.keys is None:
keep_pages = 0
elif count < len(transfer.keys):
# Aux transfers key the chain's trailing pages.
keep_pages = min(
keep_pages,
max(0, len(invalidation_hashes) - len(transfer.keys)),
)
self.storage_existence_cache.invalidate_beyond(
PoolName.KV, invalidation_hashes, keep_pages=keep_pages
PoolName.KV,
invalidation_hashes,
keep_pages=trimmed_pages + completed_tokens // self.page_size,
)
# The controller's prefetch IO thread already releases the untransferred
# tail (host_indices[completed_tokens:])
self.cache_controller.append_host_mem_release(
host_indices=host_indices[:completed_tokens],
extra_pools=pool_transfers if operation.pool_transfers_done else None,
)
self._finish_storage_prefetch(
request, fulfilled_tokens=0, reason="storage_transfer"
)
if anchor_lock_params is not None:
self.dec_host_lock_ref(last_host_node_id, anchor_lock_params)
if self.buffer_pipeline is not None:
self.buffer_pipeline.pop_prefix_ctx(request)
self.buffer_pipeline.release_anchor_lock(request)
del self.ongoing_prefetch[request]
self.cache_controller.prefetch_tokens_occupied -= (
self._prefetch_occupied_span(prefetch_key, host_indices)
)
self.prefetch_loaded_tokens_by_reqid[request] = 0
self.prefetch_loaded_storage_start_by_reqid.pop(request, None)
# Remove failed non-KV components from comp_xfers (same dict as
# the caller's) so the commit / staging path won't touch freed
# host buffers.
for ct in failed_components:
comp_xfers.pop(ct, None)
# Retain only the KV-fetched prefix; the caller re-reads these
# off the operation after the check passes.
operation.completed_tokens = completed_tokens
operation.hash_value = hash_value[: completed_tokens // self.page_size]
logger.warning(
"HiCache hybrid prefetch discarded req=%s completed=%d requested=%d "
"kv_beliefs_kept_pages=%d",
"HiCache hybrid prefetch degraded req=%s completed=%d "
"requested=%d retained_kv=%d discarded_components=%s",
request.rid,
completed_tokens,
expected_tokens,
keep_pages,
completed_tokens,
[ct.value for ct in failed_components],
)
return False
return True
return True
def _record_storage_prefetch_hit(
@@ -2897,7 +3042,7 @@ class UnifiedRadixCache(BasePrefixCache):
self.ongoing_prefetch[request] = info._replace(host_indices=host_indices)
if buffer_mode:
cc.prefetch_tokens_occupied += alloc_len
cc.prefetch_buffer.put(operation)
cc.dispatch_prefetch(operation)
return True
def _drain_and_alloc_storage_hit():
@@ -2978,6 +3123,14 @@ class UnifiedRadixCache(BasePrefixCache):
operation.pool_transfers_done = True
if ack.completed_req:
if is_current:
if _hc_timing_enabled():
_hc_enq = getattr(operation, "final_ack_enqueue_ts", None)
if _hc_enq is not None:
logger.info(
"HC_TIMING stage=ack_drain req=%s wait_ms=%.1f",
operation.request_id,
(time.monotonic() - _hc_enq) * 1e3,
)
# check_prefetch_progress() is not called for this rid yet.
# Let us insert the prefetch result into the radix tree.
self._handle_prefetch_result(operation)
@@ -3386,6 +3539,16 @@ class UnifiedRadixCache(BasePrefixCache):
def check_hicache_events(self) -> None:
"""Called per scheduler step to poll async HiCache events."""
global _hc_step
if _AR_SEQ_DEBUG:
_hc_step += 1
if _hc_step & 255 == 0:
logger.info(
"[ARDBG] hcev rank=%s step=%d linker=%s",
torch.distributed.get_rank(),
_hc_step,
self.linker is not None,
)
if self.linker is not None:
finish_counts = torch.tensor(
[
@@ -3419,6 +3582,15 @@ class UnifiedRadixCache(BasePrefixCache):
storage_queue_sizes,
extra_pool_names,
) = self._sync_hicache_ready_counts()
if _AR_SEQ_DEBUG and _hc_step & 255 == 0:
logger.info(
"[ARDBG] hcsync rank=%s step=%d write=%s load=%s storage=%s",
torch.distributed.get_rank(),
_hc_step,
write_finish_count,
load_finish_count,
storage_queue_sizes,
)
self.writing_check(finish_count=write_finish_count)
self.loading_check(finish_count=load_finish_count)
+6 -1
View File
@@ -4488,9 +4488,12 @@ class DeepseekV4Model(nn.Module):
forward_batch,
cp_all_tokens=cp_extend,
)
# Only multimodal placeholders become image_token_id, so a text-only
# batch skips this full-residual where.
if (
self.config.model_type == "deepseek_v41"
and self.config.vision_n_layers > 0
and forward_batch.contains_mm_inputs()
):
hidden_states = torch.where(
(input_ids == self.config.image_token_id)[:, None, None],
@@ -4715,7 +4718,9 @@ class DeepseekV4Model(nn.Module):
hidden_states = self.embed_tokens(input_ids)
else:
hidden_states = input_embeds
hidden_states = hidden_states.unsqueeze(1).repeat(1, self.hc_mult, 1)
from sglang.kernels.ops.layernorm.mhc import hc_broadcast
hidden_states = hc_broadcast(hidden_states, self.hc_mult)
else:
assert pp_proxy_tensors is not None
hidden_states = pp_proxy_tensors["hidden_states"]
@@ -157,7 +157,7 @@ class Session:
or carry_fill is input_ids_unpadded
):
# Unexpected type or aliased with an origin array (extending it
# below would double-append): let _refresh_fill_ids rebuild.
# below would double-append): let refresh_fill_ids rebuild.
carry_fill = None
else:
del carry_fill[self.committed_fill_len :]
+2 -1
View File
@@ -4052,7 +4052,8 @@ def get_cuda_graph_batch_size_alignment() -> int:
alignment *= 2
if require_gathered_buffer():
alignment *= get_parallel().attn_tp_size
if alignment % get_parallel().attn_cp_size != 0:
# TODO: unverified on NVIDIA; drop the gate once validated on CUDA.
if not is_hip() and alignment % get_parallel().attn_cp_size != 0:
alignment *= get_parallel().attn_cp_size
return alignment
@@ -2592,6 +2592,107 @@ fn backup_kv_action_chains_unbacked_ancestors_first() {
assert_eq!(action.node_ids, vec![c]);
}
// Mirrors UnifiedRadixCache._execute_and_commit_kv_backup: back up each chain
// node and stamp every node its transfers cover with that node's ack.
// It skips the inc_lock_ref the Python side takes; the asserts do not depend on it.
fn execute_backup_kv_for_test(tc: &mut UnifiedTreeCore<Vec<i64>>, backup: &BackupKV) {
for &node_id in &backup.node_ids {
let (device_value, mut comp_xfers) = tc.build_backup_spec(node_id).expect("backup spec");
if device_value.numel() == 0 && comp_xfers.is_empty() {
continue;
}
let mut publish_node_ids = Vec::new();
for transfer in comp_xfers.values_mut().flatten() {
transfer.host_indices = transfer.device_indices.as_ref().map(Tensor::copy);
publish_node_ids.extend(transfer.nodes_to_load.iter().flatten().copied());
}
if !publish_node_ids.contains(&node_id) {
publish_node_ids.push(node_id);
}
tc.commit_backup(node_id, device_value.copy(), comp_xfers)
.expect("live test node");
tc.mark_write_through_pending(publish_node_ids, /* ack_id = */ node_id)
.expect("live test nodes");
}
}
#[test]
fn write_back_swa_publish_leaves_an_ancestor_pending_under_another_ack_alone() {
// c(Full unbacked) -> b(Full backed) -> t1(Full backed)
// -> d(Full unbacked) -> t2(Full backed)
// Every node holds a device-only SWA value inside the window.
let mut tc = UnifiedTreeCore::new(
CacheInitParams {
is_write_back: true,
enable_hicache: true,
has_swa_host_pool: true,
swa_sliding_window_size: Some(8),
..Default::default()
},
vec![FULL, SWA],
);
let (key_t1, key_t2) = (vec![1, 2, 3, 4], vec![1, 2, 5, 6]);
tc.insert(&insert_params(&key_t1, &[10, 11, 12, 13]));
tc.insert(&insert_params(&key_t2, &[10, 11, 14, 15]));
tc.insert(&insert_params(&vec![1, 2, 3], &[10, 11, 12]));
tc.insert(&insert_params(&vec![1, 2, 5], &[10, 11, 14]));
let child = |parent, page: &[i64]| {
tc.arena
.child_on_page(parent, /* extra_key = */ None, page)
.expect("child on page")
};
let c_idx = child(tc.arena.root(), &[1]);
let b_idx = child(c_idx, &[3]);
let t1_idx = child(b_idx, &[4]);
let d_idx = child(c_idx, &[5]);
let t2_idx = child(d_idx, &[6]);
let [c, b, t1, d, t2] = [c_idx, b_idx, t1_idx, d_idx, t2_idx].map(|idx| tc.arena.node(idx).id);
for idx in [c_idx, b_idx, t1_idx, d_idx, t2_idx] {
let full = tc.arena.node(idx).device_value(FULL).copy();
tc.set_component_device_value(tc.arena.node(idx).id, SWA, full.copy())
.expect("live test node");
if [b_idx, t1_idx, t2_idx].contains(&idx) {
tc.arena.set_host_value(idx, FULL, full);
}
}
// Two requests in one batch finish at t1 and t2; neither ack drains.
for (key, target) in [(&key_t1, t1), (&key_t2, t2)] {
let result = tc.insert(&insert_params(key, &[20, 21, 22, 23]));
let backups: Vec<&BackupKV> = result
.cache_actions
.iter()
.filter_map(|action| match action {
CacheAction::BackupKV(backup) => Some(backup),
_ => None,
})
.collect();
assert_eq!(backups.len(), 1);
assert_eq!(backups[0].node_ids, vec![target]);
execute_backup_kv_for_test(&mut tc, backups[0]);
}
let pending = |node_id| {
tc.arena
.node(tc.arena.resolve(node_id).expect("live test node"))
.write_through_pending_id
};
for (node_id, ack_id) in [(c, t1), (b, t1), (t1, t1), (d, t2), (t2, t2)] {
assert_eq!(pending(node_id), Some(ack_id));
}
// Write-back defers the unbacked Full prefix to eviction.
assert!(
!tc.arena
.node(tc.arena.resolve(c).expect("live test node"))
.backuped()
);
assert!(
!tc.arena
.node(tc.arena.resolve(d).expect("live test node"))
.backuped()
);
}
#[test]
fn backup_kv_action_stops_at_an_externally_stored_or_pending_ancestor() {
let mut tc = core();
@@ -1833,7 +1833,8 @@ impl<K: ChildKeyType> UnifiedTreeCore<K> {
&& self.needs_incremental_component_backup_(target_node_id)
}
/// Refresh the LRUs and append terminal backup actions.
/// Refresh the LRUs and append the insert backup: a new-leaf write-through,
/// or an SWA window publish on an existing backed node.
fn insert_tail_step_(&mut self, state: &mut InsertWalkState<K>) {
let target_node_id = state
.target_node_id
@@ -1850,10 +1851,8 @@ impl<K: ChildKeyType> UnifiedTreeCore<K> {
}
if self.should_backup_after_insert_(state, target_node_id) {
let backup = self.build_backup_kv_action_(
self.arena.node(target_node_id),
/* write_back = */ false,
);
let backup =
self.build_backup_kv_action_(self.arena.node(target_node_id), self.is_write_back);
state.pending_actions.push(CacheAction::BackupKV(backup));
}
}
@@ -3611,7 +3610,8 @@ impl<K: ChildKeyType> UnifiedTreeCore<K> {
Ok(order)
}
/// Build the backup action for a node and its not-yet-persisted ancestors.
/// Build the backup action for a node; write-through also chains its
/// unbacked ancestors.
pub fn build_backup_kv_action_(&self, node: &Node<K>, write_back: bool) -> BackupKV {
let mut chain = vec![node.id];
if !write_back {
+6
View File
@@ -378,6 +378,12 @@ impl RouterConfigBuilder {
self
}
/// G9i pool-aware admission control + fast-P routing config (PD mode).
pub fn g9i(mut self, g9i: Option<crate::routers::http::g9i::G9iConfig>) -> Self {
self.config.g9i = g9i;
self
}
pub fn model_path<S: Into<String>>(mut self, path: S) -> Self {
self.config.model_path = Some(path.into());
self
+4
View File
@@ -99,6 +99,9 @@ pub struct RouterConfig {
/// Enable WASM support
#[serde(default)]
pub enable_wasm: bool,
/// G9i pool-aware admission control + fast-P routing (PD mode, HTTP only).
#[serde(default)]
pub g9i: Option<crate::routers::http::g9i::G9iConfig>,
}
/// Tokenizer cache configuration
@@ -551,6 +554,7 @@ impl Default for RouterConfig {
ca_certificates: vec![],
mcp_config: None,
enable_wasm: false,
g9i: None,
server_cert: None,
server_key: None,
}
+198
View File
@@ -219,6 +219,155 @@ struct CliArgs {
#[arg(long, default_value_t = 30, help_heading = "PD Disaggregation")]
worker_startup_check_interval: u64,
// ==================== G9i: Pool Admission + Fast-P Routing ====================
/// Enable G9i pool-aware admission control + fast-P routing (PD HTTP mode)
#[arg(long, default_value_t = false, help_heading = "G9i Pool Admission")]
g9i_enabled: bool,
/// Fast-P pool member base URLs (repeatable; router-side role, engine unaware)
#[arg(long, action = ArgAction::Append, help_heading = "G9i Pool Admission")]
g9i_fast_prefill_url: Vec<String>,
/// Small-prefill cutoff on NEW tokens (total - prefix hit). Tentative; tune by bench.
#[arg(long, default_value_t = 4096, help_heading = "G9i Pool Admission")]
g9i_max_new_tokens: u64,
/// 429 when the least-wait fast-P queue estimate exceeds this many seconds. Tentative.
#[arg(long, default_value_t = 5.0, help_heading = "G9i Pool Admission")]
g9i_fast_p_queue_reject_secs: f64,
/// D-pool admission ceiling: sum of in-processing requests across decode nodes. Tentative.
#[arg(long, default_value_t = 256, help_heading = "G9i Pool Admission")]
g9i_decode_max_running: u64,
/// 429 when the least-wait REGULAR pool queue estimate exceeds this many seconds. Tentative 10.0; 0=off.
#[arg(long, default_value_t = 10.0, help_heading = "G9i Pool Admission")]
g9i_regular_p_queue_reject_secs: f64,
/// Drain-rate floor (tok/s) for queue-wait estimates (div-by-zero / cold-start guard)
#[arg(long, default_value_t = 1000.0, help_heading = "G9i Pool Admission")]
g9i_drain_rate_floor: f64,
/// Snapshot poll interval (seconds)
#[arg(long, default_value_t = 1, help_heading = "G9i Pool Admission")]
g9i_poll_interval_secs: u64,
/// Snapshot TTL — stale nodes leave pool decisions (seconds)
#[arg(long, default_value_t = 5, help_heading = "G9i Pool Admission")]
g9i_snapshot_ttl_secs: u64,
/// No fresh D-pool snapshot: "allow" (default, fail-open + warn) or "reject"
#[arg(long, default_value = "allow", value_parser = ["allow", "reject"], help_heading = "G9i Pool Admission")]
g9i_d_no_snapshot: String,
/// /query_cache_hit probe timeout per request (ms)
#[arg(long, default_value_t = 2000, help_heading = "G9i Pool Admission")]
g9i_cache_hit_query_timeout_ms: u64,
/// Also least-wait route the regular P pool (default: keep configured policy)
#[arg(long, default_value_t = false, help_heading = "G9i Pool Admission")]
g9i_regular_pool_least_wait: bool,
/// Sliding window for TTFT p95/p99 statistics publication (seconds; observability only, never gates)
#[arg(long, default_value_t = 300, help_heading = "G9i Pool Admission")]
g9i_ttft_window_s: u64,
/// G9i v2 shadow mode: prefix-chain bookkeeping + shadow classifier + queue model (observability only, zero routing effect; requires --g9i-enabled)
#[arg(long, default_value_t = false, help_heading = "G9i Pool Admission")]
g9i_shadow_enabled: bool,
/// Shadow prefix-table capacity in entries (LRU + TTL evicted)
#[arg(long, default_value_t = 100_000, help_heading = "G9i Pool Admission")]
g9i_shadow_table_max_entries: usize,
/// Shadow prefix-table entry TTL (seconds)
#[arg(long, default_value_t = 86400, help_heading = "G9i Pool Admission")]
g9i_shadow_table_ttl_secs: u64,
/// est_T_D shadow estimate (D-side prealloc wait folded into shadow est_TTFT; observability only, default on; set false to roll back to the P-only model exactly)
#[arg(long, default_value_t = true, help_heading = "G9i Pool Admission")]
g9i_est_td_enabled: bool,
/// Shadow-consistent routing: live prefill pick follows the shadow decision's matched_node (requires --g9i-shadow-enabled; rejection gates unchanged; default off)
#[arg(long, default_value_t = false, help_heading = "G9i Pool Admission")]
g9i_shadow_route: bool,
/// T_D admission mode: off (default, ≡ current behavior) | shadow (log/count would-rejects, allow) | enforce (real 429s; requires --g9i-shadow-enabled)
#[arg(long, default_value = "off", help_heading = "G9i Pool Admission")]
g9i_td_admit: String,
/// T_D hard gate (seconds): when min over decode nodes of T_D exceeds this, every class gets 429 d_prealloc_saturated (0 = disabled)
#[arg(long, default_value_t = 5.0, help_heading = "G9i Pool Admission")]
g9i_td_gate_secs: f64,
/// T_PREFER (seconds): prefix-holder affinity yields to a cheaper prefill candidate when the holder's est_TTFT exceeds this
#[arg(long, default_value_t = 2.0, help_heading = "G9i Pool Admission")]
g9i_t_prefer_secs: f64,
/// Hard P-queue token cap (physical backstop; effective cap = min(this, drain_ewma x budget)) — 429 p_queue_cap only when ALL nodes over (0 = disabled; follows --g9i-td-admit mode)
#[arg(long, default_value_t = 16_000_000, help_heading = "G9i Pool Admission")]
g9i_p_queue_token_cap: u64,
/// P-queue real-queue gate: skip prefill nodes whose engine-reported waiting queue (num_waiting_reqs) exceeds this (0 = disabled; follows --g9i-td-admit mode)
#[arg(long, default_value_t = 10, help_heading = "G9i Pool Admission")]
g9i_p_queue_reqs_cap: u64,
/// Admission hold queue: park holdable rejections (d_window_full / p_queue_cap) up to this many seconds instead of an immediate 429 (0 = disabled; enforce mode only; hold time counts into TTFT)
#[arg(long, default_value_t = 1.0, help_heading = "G9i Pool Admission")]
g9i_admit_hold_secs: f64,
/// Admission hold queue bound: at most this many parked requests (full queue = immediate 429)
#[arg(long, default_value_t = 512, help_heading = "G9i Pool Admission")]
g9i_admit_hold_max: u64,
/// D-ledger pipeline booking TTL (seconds): force-release bookings that never saw a first token (wedge backstop; 0 = disabled)
#[arg(long, default_value_t = 120, help_heading = "G9i Pool Admission")]
g9i_pipeline_ttl_secs: u64,
/// D-ledger streaming booking TTL (seconds): force-release bookings stuck mid-decode (0 = disabled)
#[arg(long, default_value_t = 600, help_heading = "G9i Pool Admission")]
g9i_streaming_ttl_secs: u64,
/// Ledger reconcile: clamp the D ledger down to the engine-side in-flight when the excess is sustained >60s (subtract-only)
#[arg(long, default_value_t = true, help_heading = "G9i Pool Admission")]
g9i_ledger_reconcile: bool,
/// AIMD window initial pipeline depth per decode node (warmup start; clamped to [16, 256])
#[arg(long, default_value_t = 128.0, help_heading = "G9i Pool Admission")]
g9i_window_init: f64,
/// AIMD additive-increase step per no-congestion tick (slo_ok / blind_recover; cut & holds unchanged)
#[arg(long, default_value_t = 64.0, help_heading = "G9i Pool Admission")]
g9i_window_add: f64,
/// Pin the AIMD admission window at this value (clamped to [16, 256]); skips all window dynamics (0 = normal AIMD)
#[arg(long, default_value_t = 0.0, help_heading = "G9i Pool Admission")]
g9i_window_fixed: f64,
/// Prompt-length hard validation: 400 prompt_too_long at the front of admission when even the lower-bound estimate exceeds the context limit (default on)
#[arg(long, default_value_t = true, help_heading = "G9i Pool Admission")]
g9i_prompt_validation: bool,
/// Context-length limit (tokens) for the prompt_too_long lower-bound check (0 = disabled)
#[arg(long, default_value_t = 1_048_576, help_heading = "G9i Pool Admission")]
g9i_prompt_token_limit: u64,
/// T25 hot-channel admission A/B policy: off (default ≡ current behavior) | reservoir (Hot-class holdable rejections park in a priority reservoir up to --g9i-reservoir-max-ms instead of 429) | hardlane (cold classes may use only (1 - --g9i-hot-reserve-frac) of the d_window/p_queue budgets)
#[arg(long, default_value = "off", help_heading = "G9i Pool Admission")]
g9i_hot_policy: String,
/// T25-A reservoir wait budget in milliseconds (timeout → 429 reservoir_timeout; the wait counts into TTFT)
#[arg(long, default_value_t = 1000, help_heading = "G9i Pool Admission")]
g9i_reservoir_max_ms: u64,
/// T25-A reservoir capacity: at most this many parked requests (full → immediate 429 reservoir_full)
#[arg(long, default_value_t = 512, help_heading = "G9i Pool Admission")]
g9i_reservoir_cap: u64,
/// T25-B hardlane cold-budget reserve fraction: cold classes use at most (1 - frac) of the AIMD d_window and p_queue caps (fixed value, no adaptation)
#[arg(long, default_value_t = 0.25, help_heading = "G9i Pool Admission")]
g9i_hot_reserve_frac: f64,
// ==================== Service Discovery (Kubernetes) ====================
/// Enable Kubernetes service discovery
#[arg(
@@ -1070,11 +1219,60 @@ impl CliArgs {
.circuit_breaker(!self.disable_circuit_breaker)
.enable_wasm(self.enable_wasm)
.igw(self.enable_igw)
.g9i(self.build_g9i_config()?)
.maybe_server_cert_and_key(self.tls_cert_path.as_ref(), self.tls_key_path.as_ref());
builder.build()
}
/// G9i pool admission + fast-P routing config; None when the feature is off.
fn build_g9i_config(&self) -> ConfigResult<Option<smg::routers::http::g9i::G9iConfig>> {
if !self.g9i_enabled && self.g9i_fast_prefill_url.is_empty() {
return Ok(None);
}
let config = smg::routers::http::g9i::G9iConfig {
enabled: self.g9i_enabled,
fast_prefill_urls: self.g9i_fast_prefill_url.clone(),
max_new_tokens: self.g9i_max_new_tokens,
fast_p_queue_reject_secs: self.g9i_fast_p_queue_reject_secs,
regular_p_queue_reject_secs: self.g9i_regular_p_queue_reject_secs,
decode_max_running: self.g9i_decode_max_running,
drain_rate_floor: self.g9i_drain_rate_floor,
poll_interval_secs: self.g9i_poll_interval_secs,
snapshot_ttl_secs: self.g9i_snapshot_ttl_secs,
d_no_snapshot: self.g9i_d_no_snapshot.clone(),
cache_hit_query_timeout_ms: self.g9i_cache_hit_query_timeout_ms,
regular_pool_least_wait: self.g9i_regular_pool_least_wait,
ttft_window_s: self.g9i_ttft_window_s,
shadow_enabled: self.g9i_shadow_enabled,
shadow_table_max_entries: self.g9i_shadow_table_max_entries,
shadow_table_ttl_secs: self.g9i_shadow_table_ttl_secs,
est_td_enabled: self.g9i_est_td_enabled,
shadow_route: self.g9i_shadow_route,
td_admit: self.g9i_td_admit.clone(),
td_gate_secs: self.g9i_td_gate_secs,
t_prefer_secs: self.g9i_t_prefer_secs,
p_queue_token_cap: self.g9i_p_queue_token_cap,
p_queue_reqs_cap: self.g9i_p_queue_reqs_cap,
admit_hold_secs: self.g9i_admit_hold_secs,
admit_hold_max: self.g9i_admit_hold_max,
pipeline_ttl_secs: self.g9i_pipeline_ttl_secs,
streaming_ttl_secs: self.g9i_streaming_ttl_secs,
ledger_reconcile: self.g9i_ledger_reconcile,
window_init: self.g9i_window_init,
window_add: self.g9i_window_add,
window_fixed: self.g9i_window_fixed,
prompt_validation: self.g9i_prompt_validation,
prompt_token_limit: self.g9i_prompt_token_limit,
hot_policy: self.g9i_hot_policy.clone(),
reservoir_max_ms: self.g9i_reservoir_max_ms,
reservoir_cap: self.g9i_reservoir_cap,
hot_reserve_frac: self.g9i_hot_reserve_frac,
};
config.validate().map_err(|reason| ConfigError::ValidationFailed { reason })?;
Ok(Some(config))
}
fn to_server_config(&self, router_config: RouterConfig) -> ServerConfig {
let service_discovery_config = if self.service_discovery {
// Get router discovery config from router_config.discovery if available
+126 -2
View File
@@ -334,6 +334,64 @@ pub(crate) fn init_metrics() {
smg_mesh::metrics::init_mesh_metrics();
}
/// T25b: explicit Prometheus histogram buckets for every gateway histogram
/// NOT covered by the `duration_seconds` suffix matcher (those keep the
/// configurable duration_buckets). Root cause being fixed: without explicit
/// buckets the exporter renders histograms as client-side SUMMARIES
/// (DDSketch over a 3×20s rolling window); in the v16 deployment those
/// summaries exported all-zero quantiles while _sum/_count stayed healthy
/// (the rolling-window path silently drops samples when the quanta clock
/// reading falls behind the newest bucket begin, poisoning the summary
/// permanently). Explicit buckets render real Prometheus histograms —
/// aggregable across time and replicas with histogram_quantile, and immune
/// to the summary path entirely.
///
/// Groups by scale (keep in sync when adding a histogram!):
/// * TTFT-scale seconds (0.5–120s): TTFT, e2e, est_T_P/T_D, residuals
/// (signed values fall into the lowest bucket — acceptable),
/// bootstrap actuals;
/// * fine seconds (5ms–5s): TPOT, hold/reservoir waits, retry backoff;
/// * prefix lookup depth (ring count, 1–400);
/// * shadow-table evicted entry age (1s–1d).
fn histogram_bucket_overrides() -> Vec<(Matcher, Vec<f64>)> {
let ttft: Vec<f64> = vec![0.5, 1.0, 2.0, 5.0, 10.0, 20.0, 30.0, 60.0, 120.0];
let fine: Vec<f64> = vec![0.005, 0.01, 0.025, 0.05, 0.1, 0.25, 0.5, 1.0, 2.0, 5.0];
let depth: Vec<f64> = vec![1.0, 2.0, 3.0, 5.0, 10.0, 20.0, 50.0, 100.0, 200.0, 400.0];
let age: Vec<f64> = vec![1.0, 10.0, 60.0, 300.0, 900.0, 3600.0, 21600.0, 86400.0];
let mut out: Vec<(Matcher, Vec<f64>)> = Vec::new();
for name in [
"smg_g9i_ttft_seconds",
"smg_g9i_class_ttft_seconds",
"smg_g9i_nonstream_e2e_secs",
"smg_g9i_est_tp_secs",
"smg_g9i_est_td_secs",
"smg_g9i_est_td_slot_secs",
"smg_g9i_shadow_ttft_est_error_seconds",
"smg_g9i_est_vs_actual_residual",
"smg_g9i_bootstrap_actual_secs",
"smg_router_ttft_seconds",
] {
out.push((Matcher::Full(name.to_string()), ttft.clone()));
}
for name in [
"smg_router_tpot_seconds",
"smg_g9i_admit_hold_secs",
"smg_g9i_reservoir_wait_seconds",
"smg_worker_retry_backoff_seconds",
] {
out.push((Matcher::Full(name.to_string()), fine.clone()));
}
out.push((
Matcher::Full("smg_g9i_shadow_prefix_lookup_depth".to_string()),
depth,
));
out.push((
Matcher::Full("smg_g9i_shadow_prefix_table_evicted_age_seconds".to_string()),
age,
));
out
}
pub fn start_prometheus(config: PrometheusConfig) {
init_metrics();
@@ -351,11 +409,20 @@ pub fn start_prometheus(config: PrometheusConfig) {
.unwrap_or(IpAddr::V4(Ipv4Addr::new(0, 0, 0, 0)));
let socket_addr = SocketAddr::new(ip_addr, config.port);
PrometheusBuilder::new()
let mut builder = PrometheusBuilder::new()
.with_http_listener(socket_addr)
.upkeep_timeout(Duration::from_secs(5 * 60))
.set_buckets_for_metric(duration_matcher, &duration_bucket)
.expect("failed to set duration bucket")
.expect("failed to set duration bucket");
// T25b: every remaining gateway histogram gets explicit buckets (see
// histogram_bucket_overrides) — no metric falls back to the broken
// client-side summary path.
for (matcher, buckets) in histogram_bucket_overrides() {
builder = builder
.set_buckets_for_metric(matcher, &buckets)
.expect("failed to set g9i histogram buckets");
}
builder
.install()
.expect("failed to install Prometheus metrics exporter");
}
@@ -1513,4 +1580,61 @@ mod tests {
assert_eq!(method_to_static_str("POST"), "POST");
assert_eq!(method_to_static_str("UNKNOWN"), "OTHER");
}
/// T25b smoke: with the production bucket overrides applied, g9i
/// histograms render as real Prometheus histograms with non-zero le
/// bucket counts (histogram_quantile-able) — never the client-side
/// summary path that exported all-zero quantiles in v16.
#[test]
fn t25b_g9i_histograms_render_bucketed() {
let mut builder = PrometheusBuilder::new();
for (matcher, buckets) in histogram_bucket_overrides() {
builder = builder
.set_buckets_for_metric(matcher, &buckets)
.expect("bucket override must apply");
}
let recorder = builder.build_recorder();
let handle = recorder.handle();
metrics::with_local_recorder(&recorder, || {
let ttft = histogram!("smg_g9i_class_ttft_seconds", "class" => "hot");
for v in [0.8, 1.5, 3.0, 12.0, 45.0] {
ttft.record(v);
}
histogram!("smg_g9i_shadow_prefix_lookup_depth").record(7.0);
histogram!("smg_g9i_shadow_prefix_lookup_depth").record(120.0);
});
let out = handle.render();
// Real histogram exposition, not summary.
assert!(
out.contains("# TYPE smg_g9i_class_ttft_seconds histogram"),
"class ttft renders as histogram:\n{out}"
);
assert!(
!out.contains("smg_g9i_class_ttft_seconds{quantile="),
"no summary quantile lines:\n{out}"
);
// 3 of the 5 samples (0.8, 1.5, 3.0) land at or below le=5.
assert!(
out.contains("smg_g9i_class_ttft_seconds_bucket{class=\"hot\",le=\"5\"} 3"),
"le=5 bucket counts 3 samples:\n{out}"
);
assert!(
out.contains("smg_g9i_class_ttft_seconds_bucket{class=\"hot\",le=\"+Inf\"} 5"),
"+Inf bucket counts all 5:\n{out}"
);
// lookup_depth: 7 -> le=10; both samples (7, 120) <= 200
// (Prometheus buckets are cumulative).
assert!(
out.contains("smg_g9i_shadow_prefix_lookup_depth_bucket{le=\"10\"} 1"),
"depth 7 lands in le=10:\n{out}"
);
assert!(
out.contains("smg_g9i_shadow_prefix_lookup_depth_bucket{le=\"200\"} 2"),
"both depths land at or below le=200 (cumulative):\n{out}"
);
assert!(
out.contains("smg_g9i_shadow_prefix_lookup_depth_count 2"),
"count line present:\n{out}"
);
}
}
File diff suppressed because it is too large Load Diff
@@ -0,0 +1,979 @@
//! G9i v2 P1 shadow mode — T1: prefix hash chain (pure function layer).
//!
//! Design source: `experiments/g9i/g9i-v2-design-20261013.md` §3.2/§4.1,
//! content-handling constraints C4/C10. This module is **read-only with
//! respect to routing**: it computes hashes, it never touches dispatch.
//!
//! Why a chain: a chat body with `m` messages has exactly `m` prefix
//! boundaries, and the session bookkeeping table ("prefix → last serving
//! node + exact token count") needs a stable key per boundary. Rolling the
//! hash forward per message gives all `m` keys in one pass; lookup then
//! probes from deep to shallow and takes the deepest hit. In practice a new
//! turn is `previous body + assistant reply + new user message`, so hits
//! usually land at `h[m-2]`.
//!
//! seed = xxh3(canonical_json({model, chat_template_kwargs, tools_hash, system}))
//! h[0] = seed
//! h[k] = xxh3(h[k-1] LE bytes ‖ len(role) ‖ role ‖ tagged content
//! ‖ tagged canonical(tool_calls) ‖ tagged meta) k = 1..=m
//!
//! Encoding decisions (why, not what):
//! * Streaming `Xxh3` instead of concatenating into a `Vec`: chat bodies
//! with base64 images are routinely hundreds of KB; copying them per
//! chain step would be pure waste.
//! * Every variable-length field is length-prefixed (u64 LE) and every
//! content piece carries a type tag. Without that, `["ab","c"]` and
//! `["a","bc"]` (or role/content, part/part splits) would hash
//! identically — silent false cache hits.
//! * Missing role defaults to "user" (OpenAI clients routinely omit it);
//! missing `model` / `chat_template_kwargs` / `tools` / system message
//! serialize as JSON `null` in the seed document — a fixed placeholder
//! that cannot collide with any real value because `null` is not a
//! string/object the field could otherwise hold.
//! * Byte-level semantics for multimodal (C4): `image_url` URLs are
//! hashed as the raw string — `http(s):` URLs as-is, `data:` URLs
//! including the base64 payload, **never fetched, never decoded**.
//! Same image at different URLs hashes differently; that is correct —
//! the rendered prompt bytes differ.
//! * `is_unclassifiable` and `chain_hashes` share one traversal
//! (`walk_messages`) so the "weird content" judgement can never drift
//! between the two call sites (C10). Unknown part types still
//! contribute a deterministic tag to the chain — the chain stays
//! well-defined, the caller just must not trust it for table lookups
//! when `is_unclassifiable` is true.
//! * `content: null` / missing renders as empty content in every chat
//! template, so it hashes exactly like `""` — NOT unclassifiable.
//! Flagging it would disenfranchise tool-use conversations (assistant
//! turns with `content: null` + `tool_calls`), our main traffic shape.
//! Genuinely malformed content (number/object/bool) keeps the
//! Empty + unclassifiable treatment.
//! * `tool_calls` are rendered into the prompt by the chat template, so
//! they MUST enter the chain (as canonical JSON, key-order
//! insensitive) — otherwise two conversations with different tool
//! calls could share a hash: a false hit in the dangerous direction.
//! `reasoning_content` is deliberately excluded: templates never
//! render it, so it cannot change prompt bytes.
//! * `tool_call_id` (on `role:"tool"` results) and `name` (any role)
//! are rendered by most templates, so they enter the chain as one
//! canonical meta object. The META segment is unconditional — a
//! message without these fields contributes `TAG_META ‖ len=0`,
//! keeping the formula single-path (no presence/absence ambiguity).
//!
//! T23 (intra-message chunking): a text piece longer than `CHUNK_BYTES`
//! (2048) is first rolled into bare chunk rings `xxh3(prev ‖ chunk)` —
//! no tag, no length prefix — in content order, and the message-final
//! ring then carries only the remaining ≤2048-byte tail in its content
//! segment. Why: a per-message chain is structurally blind inside one
//! long message (one message = one ring, so a shared 100KB prefix with
//! a different tail hits nothing — v13 measured 9034/9034 depth-0
//! misses on single-message long-prompt traffic). Chunk rings give the
//! lookup table mid-message anchors at a fixed granularity. Texts of
//! ≤2048 bytes emit no chunk rings, so short-message chains stay
//! byte-identical to the pre-T23 formula (pinned by
//! `short_messages_match_legacy_chain_byte_for_byte`). Because one
//! message may now own several rings, `chain_hashes_indexed` also
//! returns each ring's message index; `classify` needs it to map a
//! hit's ring position back into the message index space of
//! `msg_bytes` / `msg_image`.
use serde_json::Value;
use xxhash_rust::xxh3::{xxh3_64, Xxh3};
/// Part-type tags mixed in before each content piece. Terminated with `\0`
/// so tags are self-delimiting; content bytes are length-prefixed on top.
const TAG_TEXT: &[u8] = b"text\0";
const TAG_IMAGE_URL: &[u8] = b"image_url\0";
const TAG_UNKNOWN: &[u8] = b"unknown\0";
const TAG_TOOL_CALLS: &[u8] = b"tool_calls\0";
const TAG_META: &[u8] = b"meta\0";
/// T23: intra-message chunk size in bytes. A text piece longer than this
/// contributes `ceil`-style bare chunk rings before its message-final
/// ring (see module docs). 2048 bytes ≈ 500-600 tokens of English/JSON —
/// fine enough to rescue shared-prefix hits, coarse enough that typical
/// multi-turn chats (short messages) never pay the extra hash passes.
const CHUNK_BYTES: usize = 2048;
/// Chain seed: folds every body field that changes the rendered prompt but
/// lives outside `messages` into one u64.
#[derive(Debug, Clone, Copy, PartialEq, Eq, Hash)]
pub struct ChainSeed {
pub value: u64,
}
impl ChainSeed {
/// Build the seed from a chat request body.
///
/// Inputs: `model`, `chat_template_kwargs`, `tools` (folded to a single
/// hash of its canonical form), and the first `role=="system"` message's
/// string content. Missing fields become JSON `null` in the seed
/// document (fixed placeholder, see module docs). Canonicalization
/// sorts object keys recursively and emits no whitespace, so semantically
/// equal bodies from different clients produce the same seed.
///
/// Note: a system message is *also* hashed as `messages[0]` inside the
/// chain — deliberately. The chain must stay 1:1 with `messages`
/// indices (the lookup table keys are `h[k]` by position), and the
/// redundancy only costs one extra mix of already-known bytes.
pub fn from_body(body: &Value) -> ChainSeed {
let system_content: Option<&str> = body
.get("messages")
.and_then(Value::as_array)
.and_then(|msgs| {
msgs.iter()
.find(|m| m.get("role").and_then(Value::as_str) == Some("system"))
})
.and_then(|m| m.get("content"))
.and_then(Value::as_str);
let tools_hash: Option<u64> = body
.get("tools")
.map(|t| xxh3_64(canonical_json(t).as_bytes()));
let seed_doc = serde_json::json!({
"model": body.get("model").cloned().unwrap_or(Value::Null),
"chat_template_kwargs": body
.get("chat_template_kwargs")
.cloned()
.unwrap_or(Value::Null),
"tools_hash": tools_hash,
"system": system_content,
});
ChainSeed {
value: xxh3_64(canonical_json(&seed_doc).as_bytes()),
}
}
}
/// Compute the chain hashes `h[1..=m]`, one ring per message (plus T23
/// intra-message chunk rings), in order.
///
/// `messages` missing / not an array / empty → empty `Vec`. The caller
/// treats an empty chain as "cannot classify" (same fall-back path as
/// `is_unclassifiable`); we never panic on malformed bodies because this
/// sits on the hot request path.
///
/// The returned chain is deterministic even when `is_unclassifiable(body)`
/// is true — unknown parts contribute their tag — but such chains are only
/// suitable for logging/shadow comparison, not for bookkeeping lookups.
///
/// This is the chain-only view of `chain_hashes_indexed`; callers that
/// index into per-message arrays (`msg_bytes`, `msg_image`) need the
/// indexed form to translate ring positions to message indices.
pub fn chain_hashes(seed: &ChainSeed, body: &Value) -> Vec<u64> {
chain_hashes_indexed(seed, body).0
}
/// `chain_hashes` plus, for each ring, the index of the message it
/// belongs to (chunk rings and the message-final ring of message `k`
/// all carry `k`). Both returned vecs always have the same length.
pub fn chain_hashes_indexed(seed: &ChainSeed, body: &Value) -> (Vec<u64>, Vec<usize>) {
let (msgs, _) = walk_messages(body);
let mut chain = Vec::with_capacity(msgs.len());
let mut ring_msg = Vec::with_capacity(msgs.len());
let mut prev = seed.value;
for (mi, msg) in msgs.iter().enumerate() {
// T23 pass 1: roll oversized text pieces into bare chunk rings
// (h = xxh3(prev ‖ chunk), no tag/length) and keep the ≤2048-byte
// tails for the message-final ring. Short texts emit nothing here
// and their tails ARE the originals — the pre-T23 byte stream.
let content = match &msg.content {
ContentView::Str(s) => {
ContentView::Str(roll_chunks(&mut prev, &mut chain, &mut ring_msg, mi, s))
}
ContentView::Parts(parts) => {
let mut viewed = Vec::with_capacity(parts.len());
for part in parts {
viewed.push(match part {
PartView::Text(t) => PartView::Text(roll_chunks(
&mut prev,
&mut chain,
&mut ring_msg,
mi,
t,
)),
PartView::ImageUrl(u) => PartView::ImageUrl(u),
PartView::Unknown => PartView::Unknown,
});
}
ContentView::Parts(viewed)
}
ContentView::Empty => ContentView::Empty,
};
// Pass 2: the message-final ring — the pre-T23 formula unchanged,
// except its content segment now hashes the pass-1 tails.
let mut h = Xxh3::new();
h.update(&prev.to_le_bytes());
let role = msg.role.as_bytes();
h.update(&(role.len() as u64).to_le_bytes());
h.update(role);
match &content {
ContentView::Str(s) => {
update_tagged(&mut h, TAG_TEXT, s.as_bytes());
}
ContentView::Parts(parts) => {
for part in parts {
match part {
PartView::Text(t) => update_tagged(&mut h, TAG_TEXT, t.as_bytes()),
PartView::ImageUrl(u) => {
update_tagged(&mut h, TAG_IMAGE_URL, u.as_bytes())
}
PartView::Unknown => h.update(TAG_UNKNOWN),
}
}
}
ContentView::Empty => {}
}
// Third segment: tool_calls are rendered into prompt bytes by the
// chat template, so they must be in the chain.
if let Some(tc) = &msg.tool_calls {
update_tagged(&mut h, TAG_TOOL_CALLS, tc.as_bytes());
}
// Fourth segment (unconditional): name / tool_call_id meta. Empty
// meta hashes as TAG_META ‖ len=0 — one formula for all messages.
update_tagged(&mut h, TAG_META, msg.extra.as_bytes());
prev = h.digest();
chain.push(prev);
ring_msg.push(mi);
}
(chain, ring_msg)
}
/// Roll `prev` forward over `text`'s full chunks, emitting one bare ring
/// per chunk (`xxh3(prev ‖ chunk bytes)`), and return the remaining tail
/// (≤ CHUNK_BYTES, possibly empty) for the message-final ring. Texts of
/// ≤ CHUNK_BYTES emit nothing and are returned untouched.
fn roll_chunks<'a>(
prev: &mut u64,
chain: &mut Vec<u64>,
ring_msg: &mut Vec<usize>,
mi: usize,
text: &'a str,
) -> &'a str {
let mut rest = text;
while rest.len() > CHUNK_BYTES {
let (chunk, tail) = take_chunk(rest);
let mut h = Xxh3::new();
h.update(&prev.to_le_bytes());
h.update(chunk.as_bytes());
*prev = h.digest();
chain.push(*prev);
ring_msg.push(mi);
rest = tail;
}
rest
}
/// Split off the leading chunk of at most CHUNK_BYTES bytes, backing off
/// to the previous UTF-8 char boundary when the cut lands mid-character.
/// The caller guarantees `s.len() > CHUNK_BYTES`, so the boundary walk
/// can never reach 0 (the first char is at most 4 bytes).
fn take_chunk(s: &str) -> (&str, &str) {
debug_assert!(s.len() > CHUNK_BYTES);
let mut end = CHUNK_BYTES;
while !s.is_char_boundary(end) {
end -= 1;
}
s.split_at(end)
}
/// Whether the body contains content this module has no hashing semantics
/// for: unknown part types (`audio_url`, `video_url`, `file`, ...),
/// malformed parts, unsupported `image_url` schemes, or genuinely
/// malformed message content (number/object/bool). Null/missing content
/// and `tool_calls`-carrying assistant turns return false — both have
/// well-defined rendering semantics. See `walk_messages` — this is the
/// same traversal, so the answer cannot disagree with what `chain_hashes`
/// actually hashed.
pub fn is_unclassifiable(body: &Value) -> bool {
walk_messages(body).1
}
/// Mix `tag ‖ len(bytes) ‖ bytes` into the hasher. Length prefix kills
/// concatenation ambiguity between adjacent pieces.
fn update_tagged(h: &mut Xxh3, tag: &[u8], bytes: &[u8]) {
h.update(tag);
h.update(&(bytes.len() as u64).to_le_bytes());
h.update(bytes);
}
/// One message, flattened to just what the hash needs. Borrows from the
/// body — zero copies on the hot path, except `tool_calls` / `extra`
/// whose canonical forms must be owned.
struct MessageView<'a> {
role: &'a str,
content: ContentView<'a>,
/// Canonical JSON of `tool_calls` when the field exists and is an
/// array (a non-array shape can't be rendered by the template either,
/// so it's treated as absent).
tool_calls: Option<String>,
/// Canonical JSON of the meta object {"name"?, "tool_call_id"?}, or
/// empty when neither field applies. See `walk_messages` for which
/// fields qualify per role.
extra: String,
}
enum ContentView<'a> {
Str(&'a str),
Parts(Vec<PartView<'a>>),
/// Genuinely malformed content (number/object/bool): contributes
/// nothing, flagged unclassifiable. Null/missing content never lands
/// here — it maps to `Str("")`.
Empty,
}
enum PartView<'a> {
Text(&'a str),
ImageUrl(&'a str),
/// Unrecognized or malformed part: deterministic placeholder.
Unknown,
}
/// Single traversal shared by `chain_hashes` and `is_unclassifiable`
/// (C10: one source of truth for "what can we hash").
fn walk_messages(body: &Value) -> (Vec<MessageView<'_>>, bool) {
let Some(msgs) = body.get("messages").and_then(Value::as_array) else {
return (Vec::new(), false);
};
let mut unclassifiable = false;
let mut out = Vec::with_capacity(msgs.len());
for m in msgs {
let role = m.get("role").and_then(Value::as_str).unwrap_or("user");
// `reasoning_content` is deliberately NOT hashed: chat templates do
// not render it, so it cannot change prompt bytes — mixing it in
// would only manufacture misses.
let tool_calls = match m.get("tool_calls") {
Some(v @ Value::Array(_)) => Some(canonical_json(v)),
_ => None,
};
// Meta segment: `name` (any role) and `tool_call_id` (tool results
// only) are rendered by most templates -> they must be in the
// chain. String and non-string values alike go through the
// canonical serializer, so wire-shape differences can't collide.
// Explicit null counts as absent (templates render it as nothing).
let mut meta = serde_json::Map::new();
if let Some(v) = m.get("name").filter(|v| !v.is_null()) {
meta.insert("name".to_string(), v.clone());
}
if role == "tool" {
if let Some(v) = m.get("tool_call_id").filter(|v| !v.is_null()) {
meta.insert("tool_call_id".to_string(), v.clone());
}
}
let extra = if meta.is_empty() {
String::new()
} else {
canonical_json(&Value::Object(meta))
};
let content = match m.get("content") {
Some(Value::String(s)) => ContentView::Str(s.as_str()),
Some(Value::Array(parts)) => {
let mut views = Vec::with_capacity(parts.len());
for p in parts {
match classify_part(p) {
Some(view) => views.push(view),
None => {
unclassifiable = true;
views.push(PartView::Unknown);
}
}
}
ContentView::Parts(views)
}
// null / absent renders as empty content in every chat
// template — hash exactly like "" and stay classifiable. This
// is the tool-use conversation shape (assistant turn with
// content:null + tool_calls), our main traffic.
None | Some(Value::Null) => ContentView::Str(""),
// number/object/bool content: genuinely malformed, no defined
// rendering semantics — fail closed.
_ => {
unclassifiable = true;
ContentView::Empty
}
};
out.push(MessageView {
role,
content,
tool_calls,
extra,
});
}
(out, unclassifiable)
}
/// Classify one content part. `None` = no hashing semantics for this part
/// (unknown type, malformed shape, or an `image_url` scheme outside
/// http/https/data — e.g. `file://` would tempt a local read we must never
/// do).
fn classify_part(p: &Value) -> Option<PartView<'_>> {
let ty = p.get("type").and_then(Value::as_str)?;
match ty {
"text" => p
.get("text")
.and_then(Value::as_str)
.map(PartView::Text),
"image_url" => {
// Both wire shapes exist in the wild:
// {"image_url":{"url":...}} and {"image_url":"..."}.
let url = match p.get("image_url") {
Some(Value::String(s)) => Some(s.as_str()),
Some(v @ Value::Object(_)) => v.get("url").and_then(Value::as_str),
_ => None,
}?;
if url.starts_with("http://") || url.starts_with("https://") || url.starts_with("data:")
{
Some(PartView::ImageUrl(url))
} else {
None
}
}
_ => None,
}
}
/// Canonical JSON: object keys sorted recursively, no whitespace. Hand-rolled
/// (no new crate) because we only need it for seed inputs, which are small —
/// `tools` is hashed from its canonical form, and the seed doc itself is
/// four fields.
fn canonical_json(v: &Value) -> String {
let mut out = String::new();
write_canonical(v, &mut out);
out
}
fn write_canonical(v: &Value, out: &mut String) {
match v {
Value::Object(map) => {
out.push('{');
let mut keys: Vec<&String> = map.keys().collect();
keys.sort();
for (i, k) in keys.iter().enumerate() {
if i > 0 {
out.push(',');
}
// Serializing a &str Value is infallible; avoid unwrap anyway.
if let Ok(k_json) = serde_json::to_string(k) {
out.push_str(&k_json);
}
out.push(':');
write_canonical(&map[*k], out);
}
out.push('}');
}
Value::Array(arr) => {
out.push('[');
for (i, item) in arr.iter().enumerate() {
if i > 0 {
out.push(',');
}
write_canonical(item, out);
}
out.push(']');
}
other => {
if let Ok(s) = serde_json::to_string(other) {
out.push_str(&s);
}
}
}
}
#[cfg(test)]
mod tests {
use super::*;
use serde_json::json;
fn text_body(msgs: Vec<Value>) -> Value {
json!({"model": "dsv41-flash", "messages": msgs})
}
fn user(t: &str) -> Value {
json!({"role": "user", "content": t})
}
fn assistant(t: &str) -> Value {
json!({"role": "assistant", "content": t})
}
fn chain(body: &Value) -> Vec<u64> {
chain_hashes(&ChainSeed::from_body(body), body)
}
#[test]
fn same_prefix_different_tail() {
let a = text_body(vec![user("hello"), assistant("hi"), user("tail A")]);
let b = text_body(vec![user("hello"), assistant("hi"), user("tail B")]);
let ca = chain(&a);
let cb = chain(&b);
assert_eq!(ca.len(), 3);
assert_eq!(cb.len(), 3);
assert_eq!(ca[..2], cb[..2], "shared prefix must hash identically");
assert_ne!(ca[2], cb[2], "different tail must diverge");
}
#[test]
fn forked_conversation_shares_prefix() {
let base = vec![user("q1"), assistant("a1"), user("q2"), assistant("a2")];
let mut fork_x = base.clone();
fork_x.push(user("branch X"));
let mut fork_y = base;
fork_y.push(user("branch Y"));
let bx = text_body(fork_x);
let by = text_body(fork_y);
let cx = chain(&bx);
let cy = chain(&by);
assert_eq!(cx[..4], cy[..4]);
assert_ne!(cx[4], cy[4]);
}
#[test]
fn editing_history_rewrites_from_that_point_on() {
let a = text_body(vec![user("q1"), assistant("old answer"), user("q2")]);
let b = text_body(vec![user("q1"), assistant("edited answer"), user("q2")]);
let ca = chain(&a);
let cb = chain(&b);
assert_eq!(ca[0], cb[0], "before the edit: unchanged");
assert_ne!(ca[1], cb[1], "at the edit: changed");
assert_ne!(ca[2], cb[2], "after the edit: still changed even with identical text");
}
#[test]
fn multimodal_byte_level_semantics() {
let mk = |url: &str| {
text_body(vec![json!({
"role": "user",
"content": [
{"type": "text", "text": "what is in this image?"},
{"type": "image_url", "image_url": {"url": url}}
]
})])
};
// Same image content, different URLs -> different chain (the rendered
// prompt bytes differ; this is the correct byte-level behaviour).
let http_a = mk("https://example.com/a.png");
let http_b = mk("https://example.com/b.png");
assert_ne!(chain(&http_a), chain(&http_b));
// Identical base64 payload -> identical chain (hit).
let data_1 = mk("data:image/png;base64,iVBORw0KGgoAAAANSUhEUg==");
let data_2 = mk("data:image/png;base64,iVBORw0KGgoAAAANSUhEUg==");
assert_eq!(chain(&data_1), chain(&data_2));
// base64 vs http form of the "same" image -> different (bytes differ).
assert_ne!(chain(&data_1), chain(&http_a));
// Wire-shape variant {"image_url":"..."} hashes like the object form.
let flat = text_body(vec![json!({
"role": "user",
"content": [
{"type": "text", "text": "what is in this image?"},
{"type": "image_url", "image_url": "https://example.com/a.png"}
]
})]);
assert_eq!(chain(&flat), chain(&http_a));
}
#[test]
fn seed_fields_change_whole_chain() {
let msgs = || vec![user("same"), assistant("same")];
let base = text_body(msgs());
let other_model = json!({"model": "other-model", "messages": msgs()});
let with_kwargs = json!({
"model": "dsv41-flash",
"chat_template_kwargs": {"thinking": true},
"messages": msgs()
});
let with_tools = json!({
"model": "dsv41-flash",
"tools": [{"type": "function", "function": {"name": "f"}}],
"messages": msgs()
});
let c_base = chain(&base);
assert_ne!(ChainSeed::from_body(&base), ChainSeed::from_body(&other_model));
assert_ne!(chain(&other_model), c_base);
assert_ne!(chain(&with_kwargs), c_base);
assert_ne!(chain(&with_tools), c_base);
}
#[test]
fn seed_canonicalization_is_key_order_insensitive() {
let a = json!({"model": "m", "chat_template_kwargs": {"x": 1, "y": 2}, "messages": [user("hi")]});
let b = json!({"messages": [user("hi")], "chat_template_kwargs": {"y": 2, "x": 1}, "model": "m"});
assert_eq!(ChainSeed::from_body(&a), ChainSeed::from_body(&b));
assert_eq!(chain(&a), chain(&b));
}
#[test]
fn missing_or_malformed_messages_give_empty_chain() {
let no_messages = json!({"model": "m"});
let empty_messages = json!({"model": "m", "messages": []});
let wrong_type = json!({"model": "m", "messages": "not-an-array"});
for b in [&no_messages, &empty_messages, &wrong_type] {
assert!(chain(b).is_empty());
}
}
#[test]
fn unknown_part_types_flag_unclassifiable() {
let audio = text_body(vec![json!({
"role": "user",
"content": [{"type": "audio_url", "audio_url": {"url": "https://x/a.mp3"}}]
})]);
assert!(is_unclassifiable(&audio));
// Chain is still full-length and deterministic — just not to be trusted.
assert_eq!(chain(&audio).len(), 1);
assert_eq!(chain(&audio), chain(&audio));
let file_scheme = text_body(vec![json!({
"role": "user",
"content": [{"type": "image_url", "image_url": {"url": "file:///etc/passwd"}}]
})]);
assert!(is_unclassifiable(&file_scheme));
let plain = text_body(vec![user("hello"), assistant("hi")]);
assert!(!is_unclassifiable(&plain));
let multimodal_ok = text_body(vec![json!({
"role": "user",
"content": [
{"type": "text", "text": "look"},
{"type": "image_url", "image_url": {"url": "https://x/a.png"}}
]
})]);
assert!(!is_unclassifiable(&multimodal_ok));
}
#[test]
fn missing_role_defaults_to_user() {
let implicit = text_body(vec![json!({"content": "hi"})]);
let explicit = text_body(vec![user("hi")]);
assert_eq!(chain(&implicit), chain(&explicit));
assert!(!is_unclassifiable(&implicit));
}
fn tool_call(name: &str, args: &str) -> Value {
json!([{
"id": format!("call_{}", name),
"type": "function",
"function": {"name": name, "arguments": args}
}])
}
#[test]
fn tool_calls_enter_the_chain() {
let mk = |tc: Value| {
text_body(vec![
user("weather in Shanghai?"),
json!({"role": "assistant", "content": null, "tool_calls": tc}),
user("and tomorrow?"),
])
};
let a = mk(tool_call("get_weather", "{\"city\":\"sh\"}"));
let b = mk(tool_call("get_weather", "{\"city\":\"bj\"}"));
let c = mk(tool_call("get_weather", "{\"city\":\"sh\"}"));
let (ca, cb, cc) = (chain(&a), chain(&b), chain(&c));
assert_eq!(ca[0], cb[0], "before the tool call: unchanged");
assert_ne!(ca[1], cb[1], "different tool_calls must diverge");
assert_ne!(ca[2], cb[2], "and stay diverged afterwards");
assert_eq!(ca, cc, "identical tool_calls must hit");
}
#[test]
fn content_null_with_tool_calls_is_classifiable() {
let tc = tool_call("get_weather", "{}");
let null_content = text_body(vec![json!({
"role": "assistant", "content": null, "tool_calls": tc.clone()
})]);
let empty_content = text_body(vec![json!({
"role": "assistant", "content": "", "tool_calls": tc
})]);
assert!(!is_unclassifiable(&null_content));
assert_eq!(
chain(&null_content),
chain(&empty_content),
"content:null renders as empty -> must hash like content:\"\""
);
}
#[test]
fn content_null_without_tool_calls_is_empty() {
let null_content = text_body(vec![json!({"role": "assistant", "content": null})]);
let missing_content = text_body(vec![json!({"role": "assistant"})]);
let empty_content = text_body(vec![assistant("")]);
assert!(!is_unclassifiable(&null_content));
assert!(!is_unclassifiable(&missing_content));
assert_eq!(chain(&null_content), chain(&empty_content));
assert_eq!(chain(&missing_content), chain(&empty_content));
// Genuinely malformed content types keep the fail-closed treatment.
let malformed = text_body(vec![json!({"role": "assistant", "content": 42})]);
assert!(is_unclassifiable(&malformed));
}
#[test]
fn tool_calls_canonicalization_key_order_insensitive() {
let a = text_body(vec![json!({
"role": "assistant",
"content": null,
"tool_calls": [{
"id": "call_1",
"type": "function",
"function": {"name": "f", "arguments": "{}"}
}]
})]);
let b = text_body(vec![json!({
"role": "assistant",
"content": null,
"tool_calls": [{
"function": {"arguments": "{}", "name": "f"},
"type": "function",
"id": "call_1"
}]
})]);
assert_eq!(chain(&a), chain(&b));
}
#[test]
fn tool_call_id_enters_chain() {
let mk = |id: &str| {
text_body(vec![
user("weather in Shanghai?"),
json!({
"role": "assistant",
"content": null,
"tool_calls": tool_call("get_weather", "{}")
}),
json!({"role": "tool", "tool_call_id": id, "content": "sunny"}),
user("thanks"),
])
};
let a = mk("call_1");
let b = mk("call_2");
let c = mk("call_1");
let (ca, cb, cc) = (chain(&a), chain(&b), chain(&c));
assert_eq!(ca[..2], cb[..2], "before the tool result: unchanged");
assert_ne!(ca[2], cb[2], "different tool_call_id must diverge");
assert_ne!(ca[3], cb[3], "and stay diverged afterwards");
assert_eq!(ca, cc, "identical tool_call_id must hit");
}
#[test]
fn name_field_enters_chain() {
let a = text_body(vec![json!({"role": "user", "name": "alice", "content": "hi"})]);
let b = text_body(vec![json!({"role": "user", "name": "bob", "content": "hi"})]);
let c = text_body(vec![json!({"role": "user", "name": "alice", "content": "hi"})]);
assert_ne!(chain(&a), chain(&b), "different name must diverge");
assert_eq!(chain(&a), chain(&c), "identical name must hit");
// Absent name and empty meta stay equivalent to a plain message.
let plain = text_body(vec![user("hi")]);
let no_name = text_body(vec![json!({"role": "user", "content": "hi"})]);
assert_eq!(chain(&plain), chain(&no_name));
}
/// `reasoning_content` must NOT change the chain: chat templates do not
/// render historical reasoning into the prompt, so hashing it would
/// manufacture misses for thinking-enabled clients that echo it back.
/// Pins the walk_messages assumption (thinking:true is our default).
#[test]
fn reasoning_content_does_not_enter_chain() {
let plain = text_body(vec![user("q"), assistant("a"), user("q2")]);
let with_reasoning = text_body(vec![
user("q"),
json!({"role": "assistant", "content": "a", "reasoning_content": "let me think..."}),
user("q2"),
]);
assert_eq!(
chain(&plain),
chain(&with_reasoning),
"reasoning_content is not rendered -> must not change the chain"
);
assert!(
!is_unclassifiable(&with_reasoning),
"reasoning-carrying turns stay classifiable"
);
}
// ==================== T23: intra-message chunking ====================
/// The pre-T23 `chain_hashes`, verbatim copy. Short-message bodies must
/// produce byte-identical chains under the chunked implementation.
fn legacy_chain_hashes(seed: &ChainSeed, body: &Value) -> Vec<u64> {
let (msgs, _) = walk_messages(body);
let mut chain = Vec::with_capacity(msgs.len());
let mut prev = seed.value;
for msg in &msgs {
let mut h = Xxh3::new();
h.update(&prev.to_le_bytes());
let role = msg.role.as_bytes();
h.update(&(role.len() as u64).to_le_bytes());
h.update(role);
match &msg.content {
ContentView::Str(s) => {
update_tagged(&mut h, TAG_TEXT, s.as_bytes());
}
ContentView::Parts(parts) => {
for part in parts {
match part {
PartView::Text(t) => update_tagged(&mut h, TAG_TEXT, t.as_bytes()),
PartView::ImageUrl(u) => {
update_tagged(&mut h, TAG_IMAGE_URL, u.as_bytes())
}
PartView::Unknown => h.update(TAG_UNKNOWN),
}
}
}
ContentView::Empty => {}
}
if let Some(tc) = &msg.tool_calls {
update_tagged(&mut h, TAG_TOOL_CALLS, tc.as_bytes());
}
update_tagged(&mut h, TAG_META, msg.extra.as_bytes());
prev = h.digest();
chain.push(prev);
}
chain
}
#[test]
fn short_messages_match_legacy_chain_byte_for_byte() {
let exact_boundary = "x".repeat(CHUNK_BYTES);
let bodies = vec![
// pure-text multi-turn
text_body(vec![user("q1"), assistant("a1"), user("q2")]),
// multimodal parts (short text + image)
text_body(vec![json!({
"role": "user",
"content": [
{"type": "text", "text": "look"},
{"type": "image_url", "image_url": {"url": "https://x/a.png"}}
]
})]),
// tool_calls + content:null + tool result (meta on both)
text_body(vec![
user("weather?"),
json!({"role": "assistant", "content": null,
"tool_calls": tool_call("get_weather", "{\"city\":\"sh\"}")}),
json!({"role": "tool", "tool_call_id": "call_1", "content": "sunny"}),
]),
// name meta
text_body(vec![json!({"role": "user", "name": "alice", "content": "hi"})]),
// null content without tool_calls
text_body(vec![json!({"role": "assistant", "content": null})]),
// exactly CHUNK_BYTES: never splits (only > CHUNK_BYTES does)
text_body(vec![user(&exact_boundary)]),
];
for b in &bodies {
let seed = ChainSeed::from_body(b);
assert_eq!(
chain_hashes(&seed, b),
legacy_chain_hashes(&seed, b),
"short-message chain must stay byte-identical to pre-T23"
);
}
}
#[test]
fn long_message_splits_into_chunk_rings() {
// CHUNK_BYTES + 1 -> one chunk ring + the message-final ring.
let text = "x".repeat(CHUNK_BYTES + 1);
let body = text_body(vec![user(&text)]);
let seed = ChainSeed::from_body(&body);
let (c, idx) = chain_hashes_indexed(&seed, &body);
assert_eq!(c.len(), 2);
assert_eq!(idx, vec![0, 0], "both rings belong to message 0");
// The chunk ring is bare: xxh3(prev_le ‖ chunk), no tag, no length.
let mut h = Xxh3::new();
h.update(&seed.value.to_le_bytes());
h.update("x".repeat(CHUNK_BYTES).as_bytes());
assert_eq!(c[0], h.digest());
// Exactly CHUNK_BYTES -> single ring (only > CHUNK_BYTES splits);
// already pinned by the legacy-equivalence battery above.
assert_eq!(chain(&text_body(vec![user(&"y".repeat(CHUNK_BYTES))])).len(), 1);
}
/// The T23 acceptance shape: two single-message bodies sharing a 4KB
/// prefix but differing in the tail now share their chunk rings — the
/// pre-T23 chain (one ring per message) could never hit here.
#[test]
fn shared_long_prefix_different_tail_shares_chunk_rings() {
let prefix = "a".repeat(2 * CHUNK_BYTES);
let a = text_body(vec![user(&format!("{prefix}{}", "b".repeat(100)))]);
let b = text_body(vec![user(&format!("{prefix}{}", "c".repeat(200)))]);
let (ca, ia) = chain_hashes_indexed(&ChainSeed::from_body(&a), &a);
let (cb, ib) = chain_hashes_indexed(&ChainSeed::from_body(&b), &b);
assert_eq!(ca.len(), 3, "2 chunk rings + message-final ring");
assert_eq!(cb.len(), 3);
assert_eq!(ia, vec![0, 0, 0]);
assert_eq!(ib, vec![0, 0, 0]);
assert_eq!(
ca[..2],
cb[..2],
"identical 4KB prefix -> identical chunk rings"
);
assert_ne!(ca[2], cb[2], "different tail -> final ring diverges");
}
#[test]
fn utf8_boundary_chunk_never_splits_a_character() {
// 700 x 3 bytes = 2100 > 2048; the 2048 cut lands mid-汉
// (2048 % 3 = 2), so the chunk must back off to 2046 bytes.
let text = "汉".repeat(700);
let (chunk, tail) = take_chunk(&text);
assert_eq!(chunk.len(), 2046);
assert_eq!(tail.len(), 54);
assert_eq!(format!("{chunk}{tail}"), text, "chunk + tail reconstruct");
let body = text_body(vec![user(&text)]);
let seed = ChainSeed::from_body(&body);
let (c, idx) = chain_hashes_indexed(&seed, &body);
assert_eq!(c.len(), 2);
assert_eq!(idx, vec![0, 0]);
// The chunk ring carries exactly the boundary-aligned bytes.
let mut h = Xxh3::new();
h.update(&seed.value.to_le_bytes());
h.update(chunk.as_bytes());
assert_eq!(c[0], h.digest());
}
#[test]
fn multi_part_long_text_chunks_per_part_in_order() {
let t1 = "p".repeat(CHUNK_BYTES + 10); // 1 chunk + 10-byte tail
let t2 = "q".repeat(CHUNK_BYTES + 20); // 1 chunk + 20-byte tail
let body = text_body(vec![json!({
"role": "user",
"content": [
{"type": "text", "text": t1},
{"type": "image_url", "image_url": {"url": "https://x/a.png"}},
{"type": "text", "text": t2}
]
})]);
let seed = ChainSeed::from_body(&body);
let (c, idx) = chain_hashes_indexed(&seed, &body);
assert_eq!(c.len(), 3, "t1 chunk + t2 chunk + message-final ring");
assert_eq!(idx, vec![0, 0, 0]);
// Chunk rings follow part order: t1's chunk first, then t2's.
let mut h = Xxh3::new();
h.update(&seed.value.to_le_bytes());
h.update("p".repeat(CHUNK_BYTES).as_bytes());
assert_eq!(c[0], h.digest());
let mut h2 = Xxh3::new();
h2.update(&c[0].to_le_bytes());
h2.update("q".repeat(CHUNK_BYTES).as_bytes());
assert_eq!(c[1], h2.digest());
}
}
File diff suppressed because it is too large Load Diff
@@ -1,5 +1,8 @@
//! HTTP router implementations
pub mod g9i;
pub mod g9i_prefix_chain;
pub mod g9i_shadow;
pub mod pd_router;
pub mod pd_types;
pub mod router;
+577 -11
View File
@@ -18,6 +18,7 @@ use serde_json::{json, Value};
use tokio_stream::wrappers::UnboundedReceiverStream;
use tracing::{debug, error, warn};
use super::g9i_shadow::{ShadowRequestContext, ShadowResponseContext};
use super::pd_types::api_path;
use crate::{
config::types::RetryConfig,
@@ -58,6 +59,8 @@ pub struct PDRouter {
pub retry_config: RetryConfig,
pub api_key: Option<String>,
pub enable_igw: bool,
/// G9i pool admission + fast-P routing state; None = feature off.
pub g9i: Option<Arc<super::g9i::G9iState>>,
}
struct PreparedWorkerRequest<'a> {
@@ -182,6 +185,14 @@ impl PDRouter {
}
pub async fn new(ctx: &Arc<crate::app_context::AppContext>) -> Result<Self, String> {
let g9i = match &ctx.router_config.g9i {
Some(config) if config.enabled => {
let state = Arc::new(super::g9i::G9iState::new(config.clone()));
state.start_monitor(Arc::clone(&ctx.worker_registry), ctx.client.clone());
Some(state)
}
_ => None,
};
Ok(PDRouter {
worker_registry: Arc::clone(&ctx.worker_registry),
policy_registry: Arc::clone(&ctx.policy_registry),
@@ -189,6 +200,7 @@ impl PDRouter {
retry_config: ctx.router_config.effective_retry_config(),
api_key: ctx.router_config.api_key.clone(),
enable_igw: ctx.router_config.enable_igw,
g9i,
})
}
@@ -366,6 +378,92 @@ impl PDRouter {
Ok((prefill_request, decode_request))
}
/// G9i admission + shadow booking, ONE attempt (T15: the caller retries
/// holdable rejections through the hold queue). None = G9i not
/// configured ("untracked"). Err = rejected (the caller counts +
/// answers 429); the rejections are structured so that a rejected
/// attempt never entered the shadow bookkeeping (T11 拒绝路径零记账).
/// T25-A: `prior_wait_secs` is the time the request already spent
/// parked in the reservoir across re-runs — it counts into TTFT, so
/// the re-judged est_TTFT carries it.
async fn g9i_admission<T: Serialize + Clone>(
&self,
context: &PDRequestContext<'_>,
original_request: &T,
prior_wait_secs: f64,
) -> Option<Result<(Option<String>, &'static str, Option<ShadowRequestContext>), super::g9i::RejectInfo>>
{
let g9i = self.g9i.as_ref()?;
if let Err(info) = g9i.check_decode_pool(&self.worker_registry).await {
return Some(Err(info));
}
if context.route != "/v1/chat/completions" {
// Non-chat routes skip classification but still land on
// the regular pool — the regular queue gate applies.
if let Err(info) = g9i.check_regular_queue(&self.worker_registry) {
return Some(Err(info));
}
return Some(Ok((None, "non_chat", None)));
}
let body = match serde_json::to_value(original_request) {
Ok(body) => body,
Err(_) => return Some(Ok((None, "regular", None))),
};
// Shadow T2: compute the prefix chain once, here, alongside the v1
// classification. Skipped entirely when shadow mode is off.
let shadow_ctx = if g9i.shadow_enabled() {
ShadowRequestContext::from_body(&body, context.is_stream)
} else {
None
};
let (plan, path, probe_new_work) = g9i
.classify_prefill(&self.worker_registry, &self.client, &body)
.await;
// T11: v1 rejection FIRST — a request rejected by any gate must
// never enter the shadow bookkeeping (拒绝路径零记账).
let v1_override = match plan.into_override() {
Ok(o) => o,
Err(info) => return Some(Err(info)),
};
let shadow_ctx = match shadow_ctx {
Some(mut ctx) => {
// T11: classify + candidate pick + cap verdict + admission
// in ONE queue-model critical section (atomic
// check-and-act) — the check→act window no longer spans
// the TTFT wait. Err = p_queue_cap (enforce), nothing
// admitted.
match g9i.shadow_classify_and_admit(&mut ctx) {
Ok(Some(mut decision)) => {
// T25-A: time already parked in the reservoir counts
// into TTFT — the re-judged estimate carries it.
if prior_wait_secs > 0.0 {
decision.est_ttft_secs += prior_wait_secs;
ctx.decision = Some(decision.clone());
}
super::g9i_shadow::record_agreement(
path,
probe_new_work,
&ctx,
&decision,
);
}
Ok(None) => {}
Err(info) => return Some(Err(info)),
}
Some(ctx)
}
None => None,
};
// T7 shadow-consistent routing (--g9i-shadow-route): the live
// prefill override follows the shadow decision's matched_node (=
// the booked node), so est vs actual is pure estimation error.
let override_url = g9i.shadow_route_pick(
v1_override,
shadow_ctx.as_ref().and_then(|c| c.decision.as_ref()),
);
Some(Ok((override_url, path, shadow_ctx)))
}
async fn execute_dual_dispatch<T: Serialize + Clone>(
&self,
headers: Option<&HeaderMap>,
@@ -390,6 +488,138 @@ impl PDRouter {
// Clone request once outside the retry loop, then use Arc to share across attempts
// This avoids O(retries) clones by sharing the same data
let shared_request = Arc::new(original_request.clone());
// G9i admission (router-side, zero engine changes). Computed once per
// request, outside the retry loop: our own 429s must never be
// retried. `ttft_path` labels the TTFT histogram dimension for this
// request's eventual first-token sample. `shadow_ctx` is the G9i v2
// shadow bookkeeping carrier (T2): Some only for chat requests with
// a classifiable prefix chain while shadow mode is on; it never
// influences routing.
//
// T15: holdable rejections (d_window_full / p_queue_cap) park the
// request in the FIFO hold queue instead of an immediate 429
// (429 太伤吞吐 — 测试端把 429 当失败). On pass the COMPLETE
// admission re-runs (every gate re-checked; the T11 booking happens
// only on a passing attempt — a parked request occupies no ledger).
// T20 (B): prompt_too_long hard validation — the FRONT-MOST check,
// before any admission gate / est_TTFT / ledger / queue bookkeeping
// and before the hold loop (an invalid request must never park).
// 400 (client bug), counted in smg_g9i_invalid_request_total, never
// in smg_g9i_rejects_total. Chat-completions only.
if context.route == "/v1/chat/completions" {
if let Some(g9i) = &self.g9i {
if let Ok(body) = serde_json::to_value(original_request) {
if let Some(resp) = g9i.prompt_too_long_rejection(&body) {
return resp;
}
}
}
}
let mut hold_deadline: Option<Instant> = None;
// T25-A reservoir park state: (deadline, seq id) across re-runs +
// the accumulated wait (counts into the est_TTFT re-judgment).
let mut reservoir_state: Option<(Instant, u64)> = None;
let mut reservoir_waited_secs: f64 = 0.0;
let (g9i_prefill_override, g9i_ttft_path, shadow_ctx): (
Option<String>,
&'static str,
Option<ShadowRequestContext>,
) = loop {
match self
.g9i_admission(&context, original_request, reservoir_waited_secs)
.await
{
None => {
// T25: per-request admission outcome (untracked = no
// G9i configured — class "none").
metrics::counter!(
super::g9i::METRIC_ADMISSION_TOTAL,
"class" => "none",
"result" => "admitted",
"reason" => "none",
)
.increment(1);
break (None, "untracked", None);
}
Some(Ok(outcome)) => {
// T25: per-request admission outcome — the shadow
// decision's class (the experiment's primary counter).
let class = outcome
.2
.as_ref()
.and_then(|c| c.decision.as_ref())
.map(|d| d.class.as_str())
.unwrap_or("none");
metrics::counter!(
super::g9i::METRIC_ADMISSION_TOTAL,
"class" => class,
"result" => "admitted",
"reason" => "none",
)
.increment(1);
break outcome;
}
Some(Err(mut info)) => {
let mut retry = false;
if let Some(g9i) = &self.g9i {
// T25-A: under hot_policy=reservoir, HOT-class
// holdable rejections park in the priority
// reservoir INSTEAD of the T15 hold (the A/B stays
// clean — every other class keeps the hold path).
if g9i.reservoir_eligible(&info) {
match g9i
.reservoir_or_reject(&info, &mut reservoir_state)
.await
{
Ok(waited) => {
reservoir_waited_secs += waited;
retry = true;
}
Err(final_info) => info = final_info,
}
} else if info.holdable() {
let body_bytes = context
.request_text
.as_deref()
.map(|s| s.len() as u64)
.unwrap_or(0);
match g9i
.hold_or_reject(&info, body_bytes, &mut hold_deadline)
.await
{
Ok(()) => retry = true,
Err(final_info) => info = final_info,
}
}
}
if retry {
continue;
}
// T25: per-request rejection outcome (class from the
// rejecting gate; "none" on non-chat/early gates).
metrics::counter!(
super::g9i::METRIC_ADMISSION_TOTAL,
"class" => info.hold_class.unwrap_or("none"),
"result" => "rejected",
"reason" => info.reason,
)
.increment(1);
metrics::counter!("smg_g9i_rejects_total", "reason" => info.reason)
.increment(1);
return super::g9i::reject_response(&info);
}
}
};
// T11: handle of the decision-time booking, extracted before the
// retry closure moves `shadow_ctx`. If the dispatch never reached
// the response path (all attempts failed before headers), the
// booking is released after the loop — exactly once, paired with
// shadow_admit's consume mark.
let admission_mark = shadow_ctx.as_ref().and_then(|c| c.admission.clone());
let response = RetryExecutor::execute_response_with_retry(
&self.retry_config,
{
@@ -397,6 +627,8 @@ impl PDRouter {
// Clone Arc (cheap reference count increment) instead of cloning the entire request
let shared_request = Arc::clone(&shared_request);
let context = context.clone();
let g9i_prefill_override = g9i_prefill_override.clone();
let shadow_ctx = shadow_ctx.clone();
async move {
let (prefill, decode) = match self
.select_pd_pair(
@@ -412,6 +644,38 @@ impl PDRouter {
}
};
// G9i least-wait override: route onto the classified
// fast-P node instead of the policy pick.
let (prefill, decode) =
if let Some(override_url) = g9i_prefill_override.as_ref() {
let matched = self
.worker_registry
.get_prefill_workers()
.into_iter()
.find(|w| {
w.url().trim_end_matches('/') == override_url.as_str()
&& w.is_available()
});
match matched {
Some(w) => {
debug!(
"G9i: overriding prefill selection with {}",
override_url
);
(w, decode)
}
None => {
warn!(
"G9i: override target {} unavailable, keeping policy pick",
override_url
);
(prefill, decode)
}
}
} else {
(prefill, decode)
};
debug!(
"PD retry attempt {} using prefill={} decode={}",
attempt,
@@ -428,6 +692,15 @@ impl PDRouter {
json_request["stream"] = Value::Bool(context.is_stream);
}
// G9i shadow T2: inject the usage / cache-details
// flags the client didn't set (chat + shadow mode
// only; `shadow_ctx` is None otherwise). Paired with
// response-side stripping: we inject exactly what we
// later strip, so clients see no change.
if let Some(sctx) = &shadow_ctx {
sctx.inject(&mut json_request);
}
json_request = match Self::inject_bootstrap_into_value(
json_request,
prefill.as_ref(),
@@ -446,6 +719,8 @@ impl PDRouter {
Arc::clone(&prefill),
Arc::clone(&decode),
start_time,
g9i_ttft_path,
shadow_ctx,
)
.await;
@@ -524,6 +799,13 @@ impl PDRouter {
);
}
// T11: release a decision-time booking that no response path ever
// consumed (all attempts failed before headers). No-op when the
// stream/non-stream path already took it over.
if let Some(g9i) = &self.g9i {
g9i.shadow_release_unconsumed(admission_mark);
}
response
}
@@ -575,6 +857,9 @@ impl PDRouter {
Some(response_headers),
prefill,
decode,
None,
// Error envelope: no usage will ever arrive — no shadow hook.
None,
)
} else {
// Handle non-streaming error response
@@ -649,6 +934,7 @@ impl PDRouter {
}
// Internal method that performs the actual dual dispatch (without retry logic)
#[allow(clippy::too_many_arguments)]
async fn execute_dual_dispatch_internal(
&self,
headers: Option<&HeaderMap>,
@@ -656,7 +942,11 @@ impl PDRouter {
context: PDRequestContext<'_>,
prefill: Arc<dyn Worker>,
decode: Arc<dyn Worker>,
_start_time: Instant,
start_time: Instant,
g9i_path: &'static str,
// G9i shadow T2: prefix chain + injection flags; None = shadow off
// or unclassifiable request — zero behavior change.
shadow_ctx: Option<ShadowRequestContext>,
) -> Response {
// For non-streaming: use guard for automatic load management
// For streaming: load will be managed in create_streaming_response
@@ -868,6 +1158,24 @@ impl PDRouter {
let response_headers = header_utils::preserve_response_headers(res.headers());
// Shadow T2: hand the chain + the ACTUAL prefill landing
// node (post-override) to the stream wrapper, which
// harvests the usage event and strips injected fields.
let shadow = shadow_ctx.map(|ctx| {
ShadowResponseContext::from_request(
ctx,
super::g9i::normalize_url(prefill.url()),
super::g9i::normalize_url(decode.url()),
)
});
// Shadow T3: admit into the queue model keyed by the
// ACTUAL landing node; the stream wrapper pairs this with
// exactly one complete (usage recorded) or release
// (abort/disconnect/error) before teardown.
if let (Some(g9i), Some(s)) = (&self.g9i, &shadow) {
g9i.shadow_admit(s);
}
self.create_streaming_response(
res.bytes_stream(),
status,
@@ -876,17 +1184,41 @@ impl PDRouter {
Some(response_headers),
prefill,
decode,
Some((start_time, g9i_path)),
shadow,
)
} else {
// Non-streaming response
if context.return_logprob {
self.process_non_streaming_response(
res,
status,
context.return_logprob,
prefill_body,
let shadow = shadow_ctx.map(|ctx| {
ShadowResponseContext::from_request(
ctx,
super::g9i::normalize_url(prefill.url()),
super::g9i::normalize_url(decode.url()),
)
.await
});
// Shadow T3: admit into the queue model keyed by the
// ACTUAL landing node; paired with complete/release in
// shadow_record_non_stream (or the read-failure paths
// below / in process_non_streaming_response).
if let (Some(g9i), Some(s)) = (&self.g9i, &shadow) {
g9i.shadow_admit(s);
}
if context.return_logprob {
let resp = self
.process_non_streaming_response(
res,
status,
context.return_logprob,
prefill_body,
shadow,
)
.await;
// T17: non-streaming e2e (observability only — the
// AIMD blind state reads local signals instead).
if let Some(g9i) = &self.g9i {
g9i.record_nonstream_e2e(start_time.elapsed().as_secs_f64());
}
resp
} else {
// Direct passthrough when no logprobs needed
let response_headers =
@@ -894,13 +1226,26 @@ impl PDRouter {
match res.bytes().await {
Ok(decode_body) => {
let mut response = Response::new(Body::from(decode_body));
*response.status_mut() = status;
*response.headers_mut() = response_headers;
let decode_body =
self.shadow_record_non_stream(decode_body, status, &shadow);
let response = Self::non_stream_passthrough_response(
status,
decode_body,
response_headers,
);
// T17: non-streaming e2e (observability only).
if let Some(g9i) = &self.g9i {
g9i.record_nonstream_e2e(start_time.elapsed().as_secs_f64());
}
response
}
Err(e) => {
error!("Failed to read decode response: {}", e);
// Shadow T3: the admitted dispatch produced no
// readable body — release the queue slot.
if let (Some(g9i), Some(s)) = (&self.g9i, &shadow) {
g9i.shadow_release(s);
}
error::internal_error(
"read_response_failed",
"Failed to read response",
@@ -1116,11 +1461,39 @@ impl PDRouter {
headers: Option<HeaderMap>,
prefill: Arc<dyn Worker>,
decode: Arc<dyn Worker>,
// TTFT sampling hook: Some((dispatch-entry instant, path label)) for
// real upstream streams; None for synthetic error envelopes (their
// "first chunk" is not a token — sampling them would poison the
// window with near-zero error samples) and for tests.
ttft: Option<(Instant, &'static str)>,
// G9i shadow T2 hook: Some only for shadowable chat requests on a
// success-status stream; None on error envelopes and tests.
shadow: Option<ShadowResponseContext>,
) -> Response {
use crate::core::AttachedBody;
let (tx, rx) = tokio::sync::mpsc::unbounded_channel();
// G9i TTFT sample: record on the first upstream chunk — that moment is
// the request's time-to-first-token as seen by the gateway (dispatch
// entry -> first token). Non-streaming requests have no observable
// first-token moment and are not sampled.
let g9i_ttft = self.g9i.clone().zip(ttft);
let mut ttft_recorded = false;
// G9i shadow T2: zip the state handle with the per-request context
// (chain + node + strip flags) for the stream loop below.
let g9i_shadow = self.g9i.clone().zip(shadow);
// Shadow T3 queue bookkeeping state. `ttft` is Option<(Instant,
// &'static str)> which is Copy, so zip() above left it usable:
// mirror the dispatch-entry instant so the first chunk can also feed
// the shadow TTFT calibration. `shadow_done` latches once usage is
// recorded; teardown releases the admission only when no completion
// happened (abort / client disconnect / upstream error).
let ttft_start = ttft.map(|(t, _)| t);
let mut shadow_ttft_secs: Option<f64> = None;
let mut shadow_done = false;
// Uses select! to race stream.next() against tx.closed() so that
// when the client disconnects the upstream HTTP connection is dropped
// promptly, allowing the engine to abort the request.
@@ -1149,8 +1522,81 @@ impl PDRouter {
chunk_result = tracked.next() => {
match chunk_result {
Some(Ok(chunk)) => {
// G9i shadow T2: harvest the usage event for
// prefix-table bookkeeping and strip exactly
// the fields we injected on the way in.
// Chunks without a usage field pass through
// byte-identically. A chunk reduced to nothing
// (usage-only event we must hide from the
// client) is dropped entirely.
let chunk = if let Some((g9i, sctx)) = &g9i_shadow {
let (usage, forward) =
super::g9i_shadow::process_sse_chunk(
&chunk,
sctx.strip_usage,
sctx.strip_cached_details,
);
if let Some(json) = usage {
// Shadow T3: exactly one completion
// per admission — the latch keeps a
// (malformed) repeat usage event from
// double-completing the queue slot.
if !shadow_done {
// Streaming harvests the usage
// event pre-strip, so the
// breakdown extraction inside
// works unaided (None).
g9i.shadow_record(sctx, &json, shadow_ttft_secs, None);
shadow_done = true;
}
}
match forward {
Some(bytes) => bytes,
None => continue,
}
} else {
chunk
};
let is_done = memmem::find(&chunk, b"data: [DONE]").is_some();
if !ttft_recorded {
ttft_recorded = true;
if let Some((g9i, (start_time, g9i_path))) = &g9i_ttft {
g9i.record_ttft(
g9i_path,
start_time.elapsed().as_secs_f64(),
);
// T25: TTFT by shadow class — the
// same gateway-observed first-token
// moment (start_time predates the
// admission loop, so reservoir /
// hold waits are included).
if let Some((_, sctx)) = &g9i_shadow {
if let Some(class) =
sctx.class.map(|c| c.as_str())
{
metrics::histogram!(
super::g9i::METRIC_CLASS_TTFT,
"class" => class,
)
.record(start_time.elapsed().as_secs_f64());
}
}
}
// Shadow T3: the same first-token moment
// calibrates est_TTFT (est − actual).
shadow_ttft_secs =
ttft_start.map(|t| t.elapsed().as_secs_f64());
// T16: first token -> the D booking moves
// from the pipeline to the streaming
// sub-ledger (the window gates pipeline
// depth only).
if let Some((g9i, sctx)) = &g9i_shadow {
g9i.shadow_promote_streaming(sctx);
}
}
let result = if return_logprob && prefill_logprobs.is_some() {
Self::merge_streaming_logprobs(prefill_logprobs.clone(), &chunk)
.unwrap_or(chunk)
@@ -1199,6 +1645,15 @@ impl PDRouter {
}
}
}
// Shadow T3: the stream ended without a recorded usage event
// (abort, client disconnect, or upstream error) — release the
// queue admission so outstanding stays conserved. No-op when the
// usage event already completed it.
if !shadow_done {
if let Some((g9i, sctx)) = &g9i_shadow {
g9i.shadow_release(sctx);
}
}
});
let stream = UnboundedReceiverStream::new(rx);
@@ -1226,16 +1681,26 @@ impl PDRouter {
status: StatusCode,
return_logprob: bool,
prefill_body: Option<bytes::Bytes>,
shadow: Option<ShadowResponseContext>,
) -> Response {
let response = res.bytes().await;
let decode_body = match response {
Ok(decode_body) => decode_body,
Err(e) => {
error!("Failed to read decode response: {}", e);
// Shadow T3: the admitted dispatch produced no readable
// body — release the queue slot.
if let (Some(g9i), Some(s)) = (&self.g9i, &shadow) {
g9i.shadow_release(s);
}
return error::internal_error("read_response_failed", "Failed to read response");
}
};
// Shadow T2: record usage + strip injected fields before any
// logprob merging (merge touches choices only, usage is untouched).
let decode_body = self.shadow_record_non_stream(decode_body, status, &shadow);
if !return_logprob {
return (status, decode_body).into_response();
}
@@ -1265,6 +1730,71 @@ impl PDRouter {
}
}
/// G9i shadow bookkeeping for non-streaming responses: parse the body,
/// record usage into the prefix table (T2), pair the T3 queue admission
/// with exactly one complete (parseable body) or release (non-success /
/// unparseable), and strip `cached_tokens_details` when (and only when)
/// we injected the request flag ourselves.
///
/// Passthrough guarantees: shadow off / non-success status / unparseable
/// body / nothing stripped -> the ORIGINAL bytes are returned untouched,
/// so clients never see a re-serialized response unless we actually
/// removed an injected field.
fn shadow_record_non_stream(
&self,
body: bytes::Bytes,
status: StatusCode,
shadow: &Option<ShadowResponseContext>,
) -> bytes::Bytes {
let (Some(g9i), Some(ctx)) = (&self.g9i, shadow) else {
return body;
};
// Only 200 responses with usage may write the table (design §3.3-3).
// A non-success response still releases the T3 queue admission.
if !status.is_success() {
g9i.shadow_release(ctx);
return body;
}
let (parsed, rewritten, breakdown) =
super::g9i_shadow::strip_and_extract(&body, ctx.strip_cached_details);
match &parsed {
// TTFT calibration is streaming-only (non-streaming has no
// observable first-token moment) -> None. The breakdown was
// extracted pre-strip and is passed through — the parsed JSON
// here is post-strip, so re-extracting from it would always
// miss (the 2026-10-17 order bug).
Some(json) => g9i.shadow_record(ctx, json, None, breakdown),
// Unparseable success body: nothing to record, but the dispatch
// is over — release rather than leak the outstanding count.
None => g9i.shadow_release(ctx),
}
match rewritten {
Some(bytes) => bytes::Bytes::from(bytes),
None => body,
}
}
/// Assemble the non-streaming passthrough response. The shadow strip
/// may have REWRITTEN (shortened) the body, so the upstream
/// Content-Length in `headers` is stale; forwarding it verbatim
/// truncated cache-hit responses at exactly the stripped length (the
/// 2026-10-17 IncompleteRead incident). Pin Content-Length to the
/// actual body length — correct for rewritten and untouched bodies
/// alike.
fn non_stream_passthrough_response(
status: StatusCode,
body: bytes::Bytes,
mut headers: HeaderMap,
) -> Response {
if let Ok(v) = HeaderValue::from_str(body.len().to_string().as_str()) {
headers.insert(CONTENT_LENGTH, v);
}
let mut response = Response::new(Body::from(body));
*response.status_mut() = status;
*response.headers_mut() = headers;
response
}
// Helper to process prefill response and extract body if needed for logprobs
async fn process_prefill_response(
&self,
@@ -1780,6 +2310,7 @@ mod tests {
retry_config: RetryConfig::default(),
api_key: Some("test_api_key".to_string()),
enable_igw: false,
g9i: None,
}
}
@@ -1791,6 +2322,39 @@ mod tests {
Box::new(worker)
}
#[test]
fn non_stream_passthrough_pins_content_length_to_body() {
// 2026-10-17 IncompleteRead regression: the shadow strip shortens
// the body on engine cache hits; the upstream Content-Length must
// never survive onto the rewritten body.
let mut headers = HeaderMap::new();
headers.insert(CONTENT_LENGTH, HeaderValue::from_static("637"));
let body = bytes::Bytes::from_static(b"{}");
let resp =
PDRouter::non_stream_passthrough_response(StatusCode::OK, body.clone(), headers);
assert_eq!(
resp.headers().get(CONTENT_LENGTH).unwrap(),
&HeaderValue::from_static("2"),
"stale upstream length is replaced by the actual body length"
);
// Untouched body: the pinned length matches the original.
let mut headers = HeaderMap::new();
headers.insert(CONTENT_LENGTH, HeaderValue::from_static("2"));
headers.insert(CONTENT_TYPE, HeaderValue::from_static("application/json"));
let resp = PDRouter::non_stream_passthrough_response(StatusCode::OK, body, headers);
assert_eq!(
resp.headers().get(CONTENT_LENGTH).unwrap(),
&HeaderValue::from_static("2")
);
assert_eq!(
resp.headers().get(CONTENT_TYPE).unwrap(),
&HeaderValue::from_static("application/json"),
"other preserved headers survive"
);
assert_eq!(resp.status(), StatusCode::OK);
}
#[test]
fn test_chat_request_text_uses_full_conversation() {
// Regression test for https://github.com/sgl-project/sglang/issues/26263
@@ -2048,6 +2612,8 @@ mod tests {
None,
prefill_ref.clone(),
decode_ref.clone(),
None,
None,
);
// Guards are now attached to response body, so load should be 1
@@ -321,6 +321,7 @@ mod pd_responses_routing_tests {
retry_config: RetryConfig::default(),
api_key: None,
enable_igw: false,
g9i: None,
}
}
@@ -0,0 +1,122 @@
"""Benchmark the fused ratio-2 decode compressor (``c2_decode_norm_rope_store``).
One CTA per token: pair-pool against the state ring, RMSNorm, RoPE, quantize and
store one compressed KV row. The line axis is the cache page format.
Shape axes:
- ``bs``: requests per batch.
- ``draft_len``: 1 is plain decode; > 1 is a target-verify block of that many
consecutive positions per request.
- ``parity``: the first position's parity. On decode, ``odd`` rows complete a
pair and run the full norm/RoPE/store path, ``even`` rows only park in the ring.
Constants follow the DeepSeek-V4 deployment: ``head_dim = 512``, ``rope_dim = 64``,
a 256-token FULL page, hence 128 compressed slots per page.
"""
import torch
from sglang.kernels.jit.benchmark import marker
from sglang.kernels.jit.benchmark.utils import (
DEFAULT_DEVICE,
create_empty,
create_random,
)
from sglang.kernels.ops.attention.dsv4.kv_layout import KVLayout
from sglang.kernels.ops.attention.dsv4.low_ratio_compress import (
c2_decode_norm_rope_store,
)
from sglang.srt.mem_cache.deepseek_v4_memory_pool import get_compress_state_ring_size
from sglang.test.ci.ci_register import register_cuda_ci
register_cuda_ci(
est_time=30, stage="base-b-kernel-benchmark", runner_config="1-gpu-large"
)
HEAD_DIM = 512
ROPE_DIM = 64
PAGE_SIZE = 256 // 2
MAX_POS = 65536
EPS = 1e-6
@marker.parametrize("parity", ["odd", "even", "half"], ["half"])
@marker.parametrize("draft_len", [1, 2, 4, 8], [1, 8])
@marker.parametrize("bs", marker.range(14, pattern="pow2"), [1, 1024])
@marker.benchmark("layout", ["V4", "V41", "V41_FP4"])
def benchmark(bs: int, draft_len: int, parity: str, layout: str):
if parity == "half" and bs % 2 != 0:
marker.skip("half parity requires even batch size")
layout = KVLayout[layout]
num_tokens = bs * draft_len
ring_size = get_compress_state_ring_size(2, draft_len if draft_len > 1 else 0)
torch.manual_seed(0)
kv_input = create_random(num_tokens, 2 * HEAD_DIM, dtype=torch.float32)
kv_state = create_random(bs * ring_size, 2 * HEAD_DIM, dtype=torch.float32)
norm_weight = create_random(HEAD_DIM)
freqs_cis = create_random(MAX_POS, ROPE_DIM, dtype=torch.float32)
req_ids = torch.randperm(bs, device=DEFAULT_DEVICE)
start = torch.randint(1, MAX_POS // 2 - draft_len, (bs,), device=DEFAULT_DEVICE)
if parity == "odd":
start = start * 2 + 1
elif parity == "even":
start = start * 2
else:
start = start * 2
start[::2] += 1 # half odd, half even
offsets = torch.arange(draft_len, device=DEFAULT_DEVICE)
positions = (start[:, None] + offsets).flatten()
req = req_ids.repeat_interleave(draft_len)
# Distinct scattered slots; slot 0 is left alone, it marks a padded row.
num_slots = num_tokens + 1
num_pages = -(-num_slots // PAGE_SIZE)
slots = torch.randperm(num_pages * PAGE_SIZE - 1, device=DEFAULT_DEVICE)
slots = slots[:num_tokens]
raw_out_loc = (slots + 1) * 2 + 1
k_cache = create_empty(
num_pages,
layout.page_bytes(PAGE_SIZE),
dtype=torch.uint8,
)
out = create_empty(num_tokens, HEAD_DIM)
# Traffic of the rows that complete a pair; parked rows write the state instead.
num_complete = int((positions % 2 == 1).sum())
num_parked = num_tokens - num_complete
row_bytes = 2 * HEAD_DIM * 4
complete_bytes = (
2 * row_bytes # kv_input + partner
+ ROPE_DIM * 4 # freqs_cis row
+ HEAD_DIM * 2 # out
+ layout.bytes_per_token
)
footprint = (
num_complete * complete_bytes
+ num_parked * 3 * row_bytes # kv_input + partner read, state write
+ num_tokens * 3 * 8 # positions, req, raw_out_loc
+ norm_weight.nbytes
)
return marker.do_bench(
c2_decode_norm_rope_store,
input_args=(kv_input, kv_state, norm_weight, positions, req, raw_out_loc),
input_kwargs=dict(
eps=EPS,
freqs_cis=freqs_cis,
k_cache=k_cache,
page_size=PAGE_SIZE,
ring_size=ring_size,
draft_len=draft_len,
layout=layout,
out=out,
),
memory_args=None,
memory_output=None,
extra_memory_footprint=footprint,
)
if __name__ == "__main__":
benchmark.run()
@@ -0,0 +1,191 @@
"""The ratio-2 compressor's eager-extend path against its decode path.
``c2_prefill_norm_rope_store`` pairs rows inside a chunk and parks each request's
last ``ring_size`` rows with a separate write-back kernel, while
``c2_decode_norm_rope_store`` takes every partner from the ring and parks every
pending row there. Driving one token stream both ways must leave the same main-KV
bytes and the same latents, so a chunk boundary that parks the wrong rows shows up
in the next chunk's output instead of staying invisible in the first chunk's.
"""
import unittest
from typing import List, Tuple
import torch
from sglang.kernels.ops.attention.dsv4.kv_layout import KVLayout
from sglang.kernels.ops.attention.dsv4.low_ratio_compress import (
c2_decode_norm_rope_store,
c2_prefill_norm_rope_store,
)
from sglang.srt.utils import is_cuda
from sglang.test.ci.ci_register import register_cuda_ci
from sglang.test.test_utils import CustomTestCase
# The c2 JIT kernels are the SM100 build; the only Blackwell runner config is the
# four-GPU one.
register_cuda_ci(est_time=60, stage="base-b-kernel-unit", runner_config="4-gpu-b200")
HEAD_DIM = 512
ROPE_DIM = 64
PAGE_SIZE = 64
NUM_PAGES = 16
EPS = 1e-6
# Request r's full slots start here, even so that a (2k, 2k+1) pair shares a
# compressed slot; never 0, which marks a padded row.
SLOT_BASE = 64
def _sm100() -> bool:
return is_cuda() and torch.cuda.get_device_capability()[0] == 10
class Stream:
"""One token stream: per-request prefix and chunk sizes, laid out as the rows
the backend would hand each call."""
def __init__(self, prefixes: List[int], chunks: List[List[int]]):
self.prefixes = prefixes
self.chunks = chunks # chunks[c][r] = rows of request r in chunk c
def chunk_rows(self, c: int) -> List[Tuple[int, int]]:
"""(request, position) of every row of chunk ``c``, requests in order."""
rows = []
for r, n in enumerate(self.chunks[c]):
start = self.prefixes[r] + sum(self.chunks[j][r] for j in range(c))
rows.extend((r, start + i) for i in range(n))
return rows
def _make_inputs(gen, rows: int) -> torch.Tensor:
"""``[rows, 2 * HEAD_DIM]`` fp32: the kv half then the score half."""
return torch.randn(
rows, 2 * HEAD_DIM, generator=gen, device="cuda", dtype=torch.float32
)
def _row_tensors(rows: List[Tuple[int, int]], pad: int):
"""``req``, ``positions`` and ``raw_out_loc`` for ``rows``, plus ``pad`` padded
rows. A padded row still has to address a live ring slot, so it takes request 0
and position 0; only ``raw_out_loc == 0`` marks it."""
req = [r for r, _ in rows] + [0] * pad
pos = [p for _, p in rows] + [0] * pad
loc = [SLOT_BASE * (1 + r) + p for r, p in rows] + [0] * pad
to = lambda v, d: torch.tensor(v, device="cuda", dtype=d)
return to(req, torch.int64), to(pos, torch.int32), to(loc, torch.int32)
class TestC2PrefillRing(CustomTestCase):
def _run(self, stream: Stream, layout: KVLayout, ring_size: int, seed: int):
"""The fused chunked path and the row-at-a-time decode path over the same
stream; returns (cache, latents) for each."""
gen = torch.Generator(device="cuda").manual_seed(seed)
num_reqs = len(stream.prefixes)
weight = torch.randn(
HEAD_DIM, generator=gen, device="cuda", dtype=torch.float32
).to(torch.bfloat16)
angles = torch.randn(4096, ROPE_DIM // 2, generator=gen, device="cuda")
freqs = torch.view_as_real(
torch.polar(torch.ones_like(angles), angles)
).flatten(-2)
# One fp32 input row per (request, position) in the stream, shared by both
# paths so any difference is the pairing, not the data.
all_rows = [
row for c in range(len(stream.chunks)) for row in stream.chunk_rows(c)
]
inputs = {row: _make_inputs(gen, 1)[0] for row in all_rows}
state_rows = num_reqs * ring_size + ring_size
caches, latents = [], []
for fused in (True, False):
cache = torch.zeros(
NUM_PAGES,
layout.page_bytes(PAGE_SIZE),
dtype=torch.uint8,
device="cuda",
)
state = torch.zeros(
state_rows, 2 * HEAD_DIM, dtype=torch.float32, device="cuda"
)
out = {}
if fused:
for c in range(len(stream.chunks)):
rows = stream.chunk_rows(c)
pad = 2
req, pos, loc = _row_tensors(rows, pad)
kv = torch.stack(
[inputs[row] for row in rows]
+ [torch.zeros(2 * HEAD_DIM, device="cuda")] * pad
)
starts, lens = [], []
at = 0
for r in range(num_reqs):
starts.append(at)
lens.append(stream.chunks[c][r])
at += stream.chunks[c][r]
latent = c2_prefill_norm_rope_store(
kv,
state,
weight,
pos,
req,
loc,
EPS,
freqs,
cache,
torch.tensor(starts, device="cuda", dtype=torch.int32),
torch.tensor(lens, device="cuda", dtype=torch.int32),
page_size=PAGE_SIZE,
ring_size=ring_size,
layout=layout,
)
for i, row in enumerate(rows):
out[row] = latent[i].clone()
else:
for row in all_rows:
req, pos, loc = _row_tensors([row], 0)
latent = c2_decode_norm_rope_store(
inputs[row][None, :],
state,
weight,
pos,
req,
loc,
EPS,
freqs,
cache,
page_size=PAGE_SIZE,
ring_size=ring_size,
layout=layout,
)
out[row] = latent[0].clone()
caches.append(cache)
# Only a completing (odd position) row writes a latent.
latents.append(
torch.stack([out[row] for row in all_rows if row[1] % 2 == 1])
)
return caches, latents
@unittest.skipUnless(_sm100(), "the c2 JIT kernels are the SM100 build")
def test_chunked_extend_matches_decode(self):
# Request 0 starts even, request 1 odd and spans more rows than the ring,
# request 2 odd with a single row in its first chunk.
stream = Stream(
prefixes=[0, 7, 1],
chunks=[[5, 4, 1], [3, 6, 2], [1, 1, 1]],
)
for layout in (KVLayout.V41, KVLayout.V41_FP4):
for ring_size in (2, 8):
with self.subTest(layout=layout.name, ring_size=ring_size):
caches, latents = self._run(stream, layout, ring_size, seed=7)
self.assertTrue(
torch.equal(caches[0], caches[1]),
f"main-KV bytes differ: {(caches[0] != caches[1]).sum()} of "
f"{caches[0].numel()}",
)
self.assertTrue(torch.equal(latents[0], latents[1]))
if __name__ == "__main__":
unittest.main()
@@ -10,6 +10,9 @@ Checks the Triton kernel against a per-row torch reference:
3. ``test_non_trailing_query_positions``: absolute ``query_pos`` that are not
the trailing extend tokens, with a cross-chunk ``query_start_loc`` offset.
4. ``test_swa_only_layer``: ``topk == 0`` writes only the window.
5. ``test_reused_buffers_are_overwritten_whole``: the kernel is handed
uninitialized and already-written buffers, so it must overwrite every row it
is given rather than rely on a ``-1`` fill pass.
"""
import pytest
@@ -80,12 +83,15 @@ def _check(**kwargs):
assert torch.equal(got_idx, ref_idx)
def _trailing_case(seq_lens, extend_lens, topk, compress_ratio, seed=0):
"""Rows are the trailing ``extend_lens[r]`` tokens of request ``r``."""
def _trailing_case(seq_lens, extend_lens, topk, compress_ratio, seed=0, pad_rows=0):
"""Rows are the trailing ``extend_lens[r]`` tokens of request ``r``, followed by
``pad_rows`` rows past the last request."""
gen = torch.Generator(device="cpu").manual_seed(seed)
query_pos = []
for seq_len, extend_len in zip(seq_lens, extend_lens):
query_pos.extend(range(seq_len - extend_len, seq_len))
# Past the last request, so the kernel never reads these positions.
query_pos.extend([0] * pad_rows)
num_tokens = len(query_pos)
topk_indices = torch.randint(
0, 1 << 20, (num_tokens, topk), generator=gen, dtype=torch.int32
@@ -163,6 +169,48 @@ def test_non_trailing_query_positions():
)
def test_reused_buffers_are_overwritten_whole():
"""Callers hand the kernel uninitialized or already-written buffers, so every
row it is given must be overwritten in full. A row that keeps part of an
earlier call's longer selection, or a row past the last request that keeps
anything at all, is a stale index attention would read as live."""
width = -(-(64 + WINDOW) // TOPK_ALIGNMENT) * TOPK_ALIGNMENT
# Long selections first: every row fills 64 + WINDOW entries.
wide = _trailing_case([512, 640], [8, 8], topk=64, compress_ratio=4)
num_tokens = wide["topk_indices"].shape[0]
out_indices = torch.empty((num_tokens, width), dtype=torch.int32, device=DEVICE)
out_lens = torch.empty(num_tokens, dtype=torch.int32, device=DEVICE)
# An uninitialized buffer is poison, not -1.
out_indices.fill_(1 << 30)
out_lens.fill_(1 << 30)
got_idx, got_lens = combine_topk_swa_indices(
**wide, out_indices=out_indices, out_lens=out_lens
)
assert got_idx.data_ptr() == out_indices.data_ptr()
ref_idx, ref_lens = _reference(**wide)
assert torch.equal(got_lens, ref_lens)
assert torch.equal(got_idx, ref_idx)
assert int(got_lens[0]) == 64 + WINDOW
# Same buffers, now holding the first call's output: shorter selections, a
# zero-length request, and rows past the last request.
narrow = _trailing_case(
[20, 36, 48], [2, 0, 3], topk=64, compress_ratio=4, pad_rows=num_tokens - 5
)
assert narrow["topk_indices"].shape[0] == num_tokens
got_idx, got_lens = combine_topk_swa_indices(
**narrow, out_indices=out_indices, out_lens=out_lens
)
ref_idx, ref_lens = _reference(**narrow)
assert torch.equal(got_lens, ref_lens), (got_lens.tolist(), ref_lens.tolist())
assert torch.equal(got_idx, ref_idx)
# The shorter rows and the padding rows must not keep the first call's tails.
assert int(got_lens[0]) < 64 + WINDOW
assert (got_lens[5:] == 0).all()
assert (got_idx[5:] == -1).all()
def test_swa_only_layer():
case = _trailing_case([200, 50], [2, 2], topk=0, compress_ratio=4)
case["topk_indices"] = torch.zeros((4, 1), dtype=torch.int32, device=DEVICE)
@@ -1,6 +1,7 @@
import struct
import threading
import unittest
from collections import defaultdict
from types import SimpleNamespace
from unittest.mock import ANY, Mock, call, patch
@@ -18,6 +19,8 @@ from sglang.srt.disaggregation.common.staging_handler import (
handle_staging_req,
)
from sglang.srt.disaggregation.common.utils import (
FastQueue,
TransferKVChunk,
group_concurrent_contiguous,
pack_int_lists,
pack_list_of_buffers,
@@ -68,6 +71,227 @@ register_cpu_ci(est_time=11, suite="base-a-test-cpu")
class TestDisaggregationWire(unittest.TestCase):
def test_mooncake_stride_mismatch_only_fails_affected_request(self):
for layer_ids in ([], [7]):
for outcome in ("mismatch", "cleared", "transport_failure"):
with self.subTest(layer_ids=layer_ids, outcome=outcome):
manager = object.__new__(MooncakeKVManager)
manager.kv_args = SimpleNamespace(
kv_data_ptrs=[],
state_types=[StateType.SWA, StateType.DSA],
state_data_ptrs=[[0x1000], [0x3000]],
state_item_lens=[[32], [16]],
state_dim_per_tensor=[[], []],
state_layer_ids=[layer_ids, layer_ids],
aux_data_ptrs=[0x5000],
aux_item_lens=[8],
)
manager.is_mla_backend = True
manager.is_hybrid_mla_backend = False
manager.attn_tp_size = manager.attn_cp_size = manager.pp_size = 1
manager.attn_tp_rank = manager.attn_cp_rank = manager.pp_rank = 0
manager.enable_trace = manager.enable_staging = False
manager.enable_custom_mem_pool = False
manager.max_transfer_batch_indices = 0
manager.bootstrap_port = 1234
manager.session_lock = threading.Lock()
manager.failure_lock = threading.Lock()
manager.failed_sessions = set()
manager.session_failures = defaultdict(int)
manager.failure_records = {}
manager._staging_outstanding = defaultdict(int)
manager.request_status = {
42: KVPoll.Transferring,
43: KVPoll.Transferring,
}
manager.req_to_decode_prefix_len = {42: 0, 43: 0}
manager.transfer_infos = {}
manager.decode_kv_args_table = {}
queue = FastQueue()
for room in (42, 43):
req = TransferInfo(
room=room,
endpoint="127.0.0.1",
dst_port=1234,
mooncake_session_id=f"decode:{room}",
dst_kv_indices=np.array([], dtype=np.int32),
dst_aux_index=5,
dst_state_indices=[[5], [5]],
required_dst_info_num=1,
is_dummy=False,
)
manager.transfer_infos[room] = {req.mooncake_session_id: req}
manager.decode_kv_args_table[req.mooncake_session_id] = (
SimpleNamespace(
requires_dcp_relayout=False,
dst_state_data_ptrs=[[0x2000], [0x4000]],
dst_state_item_lens=[
[
64
if room == 42 and outcome != "transport_failure"
else 32
],
[16],
],
dst_state_dim_per_tensor=[[], []],
dst_state_layer_ids=[layer_ids, layer_ids],
dst_attn_tp_size=1,
dst_aux_ptrs=[0x6000],
)
)
queue.put(
TransferKVChunk(
room=room,
prefill_kv_indices=np.array([], dtype=np.int32),
index_slice=slice(0, 0),
is_last_chunk=True,
prefill_aux_index=2,
state_indices=[[2], [2]],
)
)
queue.put(None)
def send_status(endpoint, parts, **kwargs):
if outcome == "cleared" and parts[0] == b"42":
manager.request_status.pop(42)
manager._send_multipart_locked = Mock(side_effect=send_status)
manager._transfer_data = Mock(
side_effect=lambda session, blocks: (
-1
if session == "decode:42" and outcome == "transport_failure"
else 0
)
)
with (
get_context().override_server_args(enable_unified_memory=False),
patch.object(
envs.SGLANG_MOONCAKE_SEND_AUX_TCP, "get", return_value=False
),
):
manager.transfer_worker(queue, executor=None)
self.assertEqual(
manager.request_status.get(42),
None if outcome == "cleared" else KVPoll.Failed,
)
self.assertEqual(manager.request_status[43], KVPoll.Success)
expected_reason = (
"Failed to send state components of 42 to 127.0.0.1:1234"
if outcome == "transport_failure"
else f"{StateType.SWA} item length mismatch for "
+ (
"paired entries src[0]=32 dst[0]=64"
if layer_ids
else "positional entry 0: prefill=32 decode=64"
)
)
self.assertEqual(manager.failure_records[42], expected_reason)
self.assertEqual(
[
c.args[1]
for c in manager._send_multipart_locked.call_args_list
],
[[b"42", b"0", b"0"], [b"43", b"4", b"0"]],
)
good_writes = [
call("decode:43", [(0x1040, 0x20A0, 32)]),
call("decode:43", [(0x3020, 0x4050, 16)]),
call("decode:43", [(0x5010, 0x6028, 8)]),
]
expected_writes = (
[
call("decode:42", [(0x1040, 0x20A0, 32)]),
call("decode:42", [(0x3020, 0x4050, 16)]),
]
if outcome == "transport_failure"
else []
) + good_writes
self.assertEqual(
manager._transfer_data.call_args_list, expected_writes
)
self.assertEqual(
manager.failed_sessions,
{"decode:42"} if outcome == "transport_failure" else set(),
)
self.assertEqual(
dict(manager.session_failures),
{"decode:42": 1} if outcome == "transport_failure" else {},
)
self.assertFalse(manager._staging_outstanding)
self.assertFalse(manager.transfer_infos)
self.assertFalse(manager.req_to_decode_prefix_len)
def test_mooncake_state_stride_matches_pp_slice(self):
cases = [
([], [64, 128], [32, 64, 128, 256], [1, 2]),
(
[4, 128, 4, 128],
[32, 64, 512, 2048],
[16, 32, 64, 128, 256, 512, 1024, 2048],
[1, 2, 5, 7],
),
]
for ratios, src_lens, dst_lens, dst_entries in cases:
with self.subTest(ratios=ratios):
manager = object.__new__(MooncakeKVManager)
manager.kv_args = SimpleNamespace(
kv_data_ptrs=[],
prefill_start_layer=1,
prefill_end_layer=3,
mla_compression_ratios=ratios,
)
manager.is_mla_backend = True
manager.pp_size = 2
manager.enable_custom_mem_pool = False
manager.max_transfer_batch_indices = 0
manager._transfer_data = Mock(return_value=0)
src_ptrs = [0x1000 * (i + 1) for i in range(len(src_lens))]
dst_ptrs = [0x1000 * (i + 9) for i in range(len(dst_lens))]
manager._send_kvcache_generic(
mooncake_session_id="decode",
executor=None,
src_data_ptrs=src_ptrs,
dst_data_ptrs=dst_ptrs,
item_lens=src_lens,
dst_item_lens=dst_lens,
prefill_data_indices=np.array([2], dtype=np.int32),
dst_data_indices=np.array([5], dtype=np.int32),
state_type=StateType.SWA,
)
manager._transfer_data.assert_called_once_with(
"decode",
[
(
src_ptrs[i] + 2 * length,
dst_ptrs[j] + 5 * length,
length,
)
for i, (j, length) in enumerate(zip(dst_entries, src_lens))
],
)
def test_sender_clear_keeps_abort_ack_until_writes_drain(self):
manager = object.__new__(MooncakeKVManager)
sender = object.__new__(MooncakeKVSender)
sender.kv_mgr, sender.bootstrap_room = manager, 42
for outstanding in (0, 1):
with self.subTest(outstanding=outstanding):
manager.request_status = {42: KVPoll.Failed}
manager._staging_outstanding = {42: outstanding}
manager._deferred_ack_targets = {42: ("127.0.0.1", 1234)}
manager._deferred_ack_fanout_snapshots = {}
manager.transfer_infos = {}
with patch.object(manager, "_send_abort_ack") as ack:
sender.clear()
if outstanding:
ack.assert_not_called()
manager._staging_outstanding[42] = 0
manager._maybe_ack_drained_abort(42)
ack.assert_called_once_with("127.0.0.1", 1234, 42)
manager._maybe_ack_drained_abort(42)
ack.assert_called_once()
def test_mooncake_registration_staging_fields(self):
msg = [
b"room",
@@ -401,6 +625,54 @@ class TestQwen4StateWire(unittest.TestCase):
[(0, 2), (1, 3)],
)
def _qsa_compressed_manager(self, *, src_item_len):
manager = object.__new__(MooncakeKVManager)
manager.request_status = {42: KVPoll.Transferring}
manager.attn_tp_size = 4
manager.pp_size = 1
manager.is_mla_backend = False
manager.is_hybrid_mla_backend = False
manager.kv_args = SimpleNamespace(
engine_rank=0,
state_types=[StateType.QSA_COMPRESSED],
state_data_ptrs=[[1000]],
state_item_lens=[[src_item_len]],
state_dim_per_tensor=[[]],
state_layer_ids=[[24]],
)
manager.sent = []
manager._send_kvcache_generic = lambda **kw: manager.sent.append(kw) or 0
return manager
def test_qsa_compressed_layout_mismatch_rejected_before_transfer(self):
"""Equal attention TP with fp8 on one peer only halves that peer's
compressed item length. The transfer uses the source length as the
destination stride, so the mismatch must be rejected before any write."""
req = SimpleNamespace(
room=42, mooncake_session_id="session", dst_state_indices=[[3]]
)
decode_info = SimpleNamespace(
dst_attn_tp_size=4,
dst_state_data_ptrs=[[2000]],
dst_state_item_lens=[[16]],
dst_state_dim_per_tensor=[[]],
dst_state_layer_ids=[[24]],
)
manager = self._qsa_compressed_manager(src_item_len=32)
with self.assertRaisesRegex(RuntimeError, "QSA_COMPRESSED layout differs"):
MooncakeKVManager.maybe_send_extra(
manager, req, [[3]], None, target_rank_registration_info=decode_info
)
self.assertEqual(manager.sent, [])
manager = self._qsa_compressed_manager(src_item_len=16)
rc = MooncakeKVManager.maybe_send_extra(
manager, req, [[3]], None, target_rank_registration_info=decode_info
)
self.assertEqual(rc, 0)
self.assertEqual(len(manager.sent), 1)
self.assertEqual(manager.sent[0]["item_lens"], [16])
def test_replicated_state_tp_policy(self):
for src_tp, dst_tp, rank, expected in (
(4, 1, 0, True),
@@ -1,14 +1,18 @@
import concurrent.futures
import ctypes
import unittest
from threading import Event
from collections import defaultdict
from queue import SimpleQueue
from threading import Event, Lock
from types import SimpleNamespace
from unittest.mock import MagicMock, call, patch
import numpy as np
from sglang.srt.disaggregation.base.conn import StateType
from sglang.srt.disaggregation.mooncake.conn import MooncakeKVManager
from sglang.srt.disaggregation.base.conn import KVPoll, StateType
from sglang.srt.disaggregation.common.utils import TransferKVChunk
from sglang.srt.disaggregation.mooncake.conn import MooncakeKVManager, MooncakeKVSender
from sglang.srt.disaggregation.utils import DisaggregationMode
from sglang.test.ci.ci_register import register_cpu_ci
from sglang.test.test_utils import CustomTestCase
@@ -407,5 +411,136 @@ class TestDcpPackLifetime(CustomTestCase):
self.assertEqual(observed, [11])
class TestMooncakeEarlySend(unittest.TestCase):
def test_sender_queues_event_once_without_waiting_on_scheduler(self):
for last in (False, True):
with self.subTest(last=last):
ready = MagicMock()
manager = MagicMock()
sender = SimpleNamespace(
kv_mgr=manager,
bootstrap_room=1,
aux_index=0,
trace_ctx=MagicMock(),
_early_send_wait_event=ready,
_prepare_send_indices=lambda ids, state: (
ids,
slice(0, 1),
last,
False,
),
_record_transfer_indices=MagicMock(),
)
MooncakeKVSender.send(sender, [0])
self.assertIs(
manager.add_transfer_request.call_args.kwargs["wait_event"], ready
)
ready.synchronize.assert_not_called()
MooncakeKVSender.send(sender, [1])
self.assertIsNone(
manager.add_transfer_request.call_args.kwargs["wait_event"]
)
def test_manager_preserves_event_in_queued_chunk(self):
queue = SimpleQueue()
ready = MagicMock()
manager = SimpleNamespace(
disaggregation_mode=DisaggregationMode.PREFILL,
request_status={1: KVPoll.WaitingForInput},
check_status=lambda room: KVPoll.WaitingForInput,
transfer_infos={1: {"127.0.0.1:80": object()}},
transfer_queues=[queue],
)
MooncakeKVManager.add_transfer_request(
manager, 1, [0], slice(0, 1), False, wait_event=ready
)
self.assertIs(queue.get_nowait().wait_event, ready)
ready.synchronize.assert_not_called()
def _worker(self, event):
chunk = TransferKVChunk(
1, [0], slice(0, 1), False, None, None, wait_event=event
)
queue = SimpleQueue()
queue.put(chunk)
queue.put(None)
peer = SimpleNamespace(
is_dummy=False,
mooncake_session_id="peer",
dst_device_kv_indices=None,
dst_kv_indices=[10],
)
registration = SimpleNamespace(
requires_dcp_relayout=False,
dst_kv_ptrs=[1000],
dst_kv_layer_ids=[0],
dst_kv_item_len=16,
dst_attn_tp_size=1,
)
manager = SimpleNamespace(
enable_trace=False,
enable_staging=False,
enable_deferred_decode_kv_release=True,
_staging_outstanding=defaultdict(int),
request_status={1: KVPoll.WaitingForInput},
transfer_infos={1: {"peer": peer}},
_prefill_unique_rank=lambda: 0,
session_lock=Lock(),
failed_sessions=set(),
# Read under session_lock alongside failed_sessions on every
# non-dummy req, and invoked by the worker's except path.
state_layout_rejections={},
poison_deferred_ack_room=MagicMock(),
decode_kv_args_table={"peer": registration},
_get_dsa_cache_transfer_skip_flags=lambda reg: (False, False),
kv_args=SimpleNamespace(kv_data_ptrs=[1]),
is_mla_backend=True,
send_kvcache=MagicMock(return_value=0),
_maybe_ack_drained_abort=MagicMock(),
bootstrap_port=8998,
)
manager.check_status = lambda room: manager.request_status[room]
return manager, queue
def test_worker_waits_before_reading_kv(self):
ready = MagicMock()
manager, queue = self._worker(ready)
def transfer(*args, **kwargs):
ready.synchronize.assert_called_once_with()
return 0
manager.send_kvcache.side_effect = transfer
MooncakeKVManager.transfer_worker(manager, queue, MagicMock())
manager.send_kvcache.assert_called_once()
def test_abort_during_wait_prevents_transfer_and_releases_ack(self):
ready = MagicMock()
manager, queue = self._worker(ready)
ready.synchronize.side_effect = lambda: manager.request_status.update(
{1: KVPoll.Failed}
)
MooncakeKVManager.transfer_worker(manager, queue, MagicMock())
ready.synchronize.assert_called_once_with()
manager.send_kvcache.assert_not_called()
manager._maybe_ack_drained_abort.assert_called_once_with(1)
self.assertNotIn(1, manager._staging_outstanding)
def test_failed_room_skips_wait(self):
ready = MagicMock()
manager, queue = self._worker(ready)
manager.request_status[1] = KVPoll.Failed
MooncakeKVManager.transfer_worker(manager, queue, MagicMock())
ready.synchronize.assert_not_called()
manager.send_kvcache.assert_not_called()
manager._maybe_ack_drained_abort.assert_called_once_with(1)
self.assertNotIn(1, manager._staging_outstanding)
def test_worker_preserves_no_event_path(self):
manager, queue = self._worker(None)
MooncakeKVManager.transfer_worker(manager, queue, MagicMock())
manager.send_kvcache.assert_called_once()
if __name__ == "__main__":
unittest.main()
@@ -219,6 +219,7 @@ class TestKPoolStreamScheduling(unittest.TestCase):
)
batch = SimpleNamespace(
forward_mode=ForwardMode.EXTEND,
seq_lens=torch.tensor([8192 + num_tokens]),
seq_lens_cpu=torch.tensor([8192 + num_tokens]),
)
prepare_qk = Mock(return_value=(x, x, None, None))
@@ -28,7 +28,7 @@ class _FakeReq:
"""Carries the fill-id state convert_decode_to_extend touches, with the
real Req methods so the array bookkeeping is not re-implemented here."""
_refresh_fill_ids = Req._refresh_fill_ids
refresh_fill_ids = Req.refresh_fill_ids
set_extend_range = Req.set_extend_range
def __init__(self, *, num_prompt_tokens: int, num_output_tokens: int):
@@ -132,7 +132,7 @@ class _FakeReq:
self.full_untruncated_fill_ids = list(range(origin_len + output_len))
self.extend_range = None
def _refresh_fill_ids(self):
def refresh_fill_ids(self):
self.full_untruncated_fill_ids = self.origin_input_ids + self.output_ids
@property
@@ -103,6 +103,19 @@ class MockReq:
def get_fill_ids(self):
return self.full_untruncated_fill_ids[: self.extend_range.end]
rid = "mock-req"
skip_radix_cache_insert = False
finished_reason = None
def finished(self):
return self.finished_reason is not None
def refresh_fill_ids(self):
pass
def owned_kv_len(self):
return self.kv.kv_committed_len
def _make_req(fill_ids, req_pool_idx=0, cache_protected_len=0, last_node=None):
return MockReq(fill_ids, req_pool_idx, cache_protected_len, last_node)
@@ -24,6 +24,21 @@ register_cpu_ci(est_time=5, suite="base-a-test-cpu")
class TestDSV4CompressedPools(CustomTestCase):
def test_swa_key_page_size_uses_physical_paged_size(self):
pool = DeepSeekV4TokenToKVPool.__new__(DeepSeekV4TokenToKVPool)
pool.request_window = None
pool.swa_page_size = 256
for physical_page_size in (256, 64):
with self.subTest(physical_page_size=physical_page_size):
pool.swa_kv_pool = SimpleNamespace(page_size=physical_page_size)
self.assertEqual(pool.get_swa_key_page_size(), physical_page_size)
def test_swa_key_page_size_without_paged_pool(self):
pool = DeepSeekV4TokenToKVPool.__new__(DeepSeekV4TokenToKVPool)
pool.swa_kv_pool = None
pool.request_window = SimpleNamespace(page_size=256)
self.assertEqual(pool.get_swa_key_page_size(), 256)
def test_physical_kv_pages_cover_reserved_logical_page(self):
size = 8192
self.assertEqual(_num_dsv4_physical_kv_pages(size, 256, 256), 33)
@@ -1,5 +1,6 @@
"""Exercise container and ancestor budgets using synthetic procfs/cgroup files."""
import functools
import tempfile
import unittest
from pathlib import Path
@@ -8,6 +9,8 @@ from unittest.mock import Mock, patch
from sglang.srt.mem_cache import host_memory
from sglang.test.ci.ci_register import register_cpu_ci
_cgroup_memory_headroom = host_memory._cgroup_memory_headroom
register_cpu_ci(est_time=1, suite="base-a-test-cpu")
@@ -31,7 +34,7 @@ class TestHostMemory(unittest.TestCase):
f"1 0 0:1 {mount_root} {escaped} rw - {filesystem} cgroup {options}\n"
)
def memory(self, path, usage, maximum="max", high="max", v1=False):
def memory(self, path, usage, maximum="max", high="max", v1=False, stat=None):
directory = self.mount / path
directory.mkdir(parents=True, exist_ok=True)
files = (
@@ -39,6 +42,11 @@ class TestHostMemory(unittest.TestCase):
if v1
else {"memory.max": maximum, "memory.high": high, "memory.current": usage}
)
prefix = "total_" if v1 else ""
stat = {f"{prefix}active_file": 0, f"{prefix}inactive_file": 0} | (stat or {})
files["memory.stat"] = "".join(
f"{key} {value}\n" for key, value in stat.items()
)
for name, value in files.items():
(directory / name).write_text(str(value))
@@ -74,6 +82,32 @@ class TestHostMemory(unittest.TestCase):
self.memory("task", 400, 1000, v1=True)
self.assertEqual(host_memory._cgroup_memory_headroom(self.proc), 600)
def test_reclaimable_page_cache_is_not_used(self):
"""Page cache charged to the cgroup, such as a checkpoint just read, is
reclaimed under the limit and must not shrink the headroom. Shared
memory is counted in file/cache but cannot be reclaimed."""
v2 = {"file": 700, "shmem": 200, "active_file": 200, "inactive_file": 300}
v1 = {
"cache": 700,
"shmem": 200,
"active_file": 0,
"inactive_file": 0,
"total_active_file": 200,
"total_inactive_file": 300,
}
for is_v1, stat, usage, expected in [
(False, v2, 900, 600),
(True, v1, 900, 600),
# v1 usage is approximate and may trail the cache counters.
(True, v1, 400, 1000),
]:
with self.subTest(v1=is_v1, usage=usage):
self.configure(v1=is_v1)
self.memory("task/engine", usage, 1000, v1=is_v1, stat=stat)
self.assertEqual(
host_memory._cgroup_memory_headroom(self.proc), expected
)
def test_independent_engines_have_separate_allowances(self):
# Both engines see the same host RAM but have different charged usage.
for task, usage, expected in [("a", 300, 700), ("b", 600, 400)]:
@@ -98,9 +132,14 @@ class TestHostMemory(unittest.TestCase):
)
def test_host_availability_is_also_a_bound(self):
for cgroup, expected in [(None, 100), (200, 100), (50, 50)]:
for cgroup, allow_fallback, expected in [
(None, False, 100),
(200, False, 100),
(50, False, 50),
(50, True, 50),
]:
with (
self.subTest(cgroup=cgroup),
self.subTest(cgroup=cgroup, allow_fallback=allow_fallback),
patch.object(
host_memory.psutil,
"virtual_memory",
@@ -110,20 +149,57 @@ class TestHostMemory(unittest.TestCase):
host_memory, "_cgroup_memory_headroom", return_value=cgroup
),
):
self.assertEqual(host_memory.available_host_memory_bytes(), expected)
self.assertEqual(
host_memory.available_host_memory_bytes(
allow_cgroup_fallback=allow_fallback
),
expected,
)
def test_unmounted_memory_cgroup_fails(self):
self.configure()
(self.proc / "self/mountinfo").write_text("")
with self.assertRaisesRegex(RuntimeError, "Cannot locate"):
host_memory._cgroup_memory_headroom(self.proc)
def available(self, host_available=5000, *, allow_cgroup_fallback=False):
"""Public sizing entry point over the synthetic procfs."""
with (
patch.object(
host_memory.psutil,
"virtual_memory",
return_value=Mock(available=host_available),
),
patch.object(
host_memory,
"_cgroup_memory_headroom",
functools.partial(_cgroup_memory_headroom, self.proc),
),
):
return host_memory.available_host_memory_bytes(
allow_cgroup_fallback=allow_cgroup_fallback
)
def test_missing_usage_for_known_limit_fails(self):
self.configure()
self.memory("task/engine", 100, 1000)
(self.mount / "task/engine/memory.current").unlink()
with self.assertRaises(FileNotFoundError):
host_memory._cgroup_memory_headroom(self.proc)
def without_cgroupfs(self, v1=False):
# A container sharing the host cgroup namespace, no cgroupfs mounted.
self.configure("/system.slice/engine.scope", v1=v1)
(self.proc / "self/mountinfo").write_text("1 0 0:1 / /proc rw - proc proc rw\n")
def test_unavailable_cgroup_requires_explicit_sizing(self):
for v1 in [False, True]:
for mounted in [False, True]:
with self.subTest(v1=v1, mounted=mounted):
if mounted:
self.configure("/system.slice/engine.scope", v1=v1)
else:
self.without_cgroupfs(v1=v1)
with self.assertRaisesRegex(RuntimeError, "set --hicache-size"):
self.available()
self.assertEqual(self.available(allow_cgroup_fallback=True), 5000)
def test_missing_counters_for_known_limit_requires_explicit_sizing(self):
for name in ("memory.current", "memory.stat"):
with self.subTest(name=name):
self.configure()
self.memory("task/engine", 100, 1000)
(self.mount / "task/engine" / name).unlink()
with self.assertRaisesRegex(RuntimeError, "set --hicache-size"):
self.available()
self.assertEqual(self.available(allow_cgroup_fallback=True), 5000)
if __name__ == "__main__":
@@ -641,6 +641,11 @@ class _GraftReq:
def get_fill_ids(self):
return array("q", self.fill_ids)
def refresh_fill_ids(self):
pass # fill ids are fixed for the stand-in
def finished(self):
return self.finished_reason is not None
class TestRotationGraftDecline(CustomTestCase):
"""The overlap disagg-prefill loop plans batch t+1 before batch t's radix
@@ -37,15 +37,17 @@ class TestPPPrefetchTicket(unittest.TestCase):
c.pp_size, c.tp_size = 4, 2
c.pp_group, c.pp_prefetch_command_group = "pp", "command"
c.pp_prefetch_command_thread = None
c.prefetch_hits_sync_groups = c.prefetch_completion_sync_groups = ["pp", "tp"]
c.prefetch_hits_sync_groups = ["pp", "tp"]
c.num_prefetch_io_lanes = 1
c.prefetch_completion_sync_groups_by_lane = [["pp", "tp"]]
c.pp_prefetch_states, c.pp_prefetch_decisions = {}, {}
c.pp_prefetch_state_lock = threading.Lock()
c.pp_prefetch_command_queue = Queue()
c.prefetch_queue = Queue()
c.prefetch_sync_queue = Queue()
c.prefetch_sync_queues = [Queue()]
c.ack_prefetch_queue = Queue()
c.host_mem_release_queue = Queue()
c.prefetch_buffer = Queue()
c.prefetch_io_queues = [Queue()]
c.storage_stop_event = threading.Event()
c.prefetch_tokens_occupied = 0
c.mem_pool_host = Mock(page_size=4)
@@ -103,7 +105,7 @@ class TestPPPrefetchTicket(unittest.TestCase):
def sync_acks(self, *acks):
for ack in acks:
self.c.prefetch_sync_queue.put(ack)
self.c.prefetch_sync_queues[0].put(ack)
with patch.object(
self.c.storage_stop_event,
"is_set",
@@ -183,7 +185,7 @@ class TestPPPrefetchTicket(unittest.TestCase):
self.assertFalse(cache.check_prefetch_progress(handle))
c._storage_hit_query.reset_mock()
self.run_commands(ticket)
operation = c.prefetch_buffer.get_nowait()
operation = c.prefetch_io_queues[0].get_nowait()
self.assertEqual(operation.handle, handle)
self.assertEqual(
operation.hash_value, get_storage_hash_str(ticket.prefetch_key, page_size=4)
@@ -237,7 +239,7 @@ class TestPPPrefetchTicket(unittest.TestCase):
),
):
c.prefetch_io_aux_func()
acks = [c.prefetch_sync_queue.get_nowait() for _ in range(8)]
acks = [c.prefetch_sync_queues[0].get_nowait() for _ in range(8)]
self.assertEqual([a.rid for a in acks], ["hit"] * 4 + ["next"] * 4)
self.assertEqual(
[a.completed_tokens for a in acks], [0, 0, None, None, 4, 8, None, None]
@@ -261,7 +263,7 @@ class TestPPPrefetchTicket(unittest.TestCase):
cache.finish(ticket.handle, CacheRequestOutcome.ABORT)
self.assertIs(c.pp_prefetch_decisions["hit"], PPPrefetchDecision.CANCELLED)
self.run_commands(ticket)
operation = c.prefetch_buffer.get_nowait()
operation = c.prefetch_io_queues[0].get_nowait()
self.sync_acks(PrefetchAck("hit", operation, completed_tokens=4))
c.mem_pool_host.free.assert_not_called()
self.sync_acks(PrefetchAck("hit", operation, completed_req=True))
@@ -293,7 +295,7 @@ class TestPPPrefetchTicket(unittest.TestCase):
c.pp_prefetch_states.clear()
c.mem_pool_host.alloc.reset_mock()
self.run_commands(ticket)
operation = c.prefetch_buffer.get_nowait()
operation = c.prefetch_io_queues[0].get_nowait()
self.assertEqual(
c.mem_pool_host.alloc.call_args_list,
[
@@ -311,7 +313,7 @@ class TestPPPrefetchTicket(unittest.TestCase):
c.mem_pool_host.get_pool.side_effect = KeyError(PoolName.SWA)
with self.assertLogs(level="ERROR"):
self.run_commands(ticket)
self.assertTrue(c.prefetch_buffer.get_nowait().is_terminated())
self.assertTrue(c.prefetch_io_queues[0].get_nowait().is_terminated())
self.assertEqual(c.mem_pool_host.free.call_count, 1)
self.assertEqual(c.mem_pool_host.free.call_args.kwargs["pool"], PoolName.KV)
@@ -387,8 +389,8 @@ class TestPPPrefetchTicket(unittest.TestCase):
self.assertEqual(invocation.kwargs, {"src": 1})
self.assertEqual(c.pp_prefetch_command_queue.unfinished_tasks, 0)
self.assertTrue(c.pp_prefetch_command_queue.empty())
self.assertIs(c.prefetch_buffer.get_nowait(), operation)
self.assertTrue(c.prefetch_buffer.empty())
self.assertIs(c.prefetch_io_queues[0].get_nowait(), operation)
self.assertTrue(c.prefetch_io_queues[0].empty())
c.mem_pool_host.alloc.assert_called_once()
def test_idle_downstream_skips_empty_broadcasts_before_ticket_and_stop(self):
@@ -404,8 +406,8 @@ class TestPPPrefetchTicket(unittest.TestCase):
patch.object(c.pp_prefetch_command_queue, "task_done") as task_done,
):
c.pp_prefetch_command_thread_func()
self.assertEqual(c.prefetch_buffer.get_nowait().request_id, "hit")
self.assertTrue(c.prefetch_buffer.empty())
self.assertEqual(c.prefetch_io_queues[0].get_nowait().request_id, "hit")
self.assertTrue(c.prefetch_io_queues[0].empty())
c.mem_pool_host.alloc.assert_called_once()
task_done.assert_not_called()
@@ -167,7 +167,7 @@ class TestSharedPrefillAdmission(unittest.TestCase):
with get_parallel().override(attn_dcp_size=1):
scheduler.init_req_max_new_tokens(req)
self.assertEqual(req.sampling_params.max_new_tokens, 1)
req._refresh_fill_ids()
req.refresh_fill_ids()
cache = SimpleNamespace(
sliding_window_size=page_size,
@@ -78,7 +78,7 @@ class TestSessionTokenShare(CustomTestCase):
if baked is None:
baked = len(output)
req.output_ids.extend(output[:baked])
req._refresh_fill_ids()
req.refresh_fill_ids()
req.output_ids.extend(output[baked:])
self.session.finish_req(req)
@@ -114,7 +114,7 @@ class TestSessionTokenShare(CustomTestCase):
r2 = self._create("r2", [50, 51])
self.assertEqual(list(r2.origin_input_ids), in1 + out1 + [50, 51])
r2.output_ids.extend([6, 7])
r2._refresh_fill_ids()
r2.refresh_fill_ids()
self.session.abort_req()
self.assertEqual(self.session.committed_origin_len, len(in1))
@@ -149,7 +149,7 @@ class TestSessionTokenShare(CustomTestCase):
# array, then output trimmed to finished_len (like _trim_overshoot)
# before finish.
r1.output_ids.extend([1, 2, 3, 4, 5, 6])
r1._refresh_fill_ids()
r1.refresh_fill_ids()
del r1.output_ids[4:]
self.session.finish_req(r1)
self.assertEqual(
@@ -161,7 +161,7 @@ class TestSessionTokenShare(CustomTestCase):
r2 = self._create("r2", [50])
self.assertEqual(list(r2.origin_input_ids), in1 + [1, 2, 3, 4] + [50])
self.assertEqual(len(r2.full_untruncated_fill_ids), 0) # carry skipped
r2._refresh_fill_ids()
r2.refresh_fill_ids()
self.assertEqual(list(r2.full_untruncated_fill_ids), list(r2.origin_input_ids))
@@ -3635,14 +3635,14 @@ class UnifiedRadixCacheSuite:
gate = threading.Event()
swa_release_q = cc.extra_host_mem_release_queues.get(PoolName.SWA)
def _page_transfer_hook(_self, op):
def _page_transfer_hook(_self, op, sync_queue):
# Signal that the worker has reached the sidecar step, then block
# until the main thread releases the gate.
entered.set()
gate.wait(timeout=10.0)
# Forward to the real implementation so the worker runs the
# terminate-aware release path itself.
original_sidecar(_self, op)
original_sidecar(_self, op, sync_queue)
# Simulate a slow prefetch IO. Hook on _page_transfer.
with mock.patch.object(
@@ -5290,7 +5290,7 @@ class UnifiedRadixCacheSuite:
cc = cache.cache_controller
# Fake _reduce_prefetch_ack to drop SWA pool_hits when drop_swa is True.
def fake_reduce(ack):
def fake_reduce(ack, groups=None):
if drop_swa and ack.pool_hits is not None:
if PoolName.SWA.value in ack.pool_hits:
ack.pool_hits[PoolName.SWA.value] = 0
@@ -5368,9 +5368,11 @@ class UnifiedRadixCacheSuite:
self.skipTest("fixture does not exercise SWA L3 prefetch")
return storage_dir, seq
def test_tp_swa_prefetch_dropped_when_peer_misses(self):
"""A peer rank missing the SWA window drops the whole prefetch result
on every rank (TP-consistent all-or-nothing)."""
def test_tp_swa_prefetch_degrades_to_kv_only_when_peer_misses(self):
"""A peer rank missing the SWA window degrades the prefetch to KV-only
on every rank (TP-consistent fail-soft: the fetched KV prefix is
retained, only the failed SWA pool is discarded and rebuilt at
load-back)."""
setup = self._setup_swa_tp_prefetch()
if setup is None:
return
@@ -5382,7 +5384,10 @@ class UnifiedRadixCacheSuite:
self._consume_prefetch(cons, seq, CacheRequestHandle("drop", 0))
m = cons.match_prefix(MatchPrefixParams(key=RadixKey(array("q", seq))))
self.assertEqual(m.host_hit_length, 0)
# Fail-soft: the KV pool fetched a usable prefix and it is retained.
self.assertEqual(m.host_hit_length, len(seq))
# TP consistency: every rank agrees SWA itself is missing — the
# degraded span carries no SWA host mirror (rebuilt from Full KV).
self.assertFalse(
self._swa_host_on_path(cons, seq), "SWA must be dropped when a peer misses"
)
@@ -5409,9 +5414,11 @@ class UnifiedRadixCacheSuite:
)
cons.sanity_check()
def test_tp_swa_prefetch_drop_frees_host_pool(self):
"""A dropped SWA prefetch must return its whole host buffer to the pool
(no leak, no over-free)."""
def test_tp_swa_prefetch_degrade_frees_failed_swa_host_pool(self):
"""A degraded prefetch (peer missing SWA) must return the failed SWA
pool's host buffer to the pool (no leak, no over-free) while the
retained KV span stays occupied — the fail-soft degrade never
double-frees the kept side (TP-consistent on every rank)."""
setup = self._setup_swa_tp_prefetch()
if setup is None:
return
@@ -5420,18 +5427,25 @@ class UnifiedRadixCacheSuite:
cons = self._l3_consumer(storage_dir)
cons.tp_world_size = 2
self._patch_tp_prefetch_sync(cons, drop_swa=True)
avail_before = cons.swa_kv_pool_host.available_size()
avail_before = self._host_avail_sizes(cons)
self._consume_prefetch(cons, seq, CacheRequestHandle("drop", 0))
self.assertEqual(
cons.match_prefix(
MatchPrefixParams(key=RadixKey(array("q", seq)))
).host_hit_length,
0,
len(seq),
)
# Whole window dropped -> its host buffer is fully released back.
cons.drain_storage_control_queues() # Drain the release queue.
self.assertEqual(cons.swa_kv_pool_host.available_size(), avail_before)
avail_after = self._host_avail_sizes(cons)
# Failed pool: its whole SWA window is released back (no leak).
self.assertEqual(
cons.swa_kv_pool_host.available_size(), avail_before[PoolName.SWA]
)
# Retained side: the KV graft keeps its host buffer occupied — the
# degrade must not release the KV it just decided to keep.
self.assertLess(avail_after[PoolName.KV], avail_before[PoolName.KV])
cons.sanity_check()
def test_hicache_write_back_evict_drops_unbacked_leaf_when_host_full(self):
"""Write-back eviction will keep freeing device KV when the host pool
@@ -6636,35 +6650,32 @@ class UnifiedRadixCacheSuite:
leaf, ComponentType.SWA, EvictLayer.HOST
)
cache._free_values(evicted.device_frees, evicted.host_frees)
full_host_pool = cache.cache_controller.mem_pool_host
swa_host_pool = cache.components[ComponentType.SWA]._swa_kv_pool_host
full_available_before = full_host_pool.available_size()
swa_available_before = swa_host_pool.available_size()
result = cache.match_prefix(MatchPrefixParams(key=RadixKey(array("q", tokens))))
# Fail-soft semantics (#38348): the both-layers-absent SWA hole on a
# backuped node is a valid match boundary — the FULL span hits on
# host (no storage round-trip), and the missing SWA window is rebuilt
# from Full KV when the continuing request's insert overlaps the hole.
self.assertEqual(result.full_kv_hit_length, len(tokens))
self.assertEqual(result.swa_branching_seqlen, len(tokens))
self.assertEqual(result.host_hit_length, leaf_len)
self.assertIsNone(result.swa_branching_seqlen)
self.assertEqual(result.swa_host_hit_length, 0)
# Load-back materializes the FULL span on device; the continuing
# request's insert overlaps the hole node and rebuilds its in-window
# SWA from Full KV.
self._load_back_node(cache, leaf)
self._insert(
cache,
allocator,
req_to_token_pool,
tokens[: result.swa_branching_seqlen],
)
cache.writing_check(write_back=True)
# Full was already backed up, so only the SWA window is allocated.
self.assertEqual(full_host_pool.available_size(), full_available_before)
self.assertEqual(
swa_host_pool.available_size(),
swa_available_before - leaf_len,
cache, allocator, req_to_token_pool, tokens + self._make_seq(2000, 1)
)
rematch = cache.match_prefix(
MatchPrefixParams(key=RadixKey(array("q", tokens)))
)
self.assertEqual(len(rematch.device_indices), result.swa_branching_seqlen)
rematch = cache.match_prefix(MatchPrefixParams(key=RadixKey(array("q", tokens))))
self.assertEqual(len(rematch.device_indices), len(tokens))
self.assertIsNone(rematch.swa_branching_seqlen)
# The hole node's in-window tail now carries device SWA again.
tail = _device_value(cache, rematch.last_device_node, ComponentType.SWA)
self.assertIsNotNone(tail)
def test_swa_branching_seqlen_caps_insert_after_forward(self):
if (
@@ -7833,10 +7844,17 @@ class UnifiedRadixCacheSuite:
self.assertEqual(xfer.nodes_to_load, [y])
self.assertEqual(int(xfer.host_indices.numel()), ps)
with self.assertRaises(AssertionError):
cache.tree_core.build_hicache_transfers(
ComponentType.SWA, y, CacheTransferPhase.LOAD_BACK
)
# Anchored on Y, the walk reaches N's both-layers-absent SWA hole:
# the fail-soft LOAD_BACK builder stops collecting there instead of
# asserting — the missing SWA window is rebuilt from Full KV at
# load-back via SWARebuild/RecoverSWAWithLockedFull.
transfers = cache.tree_core.build_hicache_transfers(
ComponentType.SWA, y, CacheTransferPhase.LOAD_BACK
)
self.assertEqual(len(transfers), 1)
xfer = transfers[0]
self.assertEqual(xfer.nodes_to_load, [y])
self.assertEqual(int(xfer.host_indices.numel()), ps)
def test_hicache_swa_finalize_anchored_on_best_match_node(self):
cache, _, _, y, x, _ = self._swa_anchor_setup()
@@ -8236,6 +8254,68 @@ class UnifiedRadixCacheSuite:
cache.sanity_check()
def test_write_back_swa_publish_leaves_an_ancestor_pending_under_another_ack_alone(
self,
):
"""Write-back builds the insert backup as [target] only, so a second SWA
window publish never re-marks an ancestor another ack still holds."""
if not self.cfg.has_swa:
self.skipTest("requires SWA")
if self._skip_unsupported_hicache_test():
return
ps = self.cfg.page_size
if self.cfg.sliding_window_size <= 2 * ps:
self.skipTest("window must span past the leaf and its parent to reach c")
cache, allocator, req_to_token_pool = build_fixture(self.cfg)
self._init_hicache(cache, write_policy="write_back")
# c(Full unbacked) -> b(Full backed) -> t1(Full backed)
# -> d(Full unbacked) -> t2(Full backed)
# Every node holds a device-only SWA value inside the window.
c_seq = self._make_seq(1, 2)
t1_seq = c_seq + self._make_seq(1000, 1) + self._make_seq(2000, 1)
t2_seq = c_seq + self._make_seq(3000, 1) + self._make_seq(4000, 1)
for seq in (t1_seq, t2_seq, t1_seq[:-ps], t2_seq[:-ps]):
self._insert(cache, allocator, req_to_token_pool, seq)
def leaf(seq):
return cache.match_prefix(
MatchPrefixParams(key=RadixKey(array("q", seq)))
).last_device_node
c, b, t1 = self._path_chain(cache, leaf(t1_seq))
_, d, t2 = self._path_chain(cache, leaf(t2_seq))
for node in (b, t1, t2):
cache.tree_core.set_component_host_value_raw(
node,
ComponentType.FULL,
_device_value(cache, node, ComponentType.FULL).clone(),
)
for node in (c, b, t1, d, t2):
self.assertIsNotNone(_device_value(cache, node, ComponentType.SWA))
self.assertIsNone(_host_value(cache, node, ComponentType.SWA))
# Two requests in one batch finish at t1 and t2; neither ack drains.
with mock.patch.object(
cache,
"_execute_and_commit_kv_backup",
wraps=cache._execute_and_commit_kv_backup,
) as backups:
for seq, target in ((t1_seq, t1), (t2_seq, t2)):
backups.reset_mock()
self._insert(cache, allocator, req_to_token_pool, seq)
self.assertEqual(
[call.args[0].node_ids for call in backups.call_args_list],
[[target]],
)
for node, ack in ((c, t1), (b, t1), (t1, t1), (d, t2), (t2, t2)):
self.assertEqual(cache.tree_core.get_write_through_pending_id(node), ack)
# Write-back defers the unbacked Full prefix to eviction.
self.assertFalse(cache.tree_core.is_backuped(c))
self.assertFalse(cache.tree_core.is_backuped(d))
class UnifiedLRUListBoundedRefreshTest(CustomTestCase):
components = (ComponentType.FULL, ComponentType.SWA)
@@ -37,10 +37,15 @@ class C_HiCacheController(BaseHook):
def wrapped_init(self, *args, **kwargs):
self.sim_prefetch_buffer = Queue()
result = original_init(self, *args, **kwargs)
# The real IO thread normally creates this queue. The simulator
# replaces that thread, so initialize the handoff queue here.
if hasattr(self, "prefetch_hit_queue"):
self.prefetch_buffer = Queue()
# The real IO threads normally create the handoff queues. The
# simulator replaces those threads, so initialize the per-lane
# handoff queues here.
if hasattr(self, "prefetch_hit_queue") and not getattr(
self, "prefetch_io_queues", None
):
self.prefetch_io_queues = [
Queue() for _ in range(getattr(self, "num_prefetch_io_lanes", 1))
]
return result
def wrapped_append_host_mem_release(self, host_indices):
@@ -182,11 +187,10 @@ class C_HiCacheController(BaseHook):
# Feed operations whose host pages were allocated by the scheduler
# into the virtual-time transfer loop.
prefetch_buffer = getattr(self, "prefetch_buffer", None)
if prefetch_buffer is not None:
while not prefetch_buffer.empty():
for io_queue in getattr(self, "prefetch_io_queues", None) or []:
while True:
try:
operation = prefetch_buffer.get(block=False)
operation = io_queue.get(block=False)
if operation is not None:
self.sim_prefetch_buffer.put(operation)
except Empty:
@@ -295,10 +299,12 @@ class C_HybridCacheController(BaseHook):
def wrapped_init(self, *args, **kwargs):
result = original_init(self, *args, **kwargs)
if hasattr(self, "prefetch_hit_queue") and not hasattr(
self, "prefetch_buffer"
if hasattr(self, "prefetch_hit_queue") and not getattr(
self, "prefetch_io_queues", None
):
self.prefetch_buffer = Queue()
self.prefetch_io_queues = [
Queue() for _ in range(getattr(self, "num_prefetch_io_lanes", 1))
]
return result
def wrapped_storage_hit_query(self, operator):