Promote the validated ds41-gist-exp baseline:
- image recipe-ns-l3ssd-433257c7b-baked-20261115 (SGLang + Mooncake 0.3.14 + L3 SSD)
- P: cache / no-optimistic / fastokens / tokenizer-worker 16 / L3 SSD 4T
- D: fastokens / tokenizer-worker 16
- gateway: normal health/circuit, no P admission caps
- c580 baseline 24.045 rps; c620 and buffer_only results recorded
- Feishu baseline doc and deploy manifests included
This is the base for the next Dynamo layer.
Backport the upstream namespace-tool stack onto the B300 recipe base:
- #46370: Function.namespace / ChatMessage.action protocol fields.
- #49625: Responses/Messages native recipe rendering (convert_native),
tool_search/defer_loading, namespace identity in/out/history replay.
Adapted to this base: skips serving_kimi_k3.py (not present), drops the
Minimax/Beam/wire-only dependencies added between 10-23 and the PR base,
and keeps the existing tool-parsing fallback where the newer
`parses_without_tools()` helper is unavailable.
Also includes the PR's Anthropic tool-reference plumbing and unit tests.
A 2 ms phase whose CPU clock reads 0.1 us over its wall clock is clock
granularity, not an accounting bug; the anomaly flag fired on nearly every
short phase in the live verification run.
Only flag (and record non_cpu_raw_ms) when the deficit exceeds both 100 us and
1% of the phase wall. Larger real anomalies keep the previous behaviour.
The frontend sglang:num_requests_total has no dp_rank label (verified from the
series registry), so per-rank RPS was not derivable from any existing metric.
Add sglang:batch_requests_total{phase,dp_rank}, incremented by the scheduler
process that actually completed the request:
phase=prefill prefill compute finished and the request entered the PD
transfer queue (not a guarantee that the transfer succeeded)
phase=decode request finished decoding
Only the canonical scheduler rank reports (attn_tp_rank==0, attn_cp_rank==0,
last PP stage) — every TP/CP rank runs a scheduler and sees the same request,
so an ungated counter would be multiplied by TP*CP in the multiprocess
registry. Both hooks count each request exactly once: the decode path counts
after _handle_finish_state_updated_req and on the sampling-mask abort branch,
skipping rows that finished on an earlier step; the prefill path counts after
the request enters the disagg inflight queue (aborted/retried requests never
reach it). The per-request finished() check is skipped entirely when
SGLANG_BATCH_TELEMETRY is off.
Covers the design's unit-level acceptance points:
* disabled/unsampled loops never read the thread clock;
* wall/CPU/non-CPU semantics, unmeasured != zero, partial measurement omits
non-CPU, busy vs sleeping windows are distinguishable;
* CPU > wall flags an accounting anomaly with a raw negative delta;
* child spans carry the CPU split;
* repeated phases (prefix indices) stay attribute-only, inclusive phases are
not summed, and eviction demand/shortfall/plan plus Mamba plan/zero results
are recorded as separate counters.
Parent span scheduler.loop.prefill_batch_init records wall + the scheduler
thread's CPU at both existing close points (empty admission and normal
return); the CPU clock is only read when loop detail is enabled and sampled,
through the safe batch_telemetry.thread_cpu_ns() wrapper so a missing clock
degrades to 'not measured' instead of failing the scheduling path.
Fixed children, in call order:
schedule_batch_init, ready_to_load_host_cache, prepare_for_extend,
alloc_for_extend (+ prefix_device_indices attribute-only, sampling_batch_info),
evict_to_free_tokens, evict_for_alloc, prefill_stats (+ pending_tokens),
mixed_chunk_tail.
* prefill_stats wraps the whole assignment so _get_num_pending_tokens() is
included; mixed_chunk_tail only exists when that branch is taken.
* prefill-only shape counters (n_reqs/n_new/n_cached) are snapshotted before
mixed-chunk folding; a mixed batch additionally publishes decode_reqs and
merged totals under explicit *_merged_* names.
* eviction is timed at the call site (outside the allocator wrapper) so
rank-consensus waits are included, and covers base, SWA (x2), UnifiedSWA
(x3) and the extend Mamba reclaim through evict_for_alloc_observed().
Demand, shortfall and plan are separate counters; Mamba plan and realized
counts are both recorded (including zero).
All of it is gated by detail on a sampled loop.
Adds an optional thread CPU clock next to the existing wall clock:
* detail_begin/end record time.thread_time_ns() only when detail is on and the
loop is sampled; detail_add() accepts an explicit thread_cpu_ns for callers
that already hold absolute start/end points (the prefill_batch_init parent).
* phase records carry [wall, count, start_wall, end_wall, cpu_sum, cpu_count].
thread_cpu_ms and thread_cpu_measured_count are emitted whenever any
occurrence was measured; non_cpu_ms only when every occurrence was, so a
partial measurement never attributes unmeasured wall to non-CPU time. A
phase whose CPU exceeds its wall keeps the clamped value, adds
non_cpu_raw_ms and accounting_anomaly, and bumps a per-loop anomaly count.
* new detail_attr_sum()/detail_attr_setdefault() accumulate low-cardinality
numeric loop attributes (eviction tokens, free-token snapshots).
* prefix_device_indices joins the attribute-only prefixes so per-request
repeats never create spans.
* child spans carry phase_thread_cpu_ms / phase_cpu_measured_count /
phase_non_cpu_ms, so the CPU split is visible without the loop attributes.
Default-off path is unchanged: no thread clock, no new dicts, no spans.
The detail-phase refactor made residual a required positional argument while
test_loop_span_accounting.py still calls _emit_pass_span with four arguments,
so that test failed on the base commit. Default it to 0.0.
Split out of the prefill_batch_init observability series to keep that review
scope focused.
Deriving each candidate's uncached estimate from its own recent hit rate is
self-reinforcing: a rank that falls behind sees its hit rate drop (fewer
chances to hit), gets charged more for every new request, and falls further
behind. Measured on c580: one rank ended 16% below the others, 30 s cross-rank
dispersion max 952x, and 9.6% lower rps than the total_tokens arm.
Charge the fleet-average uncached size instead and rank by the published
num_waiting_uncached_tokens, which already carries each rank's exact prefix
match.
DP balancing on total_tokens charges each request's whole prompt, but with
prefix caching only the unmatched part costs prefill compute. On a workload
with a 95% hit rate the prompt is ~95% noise: per request the uncached part
is a few percent of the prompt yet varies by more than an order of magnitude
(p50 1.9k / p99 59k tokens against a 95k prompt), so one cold request
outweighs many hot ones and per-rank prefill work drifts apart.
Add --load-balance-method uncached_tokens: DPBudget tracks the published
LoadSnapshot.num_waiting_uncached_tokens (prefix match minus accepted storage
hits) plus the rank's recent cache_hit_rate for a newly dispatched request.
SGLANG_MOONCAKE_CLIENT_METRICS_PORT_BASE=<base> makes MooncakeStore.setup()
pass enable_http_server=True and client_http_port=base+dp_rank, so each DP
rank's embedded Mooncake client exposes /metrics (needs MC_STORE_CLIENT_METRIC=1).
Off by default.
Phase 1/3 of the async-backup handoff instrumentation plan:
* new fwd_timeline probe (SGLANG_FWD_TIMELINE=1): one CUDA event pair around
the model execute call per forward, resolved without blocking and written as
wall-clock JSONL per DP rank. Joined with loop spans it separates host time
spent while the GPU is busy (overlapped) from GPU-idle host stalls.
* detail phases: fwd_init_new, fwd_prepare, fwd_execute, fwd_graph_load,
fwd_graph_replay, fwd_post, fwd_runner_tail,
prefill_policy, prefill_adder_init, prefill_scan, prefill_batch_init.
All spans ride the existing SGLANG_TRACE_LOOP_DETAIL path; the GPU probe is
off by default.
Adds the Phase A instrumentation from the async-backup handoff (§7.1):
* PassTimings: pending_abort / chunk / new_batch / set_sched (+ chunk_count)
* prefill.get_next_disagg_prefill_batch_to_run: time
process_pending_chunked_abort, process_prefill_chunk,
get_new_batch_prefill and set_schedule_time_batch
* scheduler.prefill.pass gains sglang.{pending_abort,chunk,new_batch,set_sched}_ms,
sglang.get_next_residual_ms (= other - these four) and sglang.chunk_count;
the existing sglang.other_ms is unchanged for dashboard compatibility.
~5 monotonic calls per pass; enabled by SGLANG_BATCH_STAGE_TIMING (already on).
Net async change (from exp-async-backup-perop-20261107) applied onto
b300-pick-46884-20261107. Runtime files applied cleanly; the new tests and
the timeline probe are added as-is.
Interaction to validate on hardware: #46884 publishes prefill read-done before
the replay, so the scheduler-stream fence that the index copy waits on is
shorter; the async worker's wait should shrink and the copy may overlap more of
the in-flight forward. Re-baseline the 4-arm A/B on this branch.
* Normalize the host-pool layout to a string (a None attribute would make
the OTel SDK drop the whole flush span).
* Drop any remaining None-valued flush attributes for the same reason.
Adds loop-detail phases outside run_batch/process_result/get_next:
* inflight_poll: process_disagg_prefill_inflight_queue() (overlap loop)
* inflight_release: release_kv_cache(checkpoint=True) + tree_cache.finish()
inside it, so the checkpoint/insert cost is separated from the poll
* launch_sample: launch_batch_sample_if_needed()
The get_next side already publishes match/adder/hicache/bootstrap/mlp_sync
and a pass residual (sglang.other_ms) on scheduler.prefill.pass.
Adds the phase-3/phase-4 instrumentation from
handoff/prefill-flush-backups-validation-20261107.md:
* process_batch_result outer phases: flush_trace, load_snapshot,
flush_backups, step_counters, metrics_report. process_residual now
subtracts them, so the 37.5% residual is attributable.
* SGLANG_TRACE_BACKUP_DETAIL=1 (default 0) traces one flush:
merge_ops / move_write (anchor + per-sidecar indices) / indices_cpu /
index_sort / index_select / write_fence / build_l2 / submit_d2h
(start-event wait, prepare, per-pool launch, ack events) / ack_enqueue,
plus token/op/pool/device dimensions and ack queue depth.
* Per-flush resolution: one scheduler.cache_flush_backup span per
non-empty flush, child of the loop span. Sub-phases never enter the
loop aggregate, so the residual identity cannot double count.
* Empty vs non-empty flushes are counted on the loop.
All timings are scheduler host wall, not device transfer time.
Review follow-up. match_kv_cache also runs on waiting-queue probes
(calc_priority), PD decode _match_prefix_and_lock and HiCache rematches.
Emitting the counter there counted the same gap once per scheduling round
and attributed "re-prefilled" tokens to decode engines.
- match_kv_cache now only records req.full_kv_hit_length (plus the match-time
gap); it no longer emits.
- the metric is emitted once in PrefillAdder._commit_prefill_admission, where
the accepted prefix is final after any host load-back.
- rename the request counter to swa_gate_reject_admissions_total and document
that a retracted and re-admitted request is counted again (it is re-prefilled).
- pre-create zero-valued children so gate-off deployments expose the series
before the first rejection.
- tests: repeated match does not emit, admission emits once, a retry counts
again, and host load-back closes the gap.
Not runtime verified (no GPU run on this build).
The unified-cache SWA component gates prefix matching on a live sliding
window: when the window is missing, the match stops before an otherwise
available Full-KV prefix. The existing hit counters do not show the tokens
lost to that gate, which makes the gate-on/gate-off trade-off invisible.
Record the gap between the longest Full-KV match and the accepted prefix
(full_kv_hit_length - (device_prefix_len + host_hit_length)) on the Req and
export it as:
sglang:swa_gate_reject_tokens_total
sglang:swa_gate_reject_reqs_total
This is the token volume that will be re-prefilled because a component
validator (SWA window; Mamba where present) rejected the match. With the
gate enabled it measures the hit length lost to a missing window; with the
gate disabled it only captures the host-only / load-back-gap fallback.
Includes unit tests for the gap and the host-hit case. Not runtime verified
yet (no GPU run on this build).
Adds an opt-in scheduler.loop span (SGLANG_TRACE_LOOP_SPAN) with recv /
get_next / run_batch / process_result / kv_send children. The span is only
created for sampled loops that have pending work; fully idle loops stay in
the counters so the 714 pass/s idle spin never floods Jaeger.
No-batch accounting is always on with SGLANG_BATCH_STAGE_TIMING:
- sglang:scheduler_no_batch_{seconds,total}{dp_rank,reason,forward_overlap}
- sglang:scheduler_dp_lockstep_idle_seconds_total{dp_rank,forward_overlap}
with the design's reason priority (result_backpressure, deferred_input,
waiting_bootstrap, waiting_prefetch, admission_blocked, waiting_pd,
fully_idle, no_ready_other) and local_batch_kind/global_outcome from the DP
sync. DP lockstep idle is counted separately from pure no-batch wait.
loop gpu_ms reuses the existing DeviceTimer intervals (no new CUDA events);
the forward.gpu lane is optional and approximate. Pass/recv spans become
children of the loop span when it is active.
Tests: test_loop_span_accounting (5) and the existing stage-timing suite (8)
pass.
_handle_prefetch_result() retires the operation with a direct
del ongoing_prefetch[request] instead of _retire_ongoing_prefetch(), so
the ready-latency hook only ever saw miss/revoke paths: with a cache-hit
workload storage_prefetch_ready_seconds{result="hit"} stayed at 0 even
though storage_prefetch_hit_tokens_total was moving ~312k tok/s.
Observe right after the del, classifying by loaded_from_storage.
The first version stamped the request on its first in-flight
check_prefetch_progress() call. A fast L3 query can be drained by
_process_hicache_events before that check runs (the op is already gone
from ongoing_prefetch), so those requests produced no sample -- exactly
the common case.
Stamp at issue time in prefetch_from_storage() and observe once at
_retire_ongoing_prefetch(), classifying by operation.storage_hit_count.
Retire is the single exit for hit completion, below-threshold revoke,
anchor loss and host-capacity fallback, so hit and miss are both covered.
No behavior change; the hook is still behind SGLANG_BATCH_STAGE_TIMING.
Under --hicache-storage-prefetch-policy wait_complete a request is not
admitted until its L3 lookup resolves, and a miss pays the same query round
trip as a hit. The scheduler only exposed a skip counter, so the actual wait
was unknown.
Add sglang:storage_prefetch_ready_seconds{dp_rank, result=hit|miss}:
* stamp cache_request_handle on its first in-flight check,
* observe at the moment check_prefetch_progress() turns usable, split by
whether prefetch_loaded_tokens_by_reqid ended up > 0,
* clear the stamp on revoke / abort / accounting discard.
Buckets cover 1 ms .. 30 s. Gated by SGLANG_BATCH_STAGE_TIMING (off by
default). The measurement is ~2 dict ops + one observe per request, not per
pass.
Tests: test/registered/unit/observability/test_batch_telemetry_stage_timing.py
adds the hit/miss split assertion (9 tests, all passing).
Runtime accounting on the prefill service (64 x 14.6k-token requests,
8-way concurrency) showed the non-forward wall time is dominated by the
scheduler receive path, not by radix match:
forward (extend+idle) 58.2% | get_next 6.9% | recv_requests 25.6%
process_batch_result 1.8% | run_batch overhead 0.9% | other ~6%
The existing pass span only covers get_next, so it cannot explain that
25.6%. Split it with the same opt-in gate:
* sglang:recv_requests_seconds{dp_rank, part} (total/pull/broadcast/
unwrap/mm), 5 new series per rank
* sglang:recv_requests_msgs_total, sglang:recv_requests_prompt_tokens_total
* OTLP span scheduler.recv_requests with the same breakdown, sampled
deterministically by a per-process counter (SGLANG_RECV_SPAN_SAMPLE,
default 0.05; SGLANG_RECV_SPAN=0 disables the span only)
Note on scope: tokenization itself does NOT run in the scheduler. The
tokenizer workers (--tokenizer-worker-num) produce TokenizedGenerateReqInput
and the scheduler receives the already-tokenized ids. What this measures is
ZMQ pull + decode, the cross-rank broadcast of the pickled objects, and
unwrap_pickle_fields() of the token-id buffers -- which is why large prompts
make recv_requests expensive.
Tests: test/registered/unit/observability/test_batch_telemetry_stage_timing.py
now also covers the recv histogram parts, counters, zero-message passes and
deterministic recv-span sampling (8 tests total, all passing).
Motivation: on the PD prefill node the wall-clock budget is dominated by
get_next_batch_to_run (~50% of the hour on 2026-11-02 16:00-17:00), and
forward_overlap=partial is only 21-87 s/h, i.e. it does not overlap the GPU
forward. We know the total but not which segment inside it costs: radix
match, PrefillAdder admission, the HiCache event collective, the bootstrap
all-reduce, or the DP metadata gather.
Measurement design (one sample per scheduling pass, including passes that
produce no batch; see handoff/prefill-batch-stage-timing-20261103.md v2/v3):
L1 7 per-pass histograms:
sglang:batch_prefill_{get_next,match,adder,hicache_events,
bootstrap,mlp_sync,other}_seconds
other = total - (match + adder + hicache + bootstrap + mlp_sync),
clamped to >= 0. Nested stages are not double counted.
L3 5 per-pass counters:
candidates / skipped_prefetch / admitted / resumed_chunk / new_chunk
L2 a dedicated per-pass span scheduler.prefill.pass (not the batch span),
emitted from end_pass() on every pass. Its sampling decision is a pure
function of mlp_sync_seq, so a sampled gather round is sampled on all
DP ranks. Idle ranks and ranks without a local batch still emit, which
is exactly what the cross-rank skew analysis needs. No links.
Hooks are installed in both scheduling loops:
* get_next_disagg_prefill_batch_to_run() for --disaggregation-mode prefill
(hicache / bootstrap / mlp_sync segments), and
* get_next_batch_to_run() for standalone prefill (hicache / mlp_sync),
so the instrumentation can be validated on ds41-prefill-dp4 without a decode
peer. match/adder/counters live in the shared get_new_batch_prefill path.
Semantics:
* match times only init_next_round_input; check_prefetch_progress counts as
skipped_prefetch, never as match.
* bootstrap and hicache_events are cross-rank sync points measured at the
outer call, not attributed to local CPU work.
* dp_gather_ms comes from the existing MLPSyncBatchInfo.sync_wait_seconds;
sglang:dp_attention_sync_wait_seconds is reused, no duplicate metric.
The value is cleared before each outer call and mlp_sync_seq only
increments when a gather actually ran, so stale values cannot leak.
* resumed vs new chunked requests are captured separately: had_chunked is
recorded at pass start (before process_prefill_chunk/admission can
consume the last chunk) and new_chunk from adder.new_chunked_req. This
fixes both the first-chunk and last-chunk mislabels.
Opt-in:
* SGLANG_BATCH_STAGE_TIMING=1 (default 0);
SGLANG_BATCH_PASS_SPAN=1 / SGLANG_BATCH_PASS_SPAN_SAMPLE=0.05.
* When off every hook is one module-level bool check; no logging, no CUDA
calls, no new collectives.
Tests: test/registered/unit/observability/test_batch_telemetry_stage_timing.py
covers first/middle/last chunk labelling, all-prefetch-skip passes without a
batch, stale-gather clearing, rank-deterministic pass-span sampling, and the
disabled no-op path. py_compile passes on all touched files.
These three files were already running in the gist production image
(gist-nvlink-ee9170e7bb-pzcfix) but were never committed: the image was
produced by docker-commit with the btelem4/5/6 files overlaid on top of
the built source tree.
Port them so the tree matches the image before layering new instrumentation:
* observability/batch_telemetry.py: per-batch histograms, per-DP-step
shape histograms and the sampled OTLP span scheduler.batch.{prefill,decode}.
* managers/scheduler_components/metrics_reporter.py: observe_prefill/
observe_decode call sites and the batch-trace log suffix.
* managers/scheduler_components/dp_attn.py: observe_dp_step().
Content is identical to the image; only line endings were normalised to LF
(the image copy was mixed CRLF/LF). No behavior change when
SGLANG_BATCH_TELEMETRY is unset (default off).
With long prompts and L3 under pressure, a matched prefix can contain a node
whose SWA component has neither a device nor a host value (host tombstone
evicted between match and load-back). build_hicache_transfers asserted on
this and killed the prefill scheduler.
Add SWAComponent.has_load_back_gap() and have build_load_back_spec return an
empty transfer spec when a gap is found, so the caller backs off and the
request is recomputed. Measured on peak15-v6 (1M, cap 64): 0 scheduler
exceptions, hit 59.4%->94.0%, TTFT p95 47.0s->3.4s with the SWA prefix-match
gate disabled.
The unified-cache SWA component gates prefix matching on a live SWA window,
so a long prefix only matches up to the last sliding window and the tail is
recomputed. Add SGLANG_DSV4_SWA_GATES_PREFIX_MATCH (default True) so the gate
can be turned off for measurement; the SWA window is then rebuilt/recovered
downstream as for the req-ring paths. Measured on 4xB300 DP4 at 160k prefix:
c32 hit tokens 40.6% -> ~100% (with HiCache L2/L3 enabled), needle test
(cold vs hit) identical.
The ctypes-based probing loaded a second libcudart copy into the process,
which corrupted CUDA/runtime state and segfaulted the scheduler processes
shortly after registration (reproducible on prefill ranks, intermittent on
decode). cuda.bindings.runtime.cudaPointerGetAttributes is a properly
typed binding and boots cleanly on both roles.
IntraNodeNvlinkTransport can only register/reach device memory. sglang's
register list includes host regions (aux buffers, some state components);
one host region failed the whole registerLocalMemoryBatch and the engine
rolled back every region, leaving the segment descriptor empty so all KV
transfers failed with 'Requested address ... not found'.
When the intra-node NVLink transport is active (MC_INTRANODE_NVLINK etc.):
- register only device-memory regions (probe via cudaPointerGetAttributes)
- route transfer blocks whose local source is host memory over the ordered
per-endpoint zmq channel (new STATE_DATA message) instead of the engine;
the receiver validates and writes the block into its buffer, relying on
the same ordering guarantee as the aux TCP path
- force aux over the TCP socket path automatically