config: the speculative workers take page_size from the bags
Seven worker constructors stored `self.page_size = server_args.page_size` off
the handed record. They all run after publish and all keep a copy of a
process-level value, which is the first row of the plan doc's supplied-instance
disposition table -- so they read `get_schedule().page_size`, and a post-publish
override now reaches them like it reaches every other consumer.
The supplied-instance census named the seven pairs; the exposure ratchet in the
next member pins what remains after this batch of conversions.
config: the post-publish chunked_prefill_size consumers read the bags
Four of the ten supplied-instance `chunked_prefill_size` reads are plain
post-publish consumers -- the EPLB recorder's buffer sizing, the deep-gemm
compile warmup (five reads), the KV-cache builder's effective size, and the
ngram embedding manager's assert. All are reached from runner init, so they read
`get_schedule()`.
Two are deliberately left: `create_kt_config_from_server_args` builds a config
*from a supplied record* by name and contract, and `CanaryLaunchCapacities.from_args`
is the same shape. Converting those would change what the function is, not where
it reads -- the plan doc's disposition table says so per field.
config: the remaining post-publish graph/limit consumers read the bags
Three more of the census's supplied-instance debts are plain post-publish reads: the dspark worker's
cuda-graph decode sizes, the dspark planner's SPS table bound
(`max_running_requests`), and the LoRA manager's cuda-graph moe buffers. The
dspark worker is the clearest of them -- it already read
`get_exec().graph.cuda_graph_config.decode.bs` thirty lines below the instance
read, so the file disagreed with itself about where the same value comes from.
Left where the function's contract is "build a config from the record you are
handed" rather than "read this process's config":
`create_kt_config_from_server_args`, `DllmConfig.from_server_args`,
`CanaryLaunchCapacities.from_args`, `build_compilation_config`. Changing those
would change what the function is.
config: the runner, scheduler and offload manager take page_size from the bags
The same `self.page_size = server_args.page_size` shape as the speculative
workers, in the three remaining process-owned constructors: `ModelRunner`,
`Scheduler`, and the decode-side KV offload manager. The scheduler process
publishes before any of them run. The one path that did not is `ModelRunner`
constructed standalone -- `python -m sglang.benchmark.one_batch` and the manual
runner tests build it with no prior publish, and the constructor's own publish
sat below this read -- so that publish moves above the constructor's first bag
read instead of leaving a window where the runner half-exists unpublished.
Left where the read belongs to something else: `utils/common`'s predicates are
called only from the resolution pipeline with a `resolved_view`,
`allocation_sizing` takes the config its callers supply by contract, and
`CudaVmmFeatureTransport` is tokenizer-owned -- one per tokenizer worker, which
is the per-instance boundary.
The conversion left the offload manager parking a record it no longer
reads; the parked copy goes with the read (the constructor parameter stays
-- its hicache sizing still reads it directly).
The previous batch counted `self.server_args.X` and called the runner surface
done. It was not: the same read spelled through a local alias --
`server_args = model_runner.server_args` (or `sa = kvc.server_args`, `args = ...`)
followed by `server_args.leaf` -- is the same process-global read wearing a
local name, and the AST census counts **57 of them** across eleven files that
the grep never saw. Census per function, following the alias.
52 were leaves and go to their bag (`spec` 11, `schedule` 9, `memory` 7,
`exec.graph` 5, `exec.moe` 5, `parallel` 4, `disagg` 4, `model` 3,
`exec.mamba` 2, `exec.overlap` 2). Five were not leaves:
three derived members on the eager runner --
`max_speculative_num_draft_tokens` and `enable_mamba_extra_buffer` already had
accessors, and `max_prefill_buffer_tokens` gets one (all its inputs are `schedule`
leaves plus the configured PP size, so it derives from the bags and follows a
post-publish override; `TestDerivedPredicatesAgreeAcrossTiers` pins it against
the member over a 48-case matrix) -- plus `get_attention_backends()`, which the
same commit routes through `attention_backends()`, and a dict that merely shares
the name (`server_args_dict.items`). That dict is the one read left behind.
`build_attention_backends` also stops resolving the pair from the record: it
runs after publish, so it asks `attention_backends()` like every other consumer.
The draft override on the runner still wins first.
`dispatch_event_loop`'s three PP checks read the *configured* PP size, not the
live topology: the MLX runner stub never initializes torch.distributed, so the
live property asserts before the MLX event loop can start (a Codex catch). The
configured leaf answers the same value wherever the live groups exist.
`flashinfer_gdn_prefill_default`'s guard is the one read here that asks what the
*operator* named rather than what the config resolved to, and the bag leaf now
answers exactly that: the per-runner auto-default is stamped on the runner and
deliberately never recorded process-wide, so nothing writes that leaf after
launch and reading it back cannot mistake another runner's default for a flag.
Three test doubles injected a `SimpleNamespace`/`MagicMock` record for exactly
these reads and now publish instead (pool configurator, cache registry, GDN
prefill policy) -- the fixture publishes what the case configures and hands the
published instance to the whole-object contracts that still take one.
The functions this sweep partially converted stop mixing sources (review
catches): the flash-attention constructor's remaining seed reads
(`speculative_eagle_topk`, `speculative_algorithm`, both deterministic gates)
read their bags next to the leaves already converted;
`_should_disable_scheduler_metadata_precompute` reads the parallel config
leaves itself instead of taking the record (its alias binding was the last
use); and the autotune gates (`disable_flashinfer_autotune`, deterministic,
`flashinfer_autotune_skip_ops`) join the moe leaves the same function already
reads from the bags. The pool-configurator fixture drops a parameter nothing
published or read.
`_is_dsa_active` asked `getattr(server_args, "_is_dsa_model_arch", False)`, and
that name has never existed on `ServerArgs` -- it arrived as a placeholder with
the CP strategy abstractions (#27313), so the getattr default has always decided
the predicate. A dynamic read of a name nothing sets is the one shape the config
census cannot follow, and it looked like a live decision while being dead.
Spelled as the constant it evaluates to, with the placeholder written down: what
it should ask (whether this process runs a DSA model arch) is the CP path's
call, and its only consumer, `ContextParallelStrategy.per_layer_attn_cp_comm`,
has no readers yet.
That was the sole entry in the read ratchet's `_INERT_DYNAMIC_READS`, so the
exemption list is gone with it -- there is no way to exempt a read from the
baselines any more, which is the invariant worth having. The `counted()`
indirection it existed for goes too (verified the three shapes it guarded still
report: direct, `getattr`, and an attribute-parked alias).
`--attention-backend` is one field of three: a launch that sets only
`--prefill-attention-backend` or `--decode-attention-backend` leaves the base
field at `None`. Seven decisions read that base field alone and therefore
answered from a field the operator never set. `attention_backends()` is the
pair with the base-field fallback already applied, so each site now asks it for
the half it actually needs:
- `inkling_common/attn` assembles backend-specific kwargs (rel_bias / score
mods) and gates its fused prologue; the backend those describe is the one
`self.attn` dispatches to, so `serving_attention_backend()` selects the pair
member by `forward_batch.forward_mode`, mirroring
`HybridAttnBackend._select_backend` exactly -- draft-extend routes through
the prefill branch like the dispatcher does -- and preferring the
runner-stamped pair, so a draft runner answers with its own backend. That
preference only works if every backend that can enter a ForwardContext
carries the stamp, so `DraftBackendFactory._create_backend` now stamps its
products with the backend it resolved (draft override first), and the
draft-extend conv-sidecar wrapper copies the wrapped backend's stamp -- the
replacement backends the spec workers install had no stamp at all and fell
back to the target's configured pair.
- The chunked-prefix-cache gate is a *prefill* feature -> prefill half. Reading
the base field switched the feature off for every prefill-only configuration.
- `init_deterministic_inference_config` maps *prefill* knobs
(SPLIT_TILE / PREFILL_TRUNCATION_ALIGN) -> prefill half; the map missed and
left truncation unset.
- `two_batch_overlap` computes extend positions -> prefill half.
- mrope's interleaved-rope kernel runs in both phases -> both halves must
support triton. This one is not conservative when it misreads:
`support_triton(None)` answers **True**, so a `--prefill-attention-backend
torch_native` launch took the triton path.
- The req-to-token writer has one caller, `alloc_for_extend` -> prefill half;
its fallback pays several `.item()` syncs per request, so gating it on the
decode half too would send every extend of a mixed launch through the slow
path. `get_last_loc` (the spec-decode allocator's helper) keeps the
both-halves reading: verify tokens are served by either half depending on
`speculative_attention_mode`.
- The flashinfer version floor is a guard; it never fired for a launch that
pinned flashinfer through a split field.
One more site the census found is not converted here: `gpt_oss` derives its
`sinks` parameter dtype from the backend, and a single parameter dtype cannot
serve a split pair (FA4 asserts bfloat16, trtllm_mha consumes float32), so
that one is a behaviour question rather than a config-source one and is fixed
in its own PR.
`test_split_attention_backend_decisions.py` pins the callable decisions by
calling them under a split-only publish, and pins the remaining ones
statically -- the file/why map fails if any of them goes back to the base field
(reverse-verified). It also asserts the `support_triton(None) is True` trap the
sweep exists for.
The stamp comes from the constructor, not the request: every factory leaf
answers ("effective_name", backend), because several map entries do not build
what their key says -- cutedsl_mla draft-extend builds the trtllm-mla backend,
"nsa" is a deprecated alias building dsa, and the hybrid-linear entries pick
fa3/intel_amx/triton by host, which no static rename table can express (a
review catch: on Blackwell the alias stamp reached Inkling's per-forward
kwargs assembly, which asserts a concrete kernel name, and crashed the first
draft-extend forward). The stamping is pinned by unit tests, not only by a
spec e2e: removing the child-stamping loop, stamping an alias from a leaf, or
dropping the wrapper copy goes red (reverse-verified), and a static guard
walks the factory source asserting no leaf answers an alias name. The child loop states its contract explicitly --
`create_decode_backend` passes `stamps_children=True` because its products
are per-step containers by construction, so a container without
`attn_backends` raises instead of being silently skipped by a defensive
probe. The `_version` invalidation names its contract (autograd's in-place
counter: private, chosen because it is the only per-tensor signal that ticks
on copy_-style updates; removal fails loudly). The version-floor guard's file
joins the pair-reader ratchet, and the one runner-seed chain read sharing the
backend's __init__ (`speculative_eagle_topk`) reads the spec bag.
Six reads were left on `self.server_args` outside the per-instance boundary the
plan reserves for the tokenizer-manager family, and each had a different reason
to be there:
- `scheduler.process_input_requests` (`mm_feature_transport`) and
`BaseSpecWorker._build_hicache_draft_plan` (`enable_hierarchical_cache`) are
plain leaves -> `get_mm()` / `get_memory()`.
- `DraftBackendFactory._create_backend` read the split backend through a
*runtime-computed name* (`getattr(self.server_args, backend_name)`) and then
fell back to the base field by hand -- the census's documented blind spot.
The two names it can be handed are exactly the pair `attention_backends()`
returns with that fallback already applied, so it reads the pair and indexes
it. The draft runner's own stamp still wins when it has one.
- `remote_instance_weight_loader_use_transfer_engine` and
`pre_capture_activation_reserve_mb` are derived members. Both are computed
from published leaves only, so both get a named accessor that derives from
the bags (and therefore follows a post-publish override).
The first of those two has all its inputs in one bag, so it follows the
established shape: one `*_of(cfg)` helper in `arg_groups/overrides.py`, the
`ServerArgs` member delegating to it, and the accessor calling it on
`get_model()`. `modelexpress_transport_of` splits out the JSON parse both
sides need. The second spans four bags plus the configured parallel sizes, so
it exists twice like the mamba pair -- and `TestDerivedPredicatesAgreeAcrossTiers`
now pins both new pairs equal over their input matrices (92 subtests).
`self.server_args.X` outside the tokenizer-manager family: 11 -> 5, and the
five that remain are the documented ones (the encode server's own record, the
nixl connector's rank arithmetic, `GrammarManager`'s handed instance).
The post-capture headroom path calls the same bag-backed
`pre_capture_activation_reserve_mb` accessor the configurator uses -- the
accessor advertises override-following, and a reserve that reads the record
while its sibling reads the bags can disagree after a post-publish override.
And the conversions' orphans go with them: `RemoteInstanceWeightTransporter`
kept a `server_args` field nothing reads, and `DraftBackendFactory` parked a
record it no longer consults -- both drop the parameter, and the four factory
call sites stop threading one.
Two problems in the same family as #33312 (a per-runner decision that one
participant answered differently), one fixed here and one guarded.
**The linear-attn kernel backends were process-wide.** `attn_backend_wrapper`
rebuilt a module-level dict once per runner, from the handed record plus a local
`prefill_default`. Two things follow, and both are wrong:
- **A draft could not hold a different choice than its target.** Only the runner
whose model is GDN gets the SM100 FlashInfer prefill default; the operator's
explicit flag belongs to the launch. The full-attention backends already model
this correctly -- the runner stamps `prefill_attention_backend_str` /
`decode_attention_backend_str` and its backend objects are built from the
stamp. Linear attn had no stamp at all.
- **The second rebuild replaced the first one's choice.** The default was also
recorded into the process-wide config, which the record does not see, so a
runner rebuilding without a default of its own resolved `prefill` back to the
base backend -- silently swapping the kernel the earlier runner selected.
Demonstrated in-process before this change: table `FLASHINFER`, then `TRITON`.
`resolve_linear_attn_backends(prefill_default=None)` returns a frozen
`LinearAttnBackends(decode, prefill, verify)` from the published `exec.mamba`
leaves; the wrapper stamps it as `runner.linear_attn_backends` before building
the backends that read it; and the three consumers (GDN, KDA, Ascend GDN) read it
off the runner they are built for. Each already took `model_runner` and cached
the result on itself, so the value now simply comes from the right place. A
backend built outside that path has no stamp and raises on the attribute, the way
the full-attention strings do -- no silent fallback to hide the wiring mistake.
The recording goes away with it. A per-runner choice in the process-wide config
has no meaning the second runner can read correctly: the leaf is how the gate
asks "did the operator name a backend", so a recorded default reads back as an
operator flag and the next runner declines its own. The leaf now keeps meaning
what was asked for at launch, and the effective choice lives in the stamp (and
in the log line the gate already emits).
Precedence is unchanged: the resolver takes the default as an argument and an
explicit `--linear-attn-prefill-backend` wins over it, with the gate declining
early so it neither probes the device nor logs.
**A draft entry class must answer the loader exactly when its target does.** The
loader asks the entry class it instantiates for the shared-experts-fusion
decision, and a draft is its own entry class. When the target family carries
auto-disable conditions and the draft's class does not expose them, the loader
installs one decision for each and the draft's weights are laid out for the wrong
one. That shipped: the DSV4 DSpark draft skipped its bundled shared-expert
tensors until #33312 gave it the gate, costing accept length 5.60 -> 2.05.
`test_fusion_gate_coverage.py` walks the same registry but asks whether an entry
class *touches* the decision -- reads the flag, names a gated class. That catches
a class once it already consumes the decision; it could not catch one that should
consume it and does not, which is what the DSpark class looked like (it built the
family's *layer* classes, so the flag reader lived in another module and its own
source named no gated class). `test_draft_entry_hook_parity.py` asks the
invariant directly: presence parity between a draft entry class and the target it
is named after. Identity is deliberately not required -- the Qwen3.5 MTP
delegates with adapted arguments (unwrapping `text_config`, using the MTP
quantization config), which is right -- and weight-name maps are out of scope,
since a draft's checkpoint has its own names.
Reverse-verified against the original defect: with the DSpark gate removed the
case names the pair and the side that is missing it; with #33312 in place it
passes.
Shenxiu Liu