[refactor] Fold FrozenKVMTPCudaGraphRunner onto the shared DecodeCudaGraphRunner base - #28081
Merged
Merged
Conversation
ch-wan
requested review from
Qiaolin-Yu,
hnyls2002,
kpham-sgl and
pyc96
as code owners
June 12, 2026 19:17
Contributor
|
Warning You have reached your daily quota limit. Please wait up to 24 hours and I will start processing your requests again! |
Collaborator
Author
|
/tag-and-rerun-ci |
Collaborator
Author
|
/rerun-test test_frozen_kv_mtp.py test_resolve_swa_kv_pool.py |
Contributor
|
Results for 🚀 |
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 504badc883
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
| @@ -174,53 +205,24 @@ def can_run(self, forward_batch: ForwardBatch): | |||
| ) | |||
|
|
|||
| is_bs_supported = ( | |||
| cuda_graph_bs in self.graphs | |||
| self.backend.can_run(forward_batch, cuda_graph_bs) | |||
There was a problem hiding this comment.
When --disable-cuda-graph-padding is enabled, this passes the raw integer cuda_graph_bs to backend.can_run, but this runner captures and replays graphs under ShapeKey(size=...) (see the nearby _make_graph_key usage in capture/replay). The full and breakable backends key their _graphs dictionaries by that exact object, so exact-size frozen-KV MTP batches that were captured will be reported unsupported and will fall back instead of using the CUDA graph.
Useful? React with 👍 / 👎.
Collaborator
Author
There was a problem hiding this comment.
Good catch — fixed in a2feafc. This was a real regression from the fold: pre-fold the runner kept self.graphs keyed by the int bs, so cuda_graph_bs in self.graphs worked under --disable-cuda-graph-padding. The fold moves graph storage to the backend, which keys _graphs by ShapeKey, so the raw int never matched and exact-size batches fell back. can_run now passes self._make_graph_key(cuda_graph_bs) (matching the BaseCudaGraphBackend.can_run(forward_batch, shape_key) contract and how capture/replay key the table), restoring the exact-size membership check. Verified by the FA4 frozen-KV MTP runner-mode unit test.
ch-wan
force-pushed
the
cheng/refactor/fold-frozen-kv-mtp-cuda-graph-runner
branch
from
June 12, 2026 20:52
504badc to
db0be2e
Compare
ch-wan
force-pushed
the
cheng/refactor/fold-frozen-kv-mtp-cuda-graph-runner
branch
from
June 12, 2026 20:56
db0be2e to
a2feafc
Compare
ch-wan
force-pushed
the
cheng/refactor/fold-frozen-kv-mtp-cuda-graph-runner
branch
2 times, most recently
from
June 12, 2026 22:32
f911a2b to
ac878d1
Compare
ch-wan
requested review from
BBuf,
Edwardf0t1,
Fridge003,
HaiShaw,
Ying1123,
alphabetc1,
hanming-lu,
hebiao064,
hzh0425,
ispobock,
merrymercy,
xiezhq-hermann,
yizhang2077 and
yuan-luo
as code owners
June 12, 2026 22:32
ch-wan
removed request for
BBuf,
ByronHsu,
CatherineSue,
Edwardf0t1,
Fridge003,
HaiShaw,
JustinTong0323,
ShangmingCai,
Ying1123,
alphabetc1,
hanming-lu,
hebiao064,
hzh0425,
ispobock,
merrymercy,
slin1237,
sogalin,
wisclmy0611,
xiezhq-hermann,
yizhang2077,
yuan-luo and
zijiexia
June 12, 2026 22:39
Collaborator
Author
|
/rerun-test registered/attention/unittests/* registered/spec/* |
Contributor
|
Results for 🚀 🚀 🚀 🚀 🚀 🚀 🚀 |
kpham-sgl
added a commit
that referenced
this pull request
Jun 22, 2026
Resolve conflicts from spec v2 refactor: - frozen_kv_mtp_worker.py: removed (replaced by frozen_kv_mtp_worker_v2.py in #27607); ported trtllm_mha draft-attn-backend branch into the v2 worker. - frozen_kv_mtp_cuda_graph_runner.py: adopt shared DecodeCudaGraphRunner surface (_replay_graph + _make_graph_key) from #28081 / #28384. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Chronostasys
pushed a commit
to MindLab-Research/sglang
that referenced
this pull request
Aug 24, 2026
jakki-amd
pushed a commit
to jakki-amd/sglang
that referenced
this pull request
Sep 9, 2026
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Motivation
FrozenKVMTPCudaGraphRunnerwas the last speculative cuda-graph runner still standalone on the legacy rawtorch.cuda.CUDAGraph/self.graphssurface. It hand-rolled the capture loop, bucket padding, and graph-pool management thatDecodeCudaGraphRunnerand therunner_backend/backends now own — the same machinery the EAGLE draft runners already share after the cuda-graph runner/backend refactor (#23906). This left the frozen-KV MTP draft loop as the lone duplicate of that scaffolding.Fold it onto
DecodeCudaGraphRunner, mirroringEAGLEDraftCudaGraphRunner, so the frozen-KV MTP draft step reuses the shared capture/replay path and drops its bespoke raw-graph code.Modifications
FrozenKVMTPCudaGraphRunnernow subclassesDecodeCudaGraphRunner(nosuper().__init__; it sets the parent's capture-contract fields directly and disables the inapplicable decode paths viacompile_bs=[]/enable_pdmux=False/record_nolora_graph=False/is_dllm=False).self.backend = resolve_decode_backend(self):capture_one_batch_size→capture_one_shape(size, forward, stream_idx, variant_label), driven by the inherited_capture_one_streamloop; it callsbackend.capture_one(which owns the two warmup passes +post_warmup_hook).replaycallsbackend.replayvia_replay_graph;_make_graph_keyreturnsShapeKey(size=bs).capture()/_capture_one_stream()and_pad_to_bucket()from the base; removes the hand-rolled_create_graph/_capture_init/_capture_graph/_replay, theself.graphs/self.output_buffersdicts, thebisectpadding, and the legacy graph-pool imports.topk*topkbucket divisor incan_run/replay, the expanded-bs (request_bs * topk) bookkeeping, the worker's_init_frozen_kv_metadata_{capture,replay}_cuda_graphhelpers, theFrozenKVMTPDraftInputspec_info, and the 3-tuple replay output.Behavior-preserving by construction; net +22 lines (legacy machinery removed, backend wiring added).
Accuracy Tests
Frozen-KV MTP draft acceptance is covered by
test/registered/spec/test_frozen_kv_mtp.py(GSM8K + average spec accept length ongoogle/gemma-4-E4B-it, CUDA graph enabled, topk 1 and 3). Local structural checks: the module compiles, imports as a non-abstractDecodeCudaGraphRunnersubclass with all runner methods resolving (capture()inherited;capture_one_shape/replay/can_run/_make_graph_key/_replay_graphoverridden), and no legacy raw-CUDAGraphsurface remains.Speed Tests and Profiling
Negligible. Capture runs once at startup; the steady-state replay path is the same backend graph replay as before — same captured graph, same buffers, same per-step draft logic. No kernel or hot-path code changed.
Checklist
test_frozen_kv_mtp.py).CI States
Latest PR Test (Base): 🚫 Run #27447122392
Latest PR Test (Extra): ❌ Run #27447122201