[TRTLLM-8521][chore] remove circular dependency between model engine and cuda graph runner #7572
Conversation
QiJune
requested review from
Wanli-Jiang,
amukkara,
byshiue,
danielafrimi and
hypdeb
|
/bot run |
|
PR_Github #17812 [ run ] triggered by Bot |
📝 WalkthroughWalkthroughRefactors CUDAGraphRunner to a dependency-injected, engine-agnostic component with expanded APIs carrying speculative-decoding context and metadata. Updates ModelEngine to wire new parameters and draft-token CUDA buffers. Migrates unit tests to a new create_mock_cuda_graph_runner helper and adapts capture/replay call signatures with an added boolean flag. Changes
Sequence Diagram(s)sequenceDiagram
autonumber
participant ME as ModelEngine
participant GR as CUDAGraphRunner
participant FW as forward_fn
participant Attn as AttentionMetadata
participant Spec as SpecMetadata
Note over ME,GR: New flow passes is_spec_decode, metadata, draft tokens
ME->>GR: maybe_get_cuda_graph(batch, iter_counter, is_spec_decode, Attn, Spec?, draft_tokens_cuda)
GR-->>ME: (can_use_graph, Attn', Spec')
alt needs capture
ME->>GR: needs_capture(batch_size, is_spec_decode)
GR-->>ME: bool
alt true
ME->>GR: capture(batch_size, is_spec_decode, FW, initial_inputs)
GR->>FW: forward(**initial_inputs) during capture
FW-->>GR: outputs (captured)
GR-->>ME: capture complete
end
end
alt can_use_graph
ME->>GR: replay(batch_size, is_spec_decode, current_inputs)
GR->>GR: update static tensors, set draft_len from Spec'
GR-->>ME: logits (replayed)
else not eligible
ME->>FW: forward(**current_inputs)
FW-->>ME: logits
end
Estimated code review effort🎯 4 (Complex) | ⏱️ ~60 minutes Possibly related PRs
Suggested labels
Suggested reviewers
✨ Finishing Touches
🧪 Generate unit tests
Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out. 🪧 TipsChatThere are 3 ways to chat with CodeRabbit:
SupportNeed help? Create a ticket on our support page for assistance with any issues or questions. CodeRabbit Commands (Invoked using PR/Issue comments)Type Other keywords and placeholders
Status, Documentation and Community
|
![coderabbitai[bot]](https://avatars.githubusercontent.com/in/347564?s=60&v=4){kind=small}
There was a problem hiding this comment.
Actionable comments posted: 12
Caution
Some comments are outside the diff and can’t be posted inline due to platform limitations.
⚠️ Outside diff range comments (8)
tests/unittest/_torch/modeling/test_modeling_llama_min_latency.py (1)
416-423: Fix int64→int32 dtype mismatch in CUDA-graph path (will crash at replay).cuda_graph_runner allocates static input_ids/position_ids as int32; current inputs use default int (arange → int64). torch.Tensor.copy_ requires matching dtypes and will error at replay. Cast both to int32 before capture/replay.
Apply this diff:
- inputs = { - "input_ids": input_ids, - "position_ids": position_ids, - "attn_metadata": attn_metadata, - } + inputs = { + "input_ids": input_ids.to(torch.int32), + "position_ids": position_ids.to(torch.int32), + "attn_metadata": attn_metadata, + }Also applies to: 429-429
tests/unittest/_torch/modeling/test_modeling_phi3.py (1)
322-329: Align dtypes with CUDA-graph static buffers (int32).position_ids from arange default to int64; runner uses int32 buffers, causing copy_ dtype mismatch at replay. Cast inputs before capture/replay.
- inputs = { - "input_ids": input_ids, - "position_ids": position_ids, - "attn_metadata": attn_metadata, - } + inputs = { + "input_ids": input_ids.to(torch.int32), + "position_ids": position_ids.to(torch.int32), + "attn_metadata": attn_metadata, + }Also applies to: 335-335
tests/unittest/_torch/modeling/test_modeling_mllama.py (1)
429-436: Prevent dtype mismatch with CUDA-graph int32 inputs.Ensure input_ids/position_ids are int32 before capture/replay to match runner’s static tensors.
- inputs = { - "input_ids": input_ids, - "position_ids": position_ids, - "attn_metadata": attn_metadata, - } + inputs = { + "input_ids": input_ids.to(torch.int32), + "position_ids": position_ids.to(torch.int32), + "attn_metadata": attn_metadata, + }Also applies to: 442-442
tests/unittest/_torch/modeling/test_modeling_qwen_moe.py (1)
327-334: Cast position_ids to int32 for CUDA-graph replay.Runner’s static tensors are int32; arange yields int64. Cast before capture to avoid copy_ errors.
- inputs = { - "input_ids": input_ids, - "position_ids": position_ids, - "attn_metadata": attn_metadata, - } + inputs = { + "input_ids": input_ids.to(torch.int32), + "position_ids": position_ids.to(torch.int32), + "attn_metadata": attn_metadata, + }Also applies to: 340-340
tests/unittest/_torch/modeling/test_modeling_mixtral.py (1)
322-329: Match int32 expectations in CUDA-graph path.Cast inputs to int32 to align with cuda_graph_runner’s buffers.
- inputs = { - "input_ids": input_ids, - "position_ids": position_ids, - "attn_metadata": attn_metadata, - } + inputs = { + "input_ids": input_ids.to(torch.int32), + "position_ids": position_ids.to(torch.int32), + "attn_metadata": attn_metadata, + }Also applies to: 335-335
tests/unittest/_torch/modeling/test_modeling_qwen.py (1)
85-93: Python 3.8 compatibility: avoid built-in genericsdict[str, Any] requires Python 3.9+. Tests target 3.8+. Use typing.Dict.
-from typing import Any +from typing import Any, Dict @@ -def reduce_qwen_config(mem_for_full_model: int, config_dict: dict[str, Any]): +def reduce_qwen_config(mem_for_full_model: int, config_dict: Dict[str, Any]):tensorrt_llm/_torch/pyexecutor/model_engine.py (1)
400-406: Fix dtype of index tensors (must be torch.long).
gather_ids_cudaandprevious_pos_indices_cudaare later used for tensor indexing (e.g.,logits[gather_ids]), which requiresLongTensorindices in PyTorch. Usingtorch.int(int32) risks runtime errors.Apply:
- self.gather_ids_cuda = torch.empty((self.max_num_tokens, ), - dtype=torch.int, - device='cuda') - self.previous_pos_indices_cuda = torch.empty( - (self.max_num_tokens, ), dtype=torch.int, device='cuda') + self.gather_ids_cuda = torch.empty((self.max_num_tokens, ), + dtype=torch.long, + device='cuda') + self.previous_pos_indices_cuda = torch.empty( + (self.max_num_tokens, ), dtype=torch.long, device='cuda')Note:
previous_batch_indices_cuda(Line 444) is also used as an index and should betorch.longfor consistency. See additional snippet below.tensorrt_llm/_torch/pyexecutor/cuda_graph_runner.py (1)
225-225: Inconsistent return value from replay methodThe
replaymethod returnsoutput_refwhich is a callable (weak reference), but the return type annotation suggests it should returnOptional[torch.Tensor]. The method should either call the reference or update the return type.Call the weak reference to get the actual tensor:
- return output_ref + return output_ref() if output_ref else None
🧹 Nitpick comments (11)
tests/unittest/_torch/modeling/test_modeling_llama_min_latency.py (1)
406-408: Release CUDA graph resources after test.Call graph_runner.clear() after use to free the graph’s memory pool.
Example (near test end):
if graph_runner is not None: graph_runner.clear()tests/unittest/_torch/modeling/test_modeling_phi3.py (1)
312-314: Exercise the CUDA-graph path in this test.scenario.use_cuda_graph is never set to True here, so the CUDA-graph code isn’t exercised. Consider parameterizing to run both paths.
tests/unittest/_torch/helpers.py (1)
171-186: Helper factory looks good; add return type and optional knobsAnnotate the return type and expose use_mrope/enable_attention_dp to avoid future test drift.
-def create_mock_cuda_graph_runner(batch_size: int): +def create_mock_cuda_graph_runner(batch_size: int) -> CUDAGraphRunner: return CUDAGraphRunner( use_cuda_graph=True, cuda_graph_padding_enabled=False, supported_batch_sizes=[batch_size], max_supported_batch_size=batch_size, max_batch_size=batch_size, max_beam_width=1, max_draft_len=0, use_mrope=False, spec_config=None, cuda_graph_mem_pool=None, enable_attention_dp=False, mapping=Mapping(), dist=None, kv_cache_manager_key=ResourceManagerType.KV_CACHE_MANAGER)Optionally:
-def create_mock_cuda_graph_runner(batch_size: int) -> CUDAGraphRunner: +def create_mock_cuda_graph_runner( + batch_size: int, + *, + use_mrope: bool = False, + enable_attention_dp: bool = False, +) -> CUDAGraphRunner: @@ - use_mrope=False, + use_mrope=use_mrope, @@ - enable_attention_dp=False, - mapping=Mapping(), + enable_attention_dp=enable_attention_dp, + mapping=Mapping(enable_attention_dp=enable_attention_dp),tests/unittest/_torch/modeling/test_modeling_exaone4.py (1)
25-25: Fix E402: keep imports at top-of-fileMove this import to the top with the other imports to satisfy Ruff E402.
+from _torch.helpers import create_mock_cuda_graph_runner @@ -from _torch.helpers import create_mock_cuda_graph_runnertests/unittest/_torch/modeling/test_modeling_nemotron.py (2)
320-321: Minor: avoid magic number for batch size in tests.Consider a local
graph_bs = 1to avoid repeating the literal and ease future edits.
343-344: Optional: assert non-None replay output.
replayreturns an Optional (weak-ref). Add a quickassert logits is not Nonefor clearer failures.tests/unittest/_torch/modeling/test_modeling_mistral.py (2)
402-402: Minor: avoid magic number for batch size.Define a local
graph_bs = 1and reuse it in capture/replay calls.
422-423: Optional: guard against None from replay.Add
assert logits is not Nonebefore comparisons to make weak-ref issues obvious.tests/unittest/_torch/modeling/test_modeling_llama.py (2)
328-329: Minor: avoid magic number for batch size.Use a local
graph_bs = 1and reuse it.
350-351: Optional: assert non-None replay output.Add
assert logits is not Noneto surface weak-ref invalidation early.tensorrt_llm/_torch/pyexecutor/model_engine.py (1)
412-413: Use instance attribute for clarity.Prefer
self.spec_config.max_draft_lenover the shadowedspec_config.- self.max_draft_len = spec_config.max_draft_len + self.max_draft_len = self.spec_config.max_draft_len
📜 Review details
Configuration used: Path: .coderabbit.yaml
Review profile: CHILL
Plan: Pro
💡 Knowledge Base configuration:
- MCP integration is disabled by default for public repositories
- Jira integration is disabled by default for public repositories
- Linear integration is disabled by default for public repositories
You can enable these sources in your CodeRabbit configuration.
📒 Files selected for processing (13)
tensorrt_llm/_torch/pyexecutor/cuda_graph_runner.py(10 hunks)tensorrt_llm/_torch/pyexecutor/model_engine.py(5 hunks)tests/unittest/_torch/helpers.py(2 hunks)tests/unittest/_torch/modeling/test_modeling_exaone4.py(3 hunks)tests/unittest/_torch/modeling/test_modeling_llama.py(3 hunks)tests/unittest/_torch/modeling/test_modeling_llama_min_latency.py(3 hunks)tests/unittest/_torch/modeling/test_modeling_mistral.py(3 hunks)tests/unittest/_torch/modeling/test_modeling_mixtral.py(3 hunks)tests/unittest/_torch/modeling/test_modeling_mllama.py(3 hunks)tests/unittest/_torch/modeling/test_modeling_nemotron.py(3 hunks)tests/unittest/_torch/modeling/test_modeling_phi3.py(3 hunks)tests/unittest/_torch/modeling/test_modeling_qwen.py(3 hunks)tests/unittest/_torch/modeling/test_modeling_qwen_moe.py(3 hunks)
🧰 Additional context used
📓 Path-based instructions (3)
**/*.{h,hpp,hh,hxx,cpp,cxx,cc,cu,cuh,py}
📄 CodeRabbit inference engine (CODING_GUIDELINES.md)
Use only spaces, no tabs; indent with 4 spaces.
Files:
tests/unittest/_torch/modeling/test_modeling_mllama.pytests/unittest/_torch/modeling/test_modeling_llama_min_latency.pytests/unittest/_torch/modeling/test_modeling_exaone4.pytests/unittest/_torch/helpers.pytests/unittest/_torch/modeling/test_modeling_llama.pytensorrt_llm/_torch/pyexecutor/model_engine.pytests/unittest/_torch/modeling/test_modeling_mistral.pytests/unittest/_torch/modeling/test_modeling_mixtral.pytests/unittest/_torch/modeling/test_modeling_phi3.pytests/unittest/_torch/modeling/test_modeling_qwen_moe.pytests/unittest/_torch/modeling/test_modeling_nemotron.pytests/unittest/_torch/modeling/test_modeling_qwen.pytensorrt_llm/_torch/pyexecutor/cuda_graph_runner.py
**/*.py
📄 CodeRabbit inference engine (CODING_GUIDELINES.md)
**/*.py: Python code must target Python 3.8+.
Indent Python code with 4 spaces; do not use tabs.
Maintain module namespace when importing; prefer 'from package.subpackage import foo' then 'foo.SomeClass()' instead of importing the class directly.
Python filenames should be snake_case (e.g., some_file.py).
Python classes use PascalCase names.
Functions and methods use snake_case names.
Local variables use snake_case; prefix 'k' for variables that start with a number (e.g., k_99th_percentile).
Global variables use upper SNAKE_CASE prefixed with 'G' (e.g., G_MY_GLOBAL).
Constants use upper SNAKE_CASE (e.g., MY_CONSTANT).
Avoid shadowing variables from an outer scope.
Initialize all externally visible members of a class in the constructor.
Prefer docstrings for interfaces that may be used outside a file; comments for in-function or file-local interfaces.
Use Google-style docstrings for classes and functions (Sphinx-parsable).
Document attributes and variables inline so they render under the class/function docstring.
Avoid reflection when a simpler, explicit approach suffices (e.g., avoid dict(**locals()) patterns).
In try/except, catch the most specific exceptions possible.
For duck-typing try/except, keep the try body minimal and use else for the main logic.
Files:
tests/unittest/_torch/modeling/test_modeling_mllama.pytests/unittest/_torch/modeling/test_modeling_llama_min_latency.pytests/unittest/_torch/modeling/test_modeling_exaone4.pytests/unittest/_torch/helpers.pytests/unittest/_torch/modeling/test_modeling_llama.pytensorrt_llm/_torch/pyexecutor/model_engine.pytests/unittest/_torch/modeling/test_modeling_mistral.pytests/unittest/_torch/modeling/test_modeling_mixtral.pytests/unittest/_torch/modeling/test_modeling_phi3.pytests/unittest/_torch/modeling/test_modeling_qwen_moe.pytests/unittest/_torch/modeling/test_modeling_nemotron.pytests/unittest/_torch/modeling/test_modeling_qwen.pytensorrt_llm/_torch/pyexecutor/cuda_graph_runner.py
**/*.{cpp,cxx,cc,h,hpp,hh,hxx,cu,cuh,py}
📄 CodeRabbit inference engine (CODING_GUIDELINES.md)
Prepend the NVIDIA Apache-2.0 copyright header with current year to the top of all source files (e.g., .cpp, .h, .cu, .py).
Files:
tests/unittest/_torch/modeling/test_modeling_mllama.pytests/unittest/_torch/modeling/test_modeling_llama_min_latency.pytests/unittest/_torch/modeling/test_modeling_exaone4.pytests/unittest/_torch/helpers.pytests/unittest/_torch/modeling/test_modeling_llama.pytensorrt_llm/_torch/pyexecutor/model_engine.pytests/unittest/_torch/modeling/test_modeling_mistral.pytests/unittest/_torch/modeling/test_modeling_mixtral.pytests/unittest/_torch/modeling/test_modeling_phi3.pytests/unittest/_torch/modeling/test_modeling_qwen_moe.pytests/unittest/_torch/modeling/test_modeling_nemotron.pytests/unittest/_torch/modeling/test_modeling_qwen.pytensorrt_llm/_torch/pyexecutor/cuda_graph_runner.py
🧠 Learnings (2)
📚 Learning: 2025-08-26T09:37:10.463Z
Learnt from: jiaganc
PR: NVIDIA/TensorRT-LLM#7031
File: tensorrt_llm/bench/dataclasses/configuration.py:90-104
Timestamp: 2025-08-26T09:37:10.463Z
Learning: In TensorRT-LLM, the `get_pytorch_perf_config()` method returns `self.pytorch_config` which can contain default `cuda_graph_config` values, so `llm_args` may already have this config before the extra options processing.
Applied to files:
tests/unittest/_torch/modeling/test_modeling_mllama.py
📚 Learning: 2025-08-26T09:37:10.463Z
Learnt from: jiaganc
PR: NVIDIA/TensorRT-LLM#7031
File: tensorrt_llm/bench/dataclasses/configuration.py:90-104
Timestamp: 2025-08-26T09:37:10.463Z
Learning: In TensorRT-LLM's bench configuration, the `get_pytorch_perf_config()` method returns `self.pytorch_config` which is a Dict[str, Any] that can contain default values including `cuda_graph_config`, making the fallback `llm_args["cuda_graph_config"]` safe to use.
Applied to files:
tests/unittest/_torch/modeling/test_modeling_mllama.py
🧬 Code graph analysis (13)
tests/unittest/_torch/modeling/test_modeling_mllama.py (2)
tests/unittest/_torch/helpers.py (1)
create_mock_cuda_graph_runner(171-186)tensorrt_llm/_torch/pyexecutor/cuda_graph_runner.py (2)
capture(143-194)replay(196-225)
tests/unittest/_torch/modeling/test_modeling_llama_min_latency.py (2)
tests/unittest/_torch/helpers.py (1)
create_mock_cuda_graph_runner(171-186)tensorrt_llm/_torch/pyexecutor/cuda_graph_runner.py (2)
capture(143-194)replay(196-225)
tests/unittest/_torch/modeling/test_modeling_exaone4.py (3)
tests/unittest/_torch/helpers.py (1)
create_mock_cuda_graph_runner(171-186)tensorrt_llm/_torch/pyexecutor/cuda_graph_runner.py (2)
capture(143-194)replay(196-225)tensorrt_llm/_torch/pyexecutor/model_engine.py (2)
forward(79-87)forward(2211-2313)
tests/unittest/_torch/helpers.py (3)
tensorrt_llm/_torch/pyexecutor/cuda_graph_runner.py (1)
CUDAGraphRunner(22-316)tensorrt_llm/_torch/pyexecutor/resource_manager.py (1)
ResourceManagerType(47-52)tensorrt_llm/mapping.py (1)
Mapping(32-513)
tests/unittest/_torch/modeling/test_modeling_llama.py (2)
tests/unittest/_torch/helpers.py (1)
create_mock_cuda_graph_runner(171-186)tensorrt_llm/_torch/pyexecutor/cuda_graph_runner.py (2)
capture(143-194)replay(196-225)
tensorrt_llm/_torch/pyexecutor/model_engine.py (1)
tensorrt_llm/_torch/pyexecutor/cuda_graph_runner.py (3)
CUDAGraphRunner(22-316)needs_capture(139-141)replay(196-225)
tests/unittest/_torch/modeling/test_modeling_mistral.py (2)
tests/unittest/_torch/helpers.py (1)
create_mock_cuda_graph_runner(171-186)tensorrt_llm/_torch/pyexecutor/cuda_graph_runner.py (2)
capture(143-194)replay(196-225)
tests/unittest/_torch/modeling/test_modeling_mixtral.py (2)
tests/unittest/_torch/helpers.py (1)
create_mock_cuda_graph_runner(171-186)tensorrt_llm/_torch/pyexecutor/cuda_graph_runner.py (2)
capture(143-194)replay(196-225)
tests/unittest/_torch/modeling/test_modeling_phi3.py (2)
tests/unittest/_torch/helpers.py (1)
create_mock_cuda_graph_runner(171-186)tensorrt_llm/_torch/pyexecutor/cuda_graph_runner.py (2)
capture(143-194)replay(196-225)
tests/unittest/_torch/modeling/test_modeling_qwen_moe.py (2)
tests/unittest/_torch/helpers.py (1)
create_mock_cuda_graph_runner(171-186)tensorrt_llm/_torch/pyexecutor/cuda_graph_runner.py (2)
capture(143-194)replay(196-225)
tests/unittest/_torch/modeling/test_modeling_nemotron.py (3)
tests/unittest/_torch/helpers.py (1)
create_mock_cuda_graph_runner(171-186)tensorrt_llm/_torch/pyexecutor/cuda_graph_runner.py (2)
capture(143-194)replay(196-225)tensorrt_llm/_torch/pyexecutor/model_engine.py (2)
forward(79-87)forward(2211-2313)
tests/unittest/_torch/modeling/test_modeling_qwen.py (3)
tests/unittest/_torch/helpers.py (1)
create_mock_cuda_graph_runner(171-186)tensorrt_llm/_torch/pyexecutor/cuda_graph_runner.py (2)
capture(143-194)replay(196-225)tensorrt_llm/_torch/pyexecutor/model_engine.py (2)
forward(79-87)forward(2211-2313)
tensorrt_llm/_torch/pyexecutor/cuda_graph_runner.py (8)
tensorrt_llm/mapping.py (1)
Mapping(32-513)tensorrt_llm/_torch/attention_backend/interface.py (2)
AttentionMetadata(39-347)create_cuda_graph_metadata(288-328)tensorrt_llm/_torch/distributed/communicator.py (3)
MPIDist(98-145)tp_size(46-47)tp_allgather(138-139)tensorrt_llm/_torch/expert_statistic.py (2)
ExpertStatistic(10-98)set_iter(32-36)tensorrt_llm/_torch/modules/multi_stream_utils.py (1)
with_multi_stream(26-32)tensorrt_llm/_torch/speculative/interface.py (2)
SpecMetadata(122-217)create_cuda_graph_metadata(181-192)tensorrt_llm/_torch/pyexecutor/resource_manager.py (1)
ResourceManagerType(47-52)tensorrt_llm/_torch/pyexecutor/scheduler.py (2)
ScheduledRequests(18-39)can_run_cuda_graph(31-32)
🪛 Ruff (0.12.2)
tests/unittest/_torch/modeling/test_modeling_exaone4.py
25-25: Module level import not at top of file
(E402)
tensorrt_llm/_torch/pyexecutor/cuda_graph_runner.py
43-43: Undefined name DecodingBaseConfig
(F821)
72-72: Undefined name Request
(F821)
🔇 Additional comments (21)
tests/unittest/_torch/modeling/test_modeling_llama_min_latency.py (1)
7-7: Good switch to factory helper.Decoupling tests from CUDAGraphRunner construction via create_mock_cuda_graph_runner improves isolation and avoids circular deps.
tests/unittest/_torch/modeling/test_modeling_phi3.py (1)
7-7: Factory-based runner import looks good.Keeps tests engine-agnostic and matches the new API surface.
tests/unittest/_torch/modeling/test_modeling_mllama.py (1)
6-6: Nice: helper import unifies CUDA-graph setup across tests.tests/unittest/_torch/modeling/test_modeling_qwen_moe.py (1)
6-6: Good move to create_mock_cuda_graph_runner.tests/unittest/_torch/modeling/test_modeling_mixtral.py (1)
6-6: Helper import LGTM.tests/unittest/_torch/modeling/test_modeling_exaone4.py (1)
355-357: CUDA graph capture/replay usage matches new APICapture with the added boolean flag and subsequent replay look correct.
Also applies to: 364-364
tests/unittest/_torch/modeling/test_modeling_qwen.py (1)
20-20: Import of helper factory looks correctAligned with new factory-based setup.
tests/unittest/_torch/modeling/test_modeling_nemotron.py (2)
7-7: LGTM: switched to factory helper.Importing
create_mock_cuda_graph_runnerdecouples tests from the runner class.
335-337: LGTM: updated capture signature correctly.
is_spec_decode=Falsehere is appropriate for pure decoding.tests/unittest/_torch/modeling/test_modeling_mistral.py (2)
10-10: LGTM: switched to factory helper.Keeps tests aligned with the new API surface.
416-416: LGTM: capture API use matches new signature.tests/unittest/_torch/modeling/test_modeling_llama.py (2)
7-7: LGTM: switched to factory helper.
343-345: LGTM: capture API updated correctly.tensorrt_llm/_torch/pyexecutor/model_engine.py (2)
460-474: LGTM: runner decoupled and fully parameterized.Constructor wiring looks correct and removes the prior dependency.
2258-2264: Confirmed safe:draft_tokens_cuda=Noneis tolerated
maybe_get_cuda_graphonly referencesdraft_tokens_cudawhenspec_metadatais non-null, so passingNonefor non-speculative paths is ignored and causes no errors.tensorrt_llm/_torch/pyexecutor/cuda_graph_runner.py (6)
32-49: Well-designed dependency injection refactorGreat work converting from engine-centric initialization to explicit dependency injection! This change successfully breaks the circular dependency by accepting individual parameters rather than the full engine object. The use of keyword-only arguments ensures clarity at call sites.
77-85: Enhanced API with speculative decoding contextThe updated
maybe_get_cuda_graphsignature properly extends support for speculative decoding by acceptingis_spec_decode,spec_metadata, anddraft_tokens_cudaparameters. This provides the necessary context for graph eligibility decisions.
101-111: Distributed batch size synchronization logic looks correctThe multi-GPU synchronization properly gathers batch information across TP ranks and validates consistency before allowing CUDA graph execution. This ensures all ranks are in sync.
127-135: Proper metadata creation for CUDA graphsThe code correctly creates graph-specific metadata for both attention and speculative decoding, properly handling the draft tokens buffer assignment.
168-171: Conditional mrope position deltas handlingGood use of the
use_mropeflag to conditionally include position deltas in the static tensors.
266-271: Comprehensive dummy request configurationThe padding dummy request creation properly includes all necessary parameters for speculative decoding support including
max_num_draft_tokens,use_mrope, andmax_beam_width.
|
PR_Github #17812 [ run ] completed with state |
QiJune
force-pushed
the
clean_cuda_graph
branch
from
c3143c4 to
b727284
Compare
|
/bot run |
|
PR_Github #19305 [ run ] triggered by Bot |
|
PR_Github #19305 [ run ] completed with state |
|
/bot run |
|
PR_Github #19310 [ run ] triggered by Bot |
|
PR_Github #19310 [ run ] completed with state |
|
/bot run |
|
PR_Github #19948 [ run ] triggered by Bot |
|
PR_Github #19948 [ run ] completed with state |
|
/bot run |
|
PR_Github #23615 [ run ] triggered by Bot. Commit: |
|
PR_Github #23615 [ run ] completed with state |
|
/bot run |
|
PR_Github #23993 [ run ] triggered by Bot. Commit: |
|
/bot run |
|
PR_Github #24023 [ run ] triggered by Bot. Commit: |
|
PR_Github #23993 [ run ] completed with state |
|
PR_Github #24023 [ run ] completed with state |
|
/bot run |
|
PR_Github #24080 [ run ] triggered by Bot. Commit: |
|
PR_Github #24080 [ run ] completed with state |
|
/bot run |
|
PR_Github #24132 [ run ] triggered by Bot. Commit: |
|
PR_Github #24132 [ run ] completed with state |
|
/bot run |
|
PR_Github #24184 [ run ] triggered by Bot. Commit: |
|
PR_Github #24184 [ run ] completed with state |
…and cuda graph runner (NVIDIA#7572) Signed-off-by: junq <22017000+QiJune@users.noreply.github.com>
7 participants
Summary by CodeRabbit
Description
Test Coverage
PR Checklist
Please review the following before submitting your PR:
PR description clearly explains what and why. If using CodeRabbit's summary, please make sure it makes sense.
PR Follows TRT-LLM CODING GUIDELINES to the best of your knowledge.
Test cases are provided for new code paths (see test instructions)
Any new dependencies have been scanned for license and vulnerabilities
CODEOWNERS updated if ownership changes
Documentation updated as needed
The reviewers assigned automatically/manually are appropriate for the PR.
Please check this after reviewing the above items as appropriate for this PR.
GitHub Bot Help
/bot [-h] ['run', 'kill', 'skip', 'reuse-pipeline'] ...Provide a user friendly way for developers to interact with a Jenkins server.
Run
/bot [-h|--help]to print this help message.See details below for each supported subcommand.
Details
run [--reuse-test (optional)pipeline-id --disable-fail-fast --skip-test --stage-list "A10-PyTorch-1, xxx" --gpu-type "A30, H100_PCIe" --test-backend "pytorch, cpp" --add-multi-gpu-test --only-multi-gpu-test --disable-multi-gpu-test --post-merge --extra-stage "H100_PCIe-TensorRT-Post-Merge-1, xxx" --detailed-log --debug(experimental)]Launch build/test pipelines. All previously running jobs will be killed.
--reuse-test (optional)pipeline-id(OPTIONAL) : Allow the new pipeline to reuse build artifacts and skip successful test stages from a specified pipeline or the last pipeline if no pipeline-id is indicated. If the Git commit ID has changed, this option will be always ignored. The DEFAULT behavior of the bot is to reuse build artifacts and successful test results from the last pipeline.--disable-reuse-test(OPTIONAL) : Explicitly prevent the pipeline from reusing build artifacts and skipping successful test stages from a previous pipeline. Ensure that all builds and tests are run regardless of previous successes.--disable-fail-fast(OPTIONAL) : Disable fail fast on build/tests/infra failures.--skip-test(OPTIONAL) : Skip all test stages, but still run build stages, package stages and sanity check stages. Note: Does NOT update GitHub check status.--stage-list "A10-PyTorch-1, xxx"(OPTIONAL) : Only run the specified test stages. Examples: "A10-PyTorch-1, xxx". Note: Does NOT update GitHub check status.--gpu-type "A30, H100_PCIe"(OPTIONAL) : Only run the test stages on the specified GPU types. Examples: "A30, H100_PCIe". Note: Does NOT update GitHub check status.--test-backend "pytorch, cpp"(OPTIONAL) : Skip test stages which don't match the specified backends. Only support [pytorch, cpp, tensorrt, triton]. Examples: "pytorch, cpp" (does not run test stages with tensorrt or triton backend). Note: Does NOT update GitHub pipeline status.--only-multi-gpu-test(OPTIONAL) : Only run the multi-GPU tests. Note: Does NOT update GitHub check status.--disable-multi-gpu-test(OPTIONAL) : Disable the multi-GPU tests. Note: Does NOT update GitHub check status.--add-multi-gpu-test(OPTIONAL) : Force run the multi-GPU tests in addition to running L0 pre-merge pipeline.--post-merge(OPTIONAL) : Run the L0 post-merge pipeline instead of the ordinary L0 pre-merge pipeline.--extra-stage "H100_PCIe-TensorRT-Post-Merge-1, xxx"(OPTIONAL) : Run the ordinary L0 pre-merge pipeline and specified test stages. Examples: --extra-stage "H100_PCIe-TensorRT-Post-Merge-1, xxx".--detailed-log(OPTIONAL) : Enable flushing out all logs to the Jenkins console. This will significantly increase the log volume and may slow down the job.--debug(OPTIONAL) : Experimental feature. Enable access to the CI container for debugging purpose. Note: Specify exactly one stage in thestage-listparameter to access the appropriate container environment. Note: Does NOT update GitHub check status.For guidance on mapping tests to stage names, see
docs/source/reference/ci-overview.mdand the
scripts/test_to_stage_mapping.pyhelper.kill
killKill all running builds associated with pull request.
skip
skip --comment COMMENTSkip testing for latest commit on pull request.
--comment "Reason for skipping build/test"is required. IMPORTANT NOTE: This is dangerous since lack of user care and validation can cause top of tree to break.reuse-pipeline
reuse-pipelineReuse a previous pipeline to validate current commit. This action will also kill all currently running builds associated with the pull request. IMPORTANT NOTE: This is dangerous since lack of user care and validation can cause top of tree to break.