Skip to content

[TRTLLM-6854][feat] Enable guided decoding with CUDA graph padding and draft model chunked prefill #6774

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
merged 1 commit into from
Aug 12, 2025
Merged

[TRTLLM-6854][feat] Enable guided decoding with CUDA graph padding and draft model chunked prefill #6774

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
9 changes: 7 additions & 2 deletions tensorrt_llm/_torch/pyexecutor/guided_decoder.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -83,7 +83,11 @@ def _require_matcher_advance(self, llm_req: LlmRequest) -> bool:
if llm_req.guided_decoding_params is None:
return False
if llm_req.py_is_draft:
return True
if llm_req.is_context_init_state and llm_req.is_last_context_chunk:
return True
if llm_req.is_generation_in_progress_state:
return True
return False
# The request is in a generation forward step.
return llm_req.is_generation_in_progress_state

Expand Down Expand Up @@ -189,7 +193,8 @@ def execute(self,
batched_bitmask.append(self.bitmask[slot, i])
offset += len(llm_req.py_draft_tokens) + 1

assert offset == logits.size(0)
# Dummy logits may exist for CUDA graph dummy requests.
assert offset <= logits.size(0)

if len(batched_logits) > 0:
torch.ops.trtllm.logits_bitmask(batched_logits, batched_bitmask)
Expand Down
6 changes: 6 additions & 0 deletions tests/integration/defs/accuracy/test_llm_api_pytorch.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -325,6 +325,7 @@ def test_guided_decoding_4gpus(self, backend: str, mocker):
def test_guided_decoding_with_eagle3(self, backend: str, mocker):
mocker.patch.dict(os.environ, {"TRTLLM_XGUIDANCE_LENIENT": "1"})
kv_cache_config = KvCacheConfig(free_gpu_memory_fraction=0.8)
cuda_graph_config = CudaGraphConfig(enable_padding=True)
spec_config = EagleDecodingConfig(
max_draft_len=3,
speculative_model_dir=
Expand All @@ -333,6 +334,8 @@ def test_guided_decoding_with_eagle3(self, backend: str, mocker):
llm = LLM(self.MODEL_PATH,
guided_decoding_backend=backend,
kv_cache_config=kv_cache_config,
cuda_graph_config=cuda_graph_config,
enable_chunked_prefill=True,
speculative_config=spec_config,
disable_overlap_scheduler=True)
with llm:
Expand All @@ -344,11 +347,14 @@ def test_guided_decoding_with_eagle3(self, backend: str, mocker):
def test_guided_decoding_with_ngram(self, backend: str, mocker):
mocker.patch.dict(os.environ, {"TRTLLM_XGUIDANCE_LENIENT": "1"})
kv_cache_config = KvCacheConfig(free_gpu_memory_fraction=0.8)
cuda_graph_config = CudaGraphConfig(enable_padding=True)
spec_config = NGramDecodingConfig(max_draft_len=3,
max_matching_ngram_size=3)
llm = LLM(self.MODEL_PATH,
guided_decoding_backend=backend,
kv_cache_config=kv_cache_config,
cuda_graph_config=cuda_graph_config,
enable_chunked_prefill=True,
speculative_config=spec_config,
disable_overlap_scheduler=True)
with llm:
Expand Down