Skip to content

[https://nvbugs/5815025][fix] Fix spec-dec mode flag and related cpp requirements #10996

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
pengbowang-nv merged 3 commits into from
Feb 4, 2026
Merged

[https://nvbugs/5815025][fix] Fix spec-dec mode flag and related cpp requirements #10996

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
45e5191
fix attention_need_spec_dec_mode flag
pengbowang-nv Jan 26, 2026
23334bf
unwaive related test
pengbowang-nv Jan 26, 2026
14b0c8f
Merge branch 'release/1.2' into fix-sm120-spec-dec
pengbowang-nv Jan 30, 2026
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
8 changes: 0 additions & 8 deletions cpp/tensorrt_llm/common/attentionOp.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1271,14 +1271,6 @@ int AttentionOp::mlaGeneration(
mXqaDispatcher->run(xqaParams, kv_cache_buffer, kv_scale_cache_buffer);
return 0;
}
else if (mIsSpecDecodingEnabled && mUseSpecDecoding)
{
TLLM_CHECK_WITH_INFO(false, "No available XQA kernels are found for speculative decoding mode.");
}
Comment thread
mikeiovine marked this conversation as resolved.
else if (mFuseFp4Quant)
{
TLLM_CHECK_WITH_INFO(false, "No available kernels are found for FP4 output.");
}
}

// Use FMHA otherwise.
Expand Down
8 changes: 2 additions & 6 deletions tensorrt_llm/_torch/pyexecutor/model_engine.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -3246,13 +3246,9 @@ def forward(self,
no_cache=kv_cache_manager
is None)
# attn_metadata now depends on spec_metadata since it determines the shape/content of spec_dec parameter Tensors
enable_mla = is_mla(self.model.model_config.pretrained_config)
is_spec_dec_mode = spec_metadata.spec_dec_mode.attention_need_spec_dec_mode(
spec_resource_manager,
self.is_draft_model,
self.attn_backend,
self.model_is_wrapped,
is_mla=enable_mla)
spec_resource_manager, self.is_draft_model, self.attn_backend,
self.model_is_wrapped)
attn_metadata.update_spec_dec_param(
batch_size=scheduled_requests.batch_size,
is_spec_decoding_enabled=is_spec_dec_mode,
Expand Down
14 changes: 6 additions & 8 deletions tensorrt_llm/_torch/speculative/interface.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -146,12 +146,11 @@ def extend_ctx(self, attention_backend: Type[AttentionBackend]):
TrtllmAttention) or not xqa_supported

def attention_need_spec_dec_mode(
self,
spec_resource_manager: Optional[BaseResourceManager],
is_draft_model: bool,
attention_backend: Type[AttentionBackend],
use_chain_drafter: bool, # CDL
is_mla: bool,
self,
spec_resource_manager: Optional[BaseResourceManager],
is_draft_model: bool,
attention_backend: Type[AttentionBackend],
use_chain_drafter: bool, # CDL
):
"""
If true, the attention backend kernel needs to run in spec-dec mode (multi-token query mode).
Expand All @@ -164,8 +163,7 @@ def attention_need_spec_dec_mode(
is_trtllm_attention = issubclass(attention_backend, TrtllmAttention)

# Always use the multi-token query mode for 1-model if the kernels are available.
xqa_supported = not is_mla or get_sm_version() < 120
use_case_1 = self.use_one_engine() and xqa_supported
use_case_1 = self.use_one_engine()
# For 2-model, we need to enable it when we process multiple tokens at once. This occurs with
# the target model (verification) or on the first draft for CDL based speculation.
use_case_2 = not self.use_one_engine() and (
Expand Down
1 change: 0 additions & 1 deletion tests/integration/test_lists/waives.txt
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -324,7 +324,6 @@ accuracy/test_llm_api_pytorch.py::TestGPTOSS::test_eagle3_vswa_reuse_4gpus[one_m
accuracy/test_llm_api_pytorch.py::TestGPTOSS::test_eagle3_vswa_reuse_4gpus[two_model] SKIP (https://nvbugs/5756028)
accuracy/test_llm_api_pytorch.py::TestLlama3_2_1B::test_fp8_prequantized SKIP (https://nvbugs/5785465)
accuracy/test_llm_api_pytorch.py::TestMinistral8BInstruct::test_fp8 SKIP (https://nvbugs/5785485)
accuracy/test_llm_api_pytorch.py::TestLlama3_3_70BInstruct::test_fp8_eagle3_tp8[eagle3_one_model=True-torch_compile=False] SKIP (https://nvbugs/5787892)
accuracy/test_llm_api_pytorch.py::TestLlama3_3_70BInstruct::test_fp8_eagle3_tp8[eagle3_one_model=False-torch_compile=False] SKIP (https://nvbugs/5787892)
accuracy/test_llm_api_pytorch.py::TestLlama4MaverickInstruct::test_fp8_chunked_prefill[tp8ep8-cuda_graph=True] SKIP (https://nvbugs/5791839)
accuracy/test_llm_api_pytorch.py::TestLlama4MaverickInstruct::test_fp8_chunked_prefill[tp8ep8-cuda_graph=False] SKIP (https://nvbugs/5795918)
Expand Down