Skip to content

[https://nvbugs/5916092][fix] Fix MTP+PP hang by preserving speculative layer weights on last PP rank #12555

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

Open
xxi-nv wants to merge 1 commit into
base: main
Choose a base branch
from
+10 −1
Open

[https://nvbugs/5916092][fix] Fix MTP+PP hang by preserving speculative layer weights on last PP rank #12555

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
10 changes: 10 additions & 0 deletions tensorrt_llm/_torch/models/modeling_utils.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -304,6 +304,16 @@ def __pp_init__(self):
pp_layer_list = mapping.pp_layers(num_hidden_layers)
has_pp_layer = len(pp_layer_list) > 0
for layer_idx, layer in enumerate(self.layers):
if layer_idx >= num_hidden_layers:
# Extra layers (e.g., MTP speculative layers) appended beyond
# the base model. Skip their forward on all ranks so they are
# no-ops in the main decoder loop, but preserve weights on the
# last PP rank where the MTP draft worker needs them.
if hasattr(layer, 'skip_forward'):
layer.forward = layer.skip_forward
if not mapping.is_last_pp_rank():
remove_weights(layer)
continue
is_last_layer = (layer_idx == num_hidden_layers - 1)
if layer_idx not in pp_layer_list:
# keep next layer's input_layernorm's weights for fusion
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 @@ -276,7 +276,6 @@ unittest/_torch/ray_orchestrator/single_gpu/test_llm_update_weights.py::test_llm
unittest/_torch/ray_orchestrator/single_gpu/test_llm_update_weights.py::test_llm_update_weights_with_quant_config[Qwen3/Qwen3-30B-A3B-Qwen3/Qwen3-30B-A3B-FP8] SKIP (https://nvbugspro.nvidia.com/bug/5911788)
unittest/_torch/ray_orchestrator/single_gpu/test_llm_update_weights.py::test_llm_partial_update_weights[Qwen3/Qwen3-30B-A3B] SKIP (https://nvbugspro.nvidia.com/bug/5911788)
unittest/_torch/ray_orchestrator/single_gpu/test_llm_update_weights.py::test_llm_partial_update_weights[Qwen3/Qwen3-8B] SKIP (https://nvbugspro.nvidia.com/bug/5911788)
accuracy/test_llm_api_pytorch.py::TestDeepSeekV3Lite::test_nvfp4_4gpus[moe_backend=CUTLASS-mtp_nextn=2-pp4-fp8kv=False-attention_dp=False-cuda_graph=False-overlap_scheduler=False-low_precision_combine=False-torch_compile=False] SKIP (https://nvbugspro.nvidia.com/bug/5916092)
accuracy/test_llm_api_pytorch.py::TestDeepSeekV3Lite::test_bfloat16_4gpus[pp4-mtp_nextn=0-attention_dp=False-cuda_graph=False-overlap_scheduler=False-torch_compile=True] SKIP (https://nvbugspro.nvidia.com/bug/5916155)
accuracy/test_llm_api_pytorch.py::TestDeepSeekV3Lite::test_bfloat16_4gpus[ep4-mtp_nextn=2-attention_dp=True-cuda_graph=True-overlap_scheduler=True-torch_compile=True] SKIP (https://nvbugspro.nvidia.com/bug/5916155)
accuracy/test_llm_api_pytorch.py::TestDeepSeekV3Lite::test_bfloat16_4gpus[pp4-mtp_nextn=2-attention_dp=True-cuda_graph=True-overlap_scheduler=True-torch_compile=True] SKIP (https://nvbugspro.nvidia.com/bug/5916155)
Expand Down
Loading