Skip to content

[https://nvbugs/5481385][fix] Fix max_seq_len in cuda graph warmup and intermediate_size in fused_moe_deepgemm #7345

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 4 commits into from
Aug 29, 2025
+12 −4
Merged

[https://nvbugs/5481385][fix] Fix max_seq_len in cuda graph warmup and intermediate_size in fused_moe_deepgemm #7345

Show file tree
Hide file tree
Changes from 3 commits
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
3ce2c13
fix max_seq_len in cuda graph warmup and intermediate_size in fused_m…
lfr-0531 Aug 28, 2025
5b1d479
Merge branch 'main' into user/fanrongl/fix_max_seq_len
lfr-0531 Aug 28, 2025
6cb8c84
Merge branch 'main' into user/fanrongl/fix_max_seq_len
litaotju Aug 28, 2025
e2dac98
Merge branch 'main' into user/fanrongl/fix_max_seq_len
litaotju Aug 29, 2025
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: 4 additions & 4 deletions tensorrt_llm/_torch/modules/fused_moe/fused_moe_deepgemm.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -411,7 +411,7 @@ def __init__(

def get_workspace(self, m_max: int, group_size: int):
hidden_size = self.hidden_size
intermediate_size = self.intermediate_size
intermediate_size = self.intermediate_size_per_partition
num_experts = self.expert_size_per_partition

# create workspace
Expand Down Expand Up @@ -564,7 +564,7 @@ def forward_chunk(
# grouped gemm 1
h1 = set_strides(workspace["workspace_1"],
self.expert_size_per_partition, m_max,
self.intermediate_size * 2)
self.intermediate_size_per_partition * 2)

deepgemm_fp8_group_blockwise_gemm(
d=h1,
Expand All @@ -579,9 +579,9 @@ def forward_chunk(
# activation and quantization
act_input_fp8 = set_strides(workspace["workspace_0"],
self.expert_size_per_partition, m_max,
self.intermediate_size)
self.intermediate_size_per_partition)

scale_k = fp8_utils.ceil_div(self.intermediate_size, 128)
scale_k = fp8_utils.ceil_div(self.intermediate_size_per_partition, 128)
scale_k_padded = fp8_utils.align(scale_k, 4)
act_input_sf = set_strides(workspace["workspace_sf"],
self.expert_size_per_partition,
Expand Down
8 changes: 8 additions & 0 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 @@ -583,7 +583,15 @@ def get_cuda_graph_warmup_request(batch_size, draft_len):

# Add one dummy request with the maximum possible sequence length.
# The sequence length is limited by both the max_seq_len and the number of available blocks.
# Also, the sequence length is limited by the max_position_embeddings.
token_num = max(1, min(available_tokens, self.max_seq_len - 1))
model_config = self.model.model_config.pretrained_config
max_position_embeddings = getattr(model_config,
'max_position_embeddings',
None)
if max_position_embeddings is not None:
token_num = min(token_num,
max_position_embeddings - draft_len)
max_seq_len_request = kv_cache_manager.add_dummy_requests(
request_ids=[batch_size - 1],
token_nums=[token_num],
Expand Down