Skip to content

[nvbugs/5336321][fix] Enable attention dp = False test case, Fix TRTLLM Gen Moe workspace allocation #5463

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
Jul 4, 2025
Merged

[nvbugs/5336321][fix] Enable attention dp = False test case, Fix TRTLLM Gen Moe workspace allocation #5463

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
390794a
Fix
yizhang-nv Jun 25, 2025
67b7206
Update test_llm_api_pytorch.py
yizhang-nv Jun 25, 2025
097dcc9
Update test_llm_api_pytorch.py
yizhang-nv Jun 25, 2025
eff3a08
Disable cutlass moe min latency mode
yizhang-nv Jul 2, 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
3 changes: 2 additions & 1 deletion cpp/tensorrt_llm/thop/fp4BlockScaleMoe.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -76,6 +76,7 @@ std::vector<torch::Tensor> fp4_block_scale_moe_runner(torch::Tensor const& routi

TORCH_CHECK(num_experts % 4 == 0, "Routing kernel expects that num_experts must be divisible by 4");
TORCH_CHECK(num_experts > top_k, "num_experts must be greater than top_k");
TORCH_CHECK(num_experts <= 256, "num_experts must be less than or equal to 256");

tensorrt_llm::kernels::trtllmGenFp8BlockScaleMoe::MoE::MoERunnerArgs args;
tensorrt_llm::kernels::trtllmGenFp8BlockScaleMoe::MoE::MoEWorkspace workspace;
Expand Down Expand Up @@ -123,7 +124,7 @@ std::vector<torch::Tensor> fp4_block_scale_moe_runner(torch::Tensor const& routi
{args.num_tokens, args.top_k}, routing_bias_dtype, routing_logits.device(), std::nullopt);
at::Tensor expert_indexes = at::detail::empty_cuda(
{args.num_tokens, args.top_k}, at::ScalarType::Int, routing_logits.device(), std::nullopt);
at::Tensor expert_count_histogram = at::detail::empty_cuda({((num_experts * 2 + 255) / 256) * 256},
at::Tensor expert_count_histogram = at::detail::empty_cuda({2 * 256},
at::ScalarType::Int, // 256 is the max number of threads per block and max number of experts
routing_logits.device(), std::nullopt);

Expand Down
7 changes: 3 additions & 4 deletions tensorrt_llm/_torch/models/modeling_deepseekv3.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -724,10 +724,9 @@ def _compute_mlp_tp_size(self, intermediate_size: int,
return mlp_tp_size

def _enable_min_latency_mode(self, num_tokens: int):
return (num_tokens <= 128 and self.fusion_config.POST_MOE_FUSION
and self.is_nvfp4 and self.model_config.moe_backend == 'CUTLASS'
and not self.mapping.is_multi_node()
and self.allreduce.mnnvl_allreduce is None)
# Disable cutlass min latency mode since it will cause illegal memory access
# Will use trtllm-gen moe in future
return False

def forward(
self,
Expand Down
3 changes: 1 addition & 2 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 @@ -882,8 +882,7 @@ def test_nvfp4_4gpus(self, fp8kv, attention_dp, cuda_graph,
pytest.skip("https://nvbugs/5252559")
if torch_compile and pp_size > 1:
pytest.skip("PP with torch.compile is not supported yet.")
if not attention_dp and (tp_size > 1 or ep_size > 1):
pytest.skip("https://nvbugs/5336321")

kv_cache_config = KvCacheConfig(free_gpu_memory_fraction=0.9)
torch_compile_config = TorchCompileConfig(
enable_fullgraph=True,
Expand Down