Skip to content

[https://nvbugs/5753788][chore] fix empty tensor cutlass moe #10352

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
leslie-fang25 wants to merge 1 commit into
base: main
Choose a base branch
from
+29 −5
Open

[https://nvbugs/5753788][chore] fix empty tensor cutlass moe #10352

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: 8 additions & 2 deletions cpp/tensorrt_llm/kernels/quantization.cu
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -146,7 +146,10 @@ void invokeFP4Quantization(int b, int m, int n, T const* input, float const* SFS
int const numBlocksPerSM = std::max(1u, 2048u / block.x);
// The number of blocks for m. The m dimension will be padded to 128 for swizzled layout.
int numBlocksForM = layout == QuantizationSFLayout::SWIZZLED ? PadUpFn(m, 128) : m;
dim3 grid(std::min(numBlocksForM, multiProcessorCount * numBlocksPerSM));
int gridSize = std::min(numBlocksForM, multiProcessorCount * numBlocksPerSM);
// Ensure gridSize is not zero.
gridSize = std::max(1, gridSize);
dim3 grid(gridSize);

// Launch the cvt kernel.
auto* kernel_instance = useUE8M0
Expand All @@ -165,7 +168,10 @@ void invokeFP4Quantization(int b, int m, int n, T const* input, float const* SFS
int const numBlocksPerSM = std::max(1u, 2048u / block.x);
// The number of blocks for m. The m dimension will be padded to 128 for swizzled layout.
int numBlocksForM = layout == QuantizationSFLayout::SWIZZLED ? PadUpFn(m, 128) : m;
dim3 grid(std::min(numBlocksForM, multiProcessorCount * numBlocksPerSM));
int gridSize = std::min(numBlocksForM, multiProcessorCount * numBlocksPerSM);
// Ensure gridSize is not zero.
gridSize = std::max(1, gridSize);
dim3 grid(gridSize);

// Launch the cvt kernel.
auto* kernel_instance = useUE8M0
Expand Down
24 changes: 21 additions & 3 deletions tensorrt_llm/_torch/modules/fused_moe/fused_moe_cutlass.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -319,11 +319,17 @@ def quantize_input(
x_row = x.shape[0]
else:
x_row = x.shape[0]
hidden_size = x.shape[-1]
x, x_sf = torch.ops.trtllm.fp4_quantize(
x, self.fc31_input_scale, self.scaling_vector_size,
False, False)
if x_sf.numel() == 0 and x_sf.dim() == 1:
# View torch.Size[0] in to (0, -1) is not supported
x_sf = x_sf.view(
(0,
hidden_size // int(self.scaling_vector_size)))
# Reshape x_sf to 2D for post-quant communication
if x_sf is not None:
if x_sf is not None and x_sf.numel() != 0:
x_sf = x_sf.view((x_row, -1))
else:
if not isinstance(x, Fp4QuantizedTensor):
Expand Down Expand Up @@ -494,8 +500,20 @@ def forward_chunk(
self._load_balancer_start_wait_gpu_stage(is_first_call)

# apply routing
token_selected_experts, token_final_scales = self.routing_method.apply(
router_logits)
if router_logits.numel() == 0:
# For dtype, refer to https://github.com/NVIDIA/TensorRT-LLM/blob/55f3cda66d05a2e5686c9c7512721beb522bc8b7/tensorrt_llm/_torch/modules/fused_moe/routing.py#L327
token_selected_experts = torch.empty(
(0, self.routing_method.experts_per_token),
dtype=torch.int32,
device=router_logits.device)
token_final_scales = torch.empty(
(0, self.routing_method.experts_per_token),
dtype=torch.float32,
device=router_logits.device)
else:
token_selected_experts, token_final_scales = self.routing_method.apply(
router_logits)

assert token_selected_experts.shape[
1] == self.routing_method.experts_per_token
assert token_selected_experts.shape == token_final_scales.shape
Expand Down