[Bugfix][Kernel] Correct FlashInfer CUTLASS MoE tuning token bound (#46838)
Signed-off-by: Haobin Guo <haobing@nvidia.com>
This commit is contained in:
@@ -93,7 +93,6 @@ class FlashInferExperts(mk.FusedMoEExpertsModular):
|
|||||||
# - pass per-block weight scales to the kernel
|
# - pass per-block weight scales to the kernel
|
||||||
# - skip input activation quantization (kernel applies scaling)
|
# - skip input activation quantization (kernel applies scaling)
|
||||||
self.use_deepseek_fp8_block_scale = quant_config.is_block_quantized
|
self.use_deepseek_fp8_block_scale = quant_config.is_block_quantized
|
||||||
self.max_capture_size = moe_config.max_capture_size
|
|
||||||
self.gemm1_clamp_limit: torch.Tensor | None = None
|
self.gemm1_clamp_limit: torch.Tensor | None = None
|
||||||
if quant_config.gemm1_clamp_limit is not None:
|
if quant_config.gemm1_clamp_limit is not None:
|
||||||
self.gemm1_clamp_limit = torch.tensor(
|
self.gemm1_clamp_limit = torch.tensor(
|
||||||
@@ -398,7 +397,6 @@ class FlashInferExperts(mk.FusedMoEExpertsModular):
|
|||||||
use_deepseek_fp8_block_scale=self.use_deepseek_fp8_block_scale,
|
use_deepseek_fp8_block_scale=self.use_deepseek_fp8_block_scale,
|
||||||
use_mxfp8_act_scaling=use_mxfp8_act_scaling,
|
use_mxfp8_act_scaling=use_mxfp8_act_scaling,
|
||||||
use_w4_group_scaling=use_w4_group_scaling,
|
use_w4_group_scaling=use_w4_group_scaling,
|
||||||
tune_max_num_tokens=max(self.max_capture_size, 1),
|
|
||||||
)
|
)
|
||||||
|
|
||||||
def moe_sum(self, input: torch.Tensor, output: torch.Tensor) -> None:
|
def moe_sum(self, input: torch.Tensor, output: torch.Tensor) -> None:
|
||||||
|
|||||||
Reference in New Issue
Block a user