biondizzle/vllm
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

Fix fallback to default tactic (flashinfer autotuner) with trtllm_fp4_block_scale_moe (#35088)

Browse Source 
Signed-off-by: Daniel Serebrenik <daserebrenik@nvidia.com>
This commit is contained in:
danisereb
2026-02-24 17:25:44 +02:00
committed by GitHub
parent 34ce0ffd1f
commit a0c7081695
1 changed files with 2 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
vllm/model_executor/layers/quantization/utils/flashinfer_fp4_moe.py
View File

@@ -348,7 +348,7 @@ def flashinfer_trtllm_fp4_moe(
hidden_states=hidden_states_fp4,
hidden_states_scale=hidden_states_scale_linear_fp4.view(
torch.float8_e4m3fn
).flatten(),
).reshape(*hidden_states_fp4.shape[:-1], -1),
gemm1_weights=layer.w13_weight.data,
gemm1_weights_scale=layer.w13_weight_scale.data.view(torch.float8_e4m3fn),
gemm1_bias=None,
@@ -432,7 +432,7 @@ def flashinfer_trtllm_fp4_routed_moe(
hidden_states=hidden_states_fp4,
hidden_states_scale=hidden_states_scale_linear_fp4.view(
torch.float8_e4m3fn
).flatten(),
).reshape(*hidden_states_fp4.shape[:-1], -1),
gemm1_weights=layer.w13_weight.data,
gemm1_weights_scale=layer.w13_weight_scale.data.view(torch.float8_e4m3fn),
gemm1_bias=None,