[Perf] dsv41: Use tuned MXFP8 tactics at decode sizes (#45630)

This commit is contained in:
Khoa Pham
2026-10-21 13:35:56 -07:00
committed by GitHub
parent a2e4ffbcea
commit 028de2c5d4
2 changed files with 3 additions and 33 deletions
+2 -2
View File
@@ -4578,8 +4578,8 @@ class DeepseekV4ForCausalLM(nn.Module):
if backend is None or not backend.is_flashinfer_cutedsl():
continue
if method.block_fp8_as_mxfp8:
# Small shapes and deterministic execution keep their pinned tactic.
method.mxfp8_prefill_autotune_min_tokens = 4096
# Deterministic execution keeps its pinned tactic.
method.mxfp8_prefill_autotune_min_tokens = 1
weight = layer.weight
scale = layer.weight_scale_inv_swizzled
key = (