[Perf] dsv41: Use tuned MXFP8 tactics at decode sizes (#45630)
This commit is contained in:
@@ -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 = (
|
||||
|
||||
Reference in New Issue
Block a user