Fix: dequantize fused_wqa_wkv instead of separate wq_a/wkv

wq_a and wkv are fused into a single MergedColumnParallelLinear
called fused_wqa_wkv. Was checking for non-existent separate attrs.
This commit is contained in:
2026-05-18 13:47:08 +00:00
parent a7216b27df
commit 2e8c3c961f

View File

@@ -1712,7 +1712,8 @@ class DeepseekV4Model(nn.Module):
"""
# All attention projections to dequantize to BF16
# wo_a is excluded — it uses fp8_einsum (no input_scale, weight-only FP8)
bf16_proj_names = {"wq_a", "wq_b", "wkv", "wo_b"}
# wq_a and wkv are fused into fused_wqa_wkv
bf16_proj_names = {"wq_b", "wo_b", "fused_wqa_wkv"}
fp8_proj_names = {"wo_a"}
bf16_converted = 0
fp8_converted = 0