[ROCm][LoRA] Fix MoE accuracy regression by preserving float32 router weight scaling (#31931)

Signed-off-by: Andreas Karatzas <akaratza@amd.com>
This commit is contained in:
Andreas Karatzas
2026-01-07 22:17:56 -06:00
committed by GitHub
parent a79079feef
commit c4041f37a4

View File

@@ -519,6 +519,12 @@ def fused_moe_kernel(
a_ptrs += BLOCK_SIZE_K * stride_ak a_ptrs += BLOCK_SIZE_K * stride_ak
b_ptrs += BLOCK_SIZE_K * stride_bk b_ptrs += BLOCK_SIZE_K * stride_bk
# Router weight multiplication MUST happen in float32 before precision
# conversion for numerical stability (especially critical on ROCm).
if MUL_ROUTED_WEIGHT:
moe_weight = tl.load(topk_weights_ptr + offs_token, mask=token_mask, other=0)
accumulator = accumulator * moe_weight[:, None]
if use_int8_w8a16: if use_int8_w8a16:
accumulator = (accumulator * b_scale).to(compute_type) accumulator = (accumulator * b_scale).to(compute_type)
elif use_fp8_w8a8 or use_int8_w8a8: elif use_fp8_w8a8 or use_int8_w8a8:
@@ -529,12 +535,10 @@ def fused_moe_kernel(
else: else:
accumulator = accumulator.to(compute_type) accumulator = accumulator.to(compute_type)
# Since bias is typically not quantized, it's added after dequantization. # Bias is added AFTER dequantization since bias is typically stored in
# the output dtype and should not be scaled by quantization factors.
if HAS_BIAS: if HAS_BIAS:
accumulator = accumulator + bias[None, :] accumulator = accumulator + bias[None, :]
if MUL_ROUTED_WEIGHT:
moe_weight = tl.load(topk_weights_ptr + offs_token, mask=token_mask, other=0)
accumulator = accumulator * moe_weight[:, None]
# ----------------------------------------------------------- # -----------------------------------------------------------
# Write back the block of the output # Write back the block of the output