[ROCm][CI] Add missing quantization methods and fix online quant test failures (#39801)

Signed-off-by: Andreas Karatzas <akaratza@amd.com>
This commit is contained in:
Andreas Karatzas
2026-04-27 15:08:57 -05:00
committed by GitHub
parent c8bbe05189
commit 5e2c37facd
9 changed files with 50 additions and 23 deletions
@@ -376,11 +376,15 @@ class QuarkOCP_MX(QuarkScheme):
dq_w = self.dequant_func(layer.weight, layer.weight_scale, x.dtype)
qdq_x = self.quant_dequant_func(x)
return F.linear(qdq_x, dq_w, bias)
else:
return torch.ops.vllm.gemm_with_dynamic_quant(
x,
layer.weight,
layer.weight_scale,
self.rocm_use_aiter_fp4_asm_gemm,
self.out_dtype,
)
y = torch.ops.vllm.gemm_with_dynamic_quant(
x,
layer.weight,
layer.weight_scale,
self.rocm_use_aiter_fp4_asm_gemm,
self.out_dtype,
)
# gemm_with_dynamic_quant has no bias argument; add it here so the
# native path matches F.linear (e.g. qkv_proj with qkv_bias=True).
if bias is not None:
y = y + bias
return y