[Bugfix] [ROCm] [AITER]: Fix aiter block quant not compatible with torch compile dynamo (#28716)

Signed-off-by: tjtanaa <tunjian.tan@embeddedllm.com>
2025-11-14 10:30:50 -08:00
parent 964d65deed
commit a425dc256e
3 changed files with 180 additions and 7 deletions
--- a/vllm/model_executor/layers/quantization/utils/fp8_utils.py
+++ b/vllm/model_executor/layers/quantization/utils/fp8_utils.py
@@ -342,7 +342,7 @@ class W8A8BlockFp8LinearOp:
            )
        # MI300 uses tuned AITER ASM/C++ kernel
        else:
-            q_input, input_scale = rocm_aiter_ops.per_1x128_fp8_quant(input_2d)
+            q_input, input_scale = rocm_aiter_ops.group_fp8_quant(input_2d)

        return gemm_a8w8_blockscale_op(
            q_input,