[ROCm] Enable DeepEP ROCm as all2allbackend for AMD GPUs. (#34692)

Signed-off-by: Tej Kiran <vpolamre@amd.com>
Co-authored-by: Tej Kiran <vpolamre@amd.com>
This commit is contained in:
Chaitanya Sri Krishna Lolla
2026-03-21 00:32:31 -07:00
committed by GitHub
co-authored by Tej Kiran
parent 02eec7ecbe
commit 3982bc2cd0
7 changed files with 68 additions and 29 deletions
@@ -25,6 +25,7 @@ from vllm.model_executor.layers.quantization.utils.mxfp8_utils import (
from vllm.model_executor.layers.quantization.utils.w8a8_utils import (
per_tensor_dequantize,
)
from vllm.platforms import current_platform
from vllm.triton_utils import tl, triton
from vllm.utils.math_utils import cdiv
from vllm.utils.torch_utils import is_torch_equal_or_newer
@@ -265,7 +266,7 @@ def moe_kernel_quantize_input(
# weights are already dequantized, and we proceed with normal
# activation quantization below.
if quant_dtype == torch.float8_e4m3fn:
if quant_dtype == current_platform.fp8_dtype():
return _fp8_quantize(A, A_scale, per_act_token_quant, block_shape)
elif quant_dtype == torch.int8:
return _int8_quantize(A, A_scale, per_act_token_quant, block_shape)