[ROCm][CI] Fix moe layer tests for fp8 dtype compatibility (#56106)

Signed-off-by: Divakar Verma <divakar.verma@amd.com>
Co-authored-by: Andreas Karatzas <akaratza@amd.com>
This commit is contained in:
Divakar Verma
2026-09-09 17:48:14 +00:00
committed by GitHub
co-authored by Andreas Karatzas
parent a207d7ce1d
commit 3fb676bfad
2 changed files with 11 additions and 8 deletions
+4 -1
View File
@@ -84,7 +84,7 @@ def on_gfx950() -> bool:
return False
fp8_dtype = torch.float8_e4m3fn # current_platform.fp8_dtype
fp8_dtype = current_platform.fp8_dtype()
SHAPE_COMBOS = [
(1, 128, 256),
@@ -191,6 +191,9 @@ def mock_normalize_e4m3fn_to_e4m3fnuz(
# NOTE: Not able to use monkeypatch because of the spawned parallel workers.
def override_normalize_e4m3fn_to_e4m3fnuz():
vllm.model_executor.layers.quantization.utils.w8a8_utils.normalize_e4m3fn_to_e4m3fnuz = mock_normalize_e4m3fn_to_e4m3fnuz # noqa: E501
vllm.model_executor.layers.quantization.fp8.normalize_e4m3fn_to_e4m3fnuz = (
mock_normalize_e4m3fn_to_e4m3fnuz # noqa: E501
)
def sp_wrapper(
+7 -7
View File
@@ -33,6 +33,7 @@ from vllm.model_executor.layers.fused_moe.prepare_finalize.batched import (
)
from vllm.model_executor.layers.fused_moe.router.fused_topk_router import fused_topk
from vllm.model_executor.layers.fused_moe.utils import moe_kernel_quantize_input
from vllm.model_executor.layers.quantization.utils.fp8_utils import is_fp8
from vllm.model_executor.layers.quantization.utils.nvfp4_emulation_utils import (
ref_nvfp4_quant,
)
@@ -271,11 +272,10 @@ def moe_quantize_weights_2d(
per_token_quant: bool,
block_shape: list[int] | None,
) -> tuple[torch.Tensor, torch.Tensor | None, torch.Tensor | None]:
assert (
quant_dtype == torch.float8_e4m3fn
or quant_dtype == torch.int8
or quant_dtype == "nvfp4"
), "only fp8/int8/nvfp4 supported"
is_fp8_dtype = is_fp8(quant_dtype)
assert is_fp8_dtype or quant_dtype == torch.int8 or quant_dtype == "nvfp4", (
"only fp8/int8/nvfp4 supported"
)
w_gs = None
@@ -283,7 +283,7 @@ def moe_quantize_weights_2d(
assert not per_token_quant
if quant_dtype == torch.int8:
w, w_s = per_block_cast_to_int8(w, block_shape)
elif quant_dtype == torch.float8_e4m3fn:
elif is_fp8_dtype:
w, w_s = per_block_cast_to_fp8(w, block_shape)
elif quant_dtype == "nvfp4":
raise RuntimeError("blocked quantization not supported for nvfp4")
@@ -294,7 +294,7 @@ def moe_quantize_weights_2d(
w, w_s = ops.scaled_int8_quant(
w, w_s, use_per_token_if_dynamic=per_token_quant
)
elif quant_dtype == torch.float8_e4m3fn:
elif is_fp8_dtype:
w, w_s = ops.scaled_fp8_quant(
w, w_s, use_per_token_if_dynamic=per_token_quant
)