[ROCm][CI] Enable AITER FP8/unquantized cases in modular-kernel sweep + fix MoRI per-tensor FP8 dispatch (#57074)

Signed-off-by: Divakar Verma <divakar.verma@amd.com>
Co-authored-by: Andreas Karatzas <akaratza@amd.com>
This commit is contained in:
Divakar Verma
2026-09-16 19:53:06 -05:00
committed by GitHub
co-authored by Andreas Karatzas
parent a615f53364
commit 95f4925c3a
6 changed files with 107 additions and 69 deletions
@@ -86,14 +86,14 @@ MK_FUSED_EXPERT_TYPES: list[mk.FusedMoEExpertsModular] = []
standard_format = mk.FusedMoEActivationFormat.Standard
batched_format = mk.FusedMoEActivationFormat.BatchedExperts
common_float_types: list[torch.dtype | str] = [
torch.float8_e4m3fn,
current_platform.fp8_dtype(),
torch.bfloat16,
torch.float16,
torch.float32,
]
common_float_and_int_types = common_float_types + [torch.int8]
nvfp4_types = ["nvfp4"]
fp8_types = [torch.float8_e4m3fn]
fp8_types = [current_platform.fp8_dtype()]
def register_prepare_and_finalize(
@@ -324,7 +324,9 @@ if has_aiter():
register_experts(
AiterExperts,
standard_format,
fp8_types,
# AiterExperts also supports the fully-unquantized (None, None)
# scheme (see SUPPORTED_W_A in rocm_aiter_moe.py), not just fp8.
common_float_types,
blocked_quantization_support=True,
needs_aiter=True,
)
@@ -397,35 +399,35 @@ MK_QUANT_CONFIGS: list[TestMoEQuantConfig | None] = [
None,
# per-channel / per-column weights and per-tensor activations
TestMoEQuantConfig(
quant_dtype=torch.float8_e4m3fn,
quant_dtype=current_platform.fp8_dtype(),
per_out_ch_quant=True,
per_act_token_quant=False,
block_shape=None,
),
# per-channel / per-column weights and per-token activations
TestMoEQuantConfig(
quant_dtype=torch.float8_e4m3fn,
quant_dtype=current_platform.fp8_dtype(),
per_out_ch_quant=True,
per_act_token_quant=True,
block_shape=None,
),
# per-tensor weights and per-tensor activations
TestMoEQuantConfig(
quant_dtype=torch.float8_e4m3fn,
quant_dtype=current_platform.fp8_dtype(),
per_out_ch_quant=False,
per_act_token_quant=False,
block_shape=None,
),
# per-tensor weights and per-token activations
TestMoEQuantConfig(
quant_dtype=torch.float8_e4m3fn,
quant_dtype=current_platform.fp8_dtype(),
per_out_ch_quant=False,
per_act_token_quant=True,
block_shape=None,
),
# block-quantized weights and 128 block per-token activations
TestMoEQuantConfig(
quant_dtype=torch.float8_e4m3fn,
quant_dtype=current_platform.fp8_dtype(),
per_out_ch_quant=False,
per_act_token_quant=False,
block_shape=[128, 128],