[ROCm][Quant] Minimax-M3: Enable fp8_per_channel for bf16 weights on mi300x (#45854)
Signed-off-by: Hongxia Yang <hongxia.yang@amd.com> Signed-off-by: tjtanaa <tunjian.tan@embeddedllm.com> Co-authored-by: tjtanaa <tunjian.tan@embeddedllm.com>
This commit is contained in:
@@ -603,6 +603,8 @@ def fp8_w8a8_moe_quant_config(
|
|||||||
a2_gscale: torch.Tensor | None = None,
|
a2_gscale: torch.Tensor | None = None,
|
||||||
g1_alphas: torch.Tensor | None = None,
|
g1_alphas: torch.Tensor | None = None,
|
||||||
g2_alphas: torch.Tensor | None = None,
|
g2_alphas: torch.Tensor | None = None,
|
||||||
|
gemm1_alpha: float | None = None,
|
||||||
|
gemm1_beta: float | None = None,
|
||||||
gemm1_clamp_limit: float | None = None,
|
gemm1_clamp_limit: float | None = None,
|
||||||
) -> FusedMoEQuantConfig:
|
) -> FusedMoEQuantConfig:
|
||||||
"""
|
"""
|
||||||
@@ -623,6 +625,8 @@ def fp8_w8a8_moe_quant_config(
|
|||||||
per_act_token_quant=per_act_token_quant,
|
per_act_token_quant=per_act_token_quant,
|
||||||
per_out_ch_quant=per_out_ch_quant,
|
per_out_ch_quant=per_out_ch_quant,
|
||||||
block_shape=block_shape,
|
block_shape=block_shape,
|
||||||
|
gemm1_alpha=gemm1_alpha,
|
||||||
|
gemm1_beta=gemm1_beta,
|
||||||
gemm1_clamp_limit=gemm1_clamp_limit,
|
gemm1_clamp_limit=gemm1_clamp_limit,
|
||||||
)
|
)
|
||||||
|
|
||||||
|
|||||||
@@ -568,6 +568,8 @@ def make_fp8_moe_quant_config(
|
|||||||
block_shape=block_shape,
|
block_shape=block_shape,
|
||||||
per_act_token_quant=per_act_token_quant,
|
per_act_token_quant=per_act_token_quant,
|
||||||
per_out_ch_quant=per_out_ch_quant,
|
per_out_ch_quant=per_out_ch_quant,
|
||||||
|
gemm1_alpha=gemm1_alpha,
|
||||||
|
gemm1_beta=gemm1_beta,
|
||||||
gemm1_clamp_limit=swiglu_limit,
|
gemm1_clamp_limit=swiglu_limit,
|
||||||
)
|
)
|
||||||
|
|
||||||
|
|||||||
@@ -787,6 +787,8 @@ class Fp8MoEMethod(FusedMoEMethodBase):
|
|||||||
a2_scale=a2_scale,
|
a2_scale=a2_scale,
|
||||||
block_shape=self.weight_block_size,
|
block_shape=self.weight_block_size,
|
||||||
swiglu_limit=getattr(layer, "swiglu_limit", None),
|
swiglu_limit=getattr(layer, "swiglu_limit", None),
|
||||||
|
gemm1_alpha=getattr(layer, "swiglu_alpha", None),
|
||||||
|
gemm1_beta=getattr(layer, "swiglu_beta", None),
|
||||||
)
|
)
|
||||||
|
|
||||||
# Inject biases into the quant config if the model has them
|
# Inject biases into the quant config if the model has them
|
||||||
|
|||||||
@@ -483,6 +483,8 @@ class _Fp8OnlineMoEBase(OnlineMoEMethodBase):
|
|||||||
per_act_token_quant=self.per_act_token_quant,
|
per_act_token_quant=self.per_act_token_quant,
|
||||||
per_out_ch_quant=self.per_out_ch_quant,
|
per_out_ch_quant=self.per_out_ch_quant,
|
||||||
swiglu_limit=getattr(layer, "swiglu_limit", None),
|
swiglu_limit=getattr(layer, "swiglu_limit", None),
|
||||||
|
gemm1_alpha=getattr(layer, "swiglu_alpha", None),
|
||||||
|
gemm1_beta=getattr(layer, "swiglu_beta", None),
|
||||||
)
|
)
|
||||||
|
|
||||||
|
|
||||||
|
|||||||
@@ -459,6 +459,7 @@ class RocmPlatform(Platform):
|
|||||||
"modelopt_mixed",
|
"modelopt_mixed",
|
||||||
"fp8_per_tensor",
|
"fp8_per_tensor",
|
||||||
"fp8_per_block",
|
"fp8_per_block",
|
||||||
|
"fp8_per_channel",
|
||||||
"online",
|
"online",
|
||||||
"gpt_oss_mxfp4",
|
"gpt_oss_mxfp4",
|
||||||
]
|
]
|
||||||
|
|||||||
Reference in New Issue
Block a user