[Kernel] Add b12x FP4 MoE backend (#52018)
Signed-off-by: Luke Alonso <lalonso@gmail.com> Signed-off-by: mgoin <mgoin64@gmail.com> Signed-off-by: Misha Goin <mgoin64@gmail.com> Co-authored-by: mgoin <mgoin64@gmail.com>
This commit is contained in:
@@ -190,6 +190,7 @@ if TYPE_CHECKING:
|
||||
VLLM_HUMMING_INPUT_QUANT_CONFIG: dict[str, Any] | None = None
|
||||
VLLM_HUMMING_USE_F16_ACCUM: bool = False
|
||||
VLLM_HUMMING_MOE_GEMM_TYPE: Literal["indexed", "grouped", "auto"] | None = None
|
||||
VLLM_B12X_MOE_FP4_FORCE_A16: bool = False
|
||||
VLLM_DEEPEPLL_NVFP4_DISPATCH: bool = False
|
||||
VLLM_V1_USE_OUTLINES_CACHE: bool = False
|
||||
VLLM_TPU_USING_PATHWAYS: bool = False
|
||||
@@ -1621,6 +1622,10 @@ environment_variables: dict[str, Callable[[], Any]] = {
|
||||
"VLLM_BLOCKSCALE_FP8_GEMM_FLASHINFER": lambda: bool(
|
||||
int(os.getenv("VLLM_BLOCKSCALE_FP8_GEMM_FLASHINFER", "1"))
|
||||
),
|
||||
# Force b12x FP4 MoE to use BF16 activations.
|
||||
"VLLM_B12X_MOE_FP4_FORCE_A16": lambda: bool(
|
||||
int(os.getenv("VLLM_B12X_MOE_FP4_FORCE_A16", "0"))
|
||||
),
|
||||
# Allow use of FlashInfer MxInt4 MoE kernels for fused moe ops.
|
||||
"VLLM_USE_FLASHINFER_MOE_INT4": lambda: bool(
|
||||
int(os.getenv("VLLM_USE_FLASHINFER_MOE_INT4", "0"))
|
||||
|
||||
Reference in New Issue
Block a user