[Feature] Add SM103 (Blackwell Ultra) Support to vLLM (#30484)
Signed-off-by: LopezCastroRoberto <robertol.c510@gmail.com> Signed-off-by: Roberto L. Castro <38211239+LopezCastroRoberto@users.noreply.github.com> Co-authored-by: youkaichao <youkaichao@gmail.com>
This commit is contained in:
co-authored by
youkaichao
parent
57e9bf1864
commit
4fa7ce46f3
@@ -137,7 +137,7 @@ def get_fp8_moe_backend(
|
||||
if (
|
||||
current_platform.is_cuda()
|
||||
and (
|
||||
current_platform.is_device_capability(100)
|
||||
current_platform.is_device_capability_family(100)
|
||||
or current_platform.is_device_capability(90)
|
||||
)
|
||||
and envs.VLLM_USE_FLASHINFER_MOE_FP8
|
||||
@@ -148,7 +148,7 @@ def get_fp8_moe_backend(
|
||||
logger.info_once("Using FlashInfer FP8 MoE TRTLLM backend for SM100")
|
||||
return Fp8MoeBackend.FLASHINFER_TRTLLM
|
||||
else:
|
||||
if block_quant and current_platform.is_device_capability(100):
|
||||
if block_quant and current_platform.is_device_capability_family(100):
|
||||
raise ValueError(
|
||||
"FlashInfer FP8 MoE throughput backend does not "
|
||||
"support block quantization. Please use "
|
||||
@@ -193,7 +193,7 @@ def get_fp8_moe_backend(
|
||||
# CUTLASS BlockScaled GroupedGemm on SM100 with block-quantized weights
|
||||
if (
|
||||
current_platform.is_cuda()
|
||||
and current_platform.is_device_capability(100)
|
||||
and current_platform.is_device_capability_family(100)
|
||||
and block_quant
|
||||
):
|
||||
logger.info_once(
|
||||
|
||||
@@ -118,19 +118,19 @@ def get_mxfp4_backend(with_lora_support: bool) -> Mxfp4Backend:
|
||||
logger.info_once("Using FlashInfer MXFP4 BF16 backend for SM90")
|
||||
return Mxfp4Backend.SM90_FI_MXFP4_BF16
|
||||
elif (
|
||||
current_platform.is_device_capability(100)
|
||||
current_platform.is_device_capability_family(100)
|
||||
and has_flashinfer()
|
||||
and envs.VLLM_USE_FLASHINFER_MOE_MXFP4_MXFP8_CUTLASS
|
||||
):
|
||||
logger.info_once("Using FlashInfer MXFP4 MXFP8 CUTLASS backend for SM100")
|
||||
return Mxfp4Backend.SM100_FI_MXFP4_MXFP8_CUTLASS
|
||||
elif (
|
||||
current_platform.is_device_capability(100)
|
||||
current_platform.is_device_capability_family(100)
|
||||
and has_flashinfer()
|
||||
and envs.VLLM_USE_FLASHINFER_MOE_MXFP4_MXFP8
|
||||
):
|
||||
return Mxfp4Backend.SM100_FI_MXFP4_MXFP8_TRTLLM
|
||||
elif current_platform.is_device_capability(100) and has_flashinfer():
|
||||
elif current_platform.is_device_capability_family(100) and has_flashinfer():
|
||||
logger.info_once(
|
||||
"Using FlashInfer MXFP4 BF16 backend for SM100, "
|
||||
"For faster performance on SM100, consider setting "
|
||||
@@ -139,7 +139,7 @@ def get_mxfp4_backend(with_lora_support: bool) -> Mxfp4Backend:
|
||||
)
|
||||
return Mxfp4Backend.SM100_FI_MXFP4_BF16
|
||||
elif (
|
||||
current_platform.is_device_capability(100)
|
||||
current_platform.is_device_capability_family(100)
|
||||
or current_platform.is_device_capability(90)
|
||||
) and not has_flashinfer():
|
||||
logger.warning_once(
|
||||
|
||||
@@ -50,7 +50,7 @@ def is_flashinfer_fp4_cutedsl_moe_available() -> bool:
|
||||
envs.VLLM_USE_FLASHINFER_MOE_FP4
|
||||
and has_flashinfer_cutedsl_grouped_gemm_nt_masked()
|
||||
and current_platform.is_cuda()
|
||||
and current_platform.is_device_capability(100)
|
||||
and current_platform.is_device_capability_family(100)
|
||||
)
|
||||
|
||||
|
||||
|
||||
@@ -290,7 +290,7 @@ def get_flashinfer_moe_backend() -> FlashinferMoeBackend:
|
||||
if flashinfer_moe_backend in backend_map:
|
||||
if (
|
||||
flashinfer_moe_backend == "latency"
|
||||
and not current_platform.has_device_capability(100)
|
||||
and not current_platform.is_device_capability_family(100)
|
||||
):
|
||||
logger.info_once(
|
||||
"Flashinfer TRTLLM MOE backend is only supported on "
|
||||
|
||||
@@ -247,7 +247,7 @@ class W8A8BlockFp8LinearOp:
|
||||
self.act_quant_group_shape = act_quant_group_shape
|
||||
self.is_deep_gemm_supported = is_deep_gemm_supported()
|
||||
self.is_hopper = current_platform.is_device_capability(90)
|
||||
self.is_blackwell = current_platform.is_device_capability(100)
|
||||
self.is_blackwell = current_platform.is_device_capability_family(100)
|
||||
self.use_deep_gemm_e8m0 = is_deep_gemm_e8m0_used()
|
||||
|
||||
# Get the correct blockscale mul and input quant operations.
|
||||
|
||||
@@ -63,7 +63,7 @@ def _swizzle_mxfp4(quant_tensor, scale, num_warps):
|
||||
"split_k": 1,
|
||||
}
|
||||
opt_flags.update_opt_flags_constraints(constraints)
|
||||
elif current_platform.is_device_capability(100):
|
||||
elif current_platform.is_device_capability_family(100):
|
||||
constraints = {
|
||||
"is_persistent": True,
|
||||
"epilogue_subtile": 1,
|
||||
|
||||
Reference in New Issue
Block a user