[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:
Roberto L. Castro
2025-12-12 19:34:23 -08:00
committed by GitHub
co-authored by youkaichao
parent 57e9bf1864
commit 4fa7ce46f3
21 changed files with 53 additions and 33 deletions
@@ -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,