[sgl-kernel][Feat][B200][1/N] Support MXFP8 Grouped GEMM in Blackwell (#13731)

Co-authored-by: Yineng Zhang <me@zhyncs.com>

sgl-kernel/python/sgl_kernel/__init__.py

@@ -32,7 +32,11 @@ from sgl_kernel.elementwise import (
     rmsnorm,
     silu_and_mul,
 )
-from sgl_kernel.expert_specialization import es_fp8_blockwise_scaled_grouped_mm
+from sgl_kernel.expert_specialization import (
+    es_fp8_blockwise_scaled_grouped_mm,
+    es_sm100_mxfp8_blockscaled_grouped_mm,
+    es_sm100_mxfp8_blockscaled_grouped_quant,
+)
 from sgl_kernel.fused_moe import moe_wna16_marlin_gemm
 from sgl_kernel.gemm import (
     awq_dequantize,

sgl-kernel/python/sgl_kernel/expert_specialization.py

@@ -27,3 +27,24 @@ def es_fp8_blockwise_scaled_grouped_mm(
         expert_offsets,
         workspace,
     )
+
+
+def es_sm100_mxfp8_blockscaled_grouped_mm(
+    output, a, b, sfa, sfb, problem_sizes, expert_offsets, blockscale_offsets
+):
+    torch.ops.sgl_kernel.es_sm100_mxfp8_blockscaled_grouped_mm.default(
+        a, b, sfa, sfb, output, problem_sizes, expert_offsets, blockscale_offsets
+    )
+
+
+def es_sm100_mxfp8_blockscaled_grouped_quant(
+    input, problem_sizes, expert_offsets, blockscale_offsets, quant_output, scale_factor
+):
+    torch.ops.sgl_kernel.es_sm100_mxfp8_blockscaled_grouped_quant.default(
+        input,
+        problem_sizes,
+        expert_offsets,
+        blockscale_offsets,
+        quant_output,
+        scale_factor,
+    )