optimize: reduce shulffle and quantization overhead in cutlass_moe sm90 (#8962)

Co-authored-by: 戚余航 <qiyuhang@bytedance.com>

python/sglang/srt/layers/moe/cutlass_moe.py

@@ -9,7 +9,7 @@ from typing import Any, Callable, Dict, List, Optional, Tuple
 import torch
 
 from sglang.srt.layers.moe.cutlass_moe_params import CutlassMoEParams
-from sglang.srt.layers.utils import is_sm100_supported
+from sglang.srt.layers.utils import is_sm90_supported, is_sm100_supported
 from sglang.srt.utils import is_cuda
 
 _is_cuda = is_cuda()
@@ -124,6 +124,7 @@ def cutlass_fused_experts_fp8(
 
     if is_cuda:
         from sglang.srt.layers.quantization.fp8_kernel import (
+            per_group_transpose,
             per_token_group_quant_fp8_hopper_moe_mn_major,
             sglang_per_token_group_quant_fp8,
         )
@@ -152,15 +153,12 @@ def cutlass_fused_experts_fp8(
         k,
     )
 
-    if is_sm100_supported():
     a_q, a1_scale = sglang_per_token_group_quant_fp8(a, 128)
     rep_a_q = shuffle_rows(a_q, a_map, (m * topk, k))
     rep_a1_scales = shuffle_rows(a1_scale, a_map, (m * topk, int(k / 128)))
-    else:
-        rep_a = shuffle_rows(a, a_map, (m * topk, k))
-        rep_a_q, rep_a1_scales = per_token_group_quant_fp8_hopper_moe_mn_major(
-            rep_a, expert_offsets, problem_sizes1, 128
-        )
+
+    if not is_sm100_supported():
+        rep_a1_scales = per_group_transpose(rep_a1_scales, expert_offsets)
         w1_scale = w1_scale.contiguous()
 
     c1 = torch.empty((m * topk, n * 2), device=device, dtype=out_dtype)
@@ -193,12 +191,9 @@ def cutlass_fused_experts_fp8(
     intermediate = torch.empty((m * topk, n), device=device, dtype=out_dtype)
     silu_and_mul(c1, intermediate)
 
-    if is_sm100_supported():
     intemediate_q, a2_scale = sglang_per_token_group_quant_fp8(intermediate, 128)
-    else:
-        intemediate_q, a2_scale = per_token_group_quant_fp8_hopper_moe_mn_major(
-            intermediate, expert_offsets, problem_sizes2, 128
-        )
+    if not is_sm100_supported():
+        a2_scale = per_group_transpose(a2_scale, expert_offsets)
         w2_scale = w2_scale.contiguous()
 
     fp8_blockwise_scaled_grouped_mm(

python/sglang/srt/layers/quantization/fp8_kernel.py

@@ -1356,3 +1356,62 @@ def per_token_group_quant_fp8_hopper_moe_mn_major(
         expert_tokens_alignment,
     )
     return a_q, sfa
+
+
+@triton.jit
+def _per_group_transpose(
+    data_ptr: torch.Tensor,
+    trans_data_ptr: torch.Tensor,
+    expert_offsets: torch.Tensor,
+    k: int,
+    M_ALIGNMENT: tl.constexpr,
+    BLOCK_SIZE_M: tl.constexpr,
+    BLOCK_SIZE_K: tl.constexpr,
+):
+    expert_id = tl.program_id(0)
+    m_id = tl.program_id(1)
+    k_id = tl.program_id(2)
+
+    curr_expert_offset = tl.load(expert_offsets + expert_id)
+    next_expert_offset = tl.load(expert_offsets + expert_id + 1)
+    num_tokens_of_expert = next_expert_offset - curr_expert_offset
+    tl.multiple_of(curr_expert_offset, M_ALIGNMENT)
+    tl.multiple_of(next_expert_offset, M_ALIGNMENT)
+
+    data_start_ptr = data_ptr + curr_expert_offset * k
+    trans_data_start_ptr = trans_data_ptr + curr_expert_offset * k
+
+    k_coord = k_id * BLOCK_SIZE_K + tl.arange(0, BLOCK_SIZE_K)
+    k_mask = k_coord < k
+    for start_m in tl.range(0, num_tokens_of_expert, BLOCK_SIZE_M * tl.num_programs(1)):
+        m_coord = start_m + m_id * BLOCK_SIZE_M + tl.arange(0, BLOCK_SIZE_M)
+        m_mask = m_coord < num_tokens_of_expert
+        off = m_coord[:, None] * k + k_coord[None, :]
+        trans_off = m_coord[:, None] + k_coord[None, :] * num_tokens_of_expert
+        mask = m_mask[:, None] & k_mask[None, :]
+
+        data = tl.load(data_start_ptr + off, mask=mask)
+        tl.store(trans_data_start_ptr + trans_off, data, mask=mask)
+
+
+def per_group_transpose(
+    a: torch.Tensor,
+    expert_offsets: torch.Tensor,
+    M_ALIGNMENT: int = 1,
+) -> torch.Tensor:
+    assert a.dim() == 2
+    assert a.is_contiguous(), "`a` is not contiguous"
+
+    m, k = a.size()
+    trans_a = torch.empty_like(a)
+    num_experts = expert_offsets.size(0) - 1
+
+    grid = lambda META: (
+        num_experts,
+        triton.cdiv((m + num_experts - 1) // num_experts, META["BLOCK_SIZE_M"]),
+        triton.cdiv(k, META["BLOCK_SIZE_K"]),
+    )
+    _per_group_transpose[grid](
+        a, trans_a, expert_offsets, k, M_ALIGNMENT, BLOCK_SIZE_M=16, BLOCK_SIZE_K=8
+    )
+    return trans_a