[Kernel] port sgl moe_align_block_size kernels (#12574)

sgl_moe_align_block_size is based on:


ded9fcd09a

moe_align_block_size is based on:


ba5112ff69

Signed-off-by: Yang Chen <yangche@fb.com>

vllm/model_executor/layers/fused_moe/fused_moe.py

@@ -405,6 +405,144 @@ def fused_moe_kernel(
     tl.store(c_ptrs, accumulator, mask=c_mask)
 
 
+def ceil_div(a, b):
+    return (a + b - 1) // b
+
+
+@triton.jit
+def moe_align_block_size_stage1(
+    topk_ids_ptr,
+    tokens_cnts_ptr,
+    num_experts: tl.constexpr,
+    numel: tl.constexpr,
+    tokens_per_thread: tl.constexpr,
+):
+    pid = tl.program_id(0)
+
+    start_idx = pid * tokens_per_thread
+
+    off_c = (pid + 1) * num_experts
+
+    for i in range(tokens_per_thread):
+        if start_idx + i < numel:
+            idx = tl.load(topk_ids_ptr + start_idx + i)
+            token_cnt = tl.load(tokens_cnts_ptr + off_c + idx)
+            tl.store(tokens_cnts_ptr + off_c + idx, token_cnt + 1)
+
+
+@triton.jit
+def moe_align_block_size_stage2(
+    tokens_cnts_ptr,
+    num_experts: tl.constexpr,
+):
+    pid = tl.program_id(0)
+
+    last_cnt = 0
+    for i in range(1, num_experts + 1):
+        token_cnt = tl.load(tokens_cnts_ptr + i * num_experts + pid)
+        last_cnt = last_cnt + token_cnt
+        tl.store(tokens_cnts_ptr + i * num_experts + pid, last_cnt)
+
+
+@triton.jit
+def moe_align_block_size_stage3(
+    total_tokens_post_pad_ptr,
+    tokens_cnts_ptr,
+    cumsum_ptr,
+    num_experts: tl.constexpr,
+    block_size: tl.constexpr,
+):
+    last_cumsum = 0
+    off_cnt = num_experts * num_experts
+    for i in range(1, num_experts + 1):
+        token_cnt = tl.load(tokens_cnts_ptr + off_cnt + i - 1)
+        last_cumsum = last_cumsum + tl.cdiv(token_cnt, block_size) * block_size
+        tl.store(cumsum_ptr + i, last_cumsum)
+    tl.store(total_tokens_post_pad_ptr, last_cumsum)
+
+
+@triton.jit
+def moe_align_block_size_stage4(
+    topk_ids_ptr,
+    sorted_token_ids_ptr,
+    expert_ids_ptr,
+    tokens_cnts_ptr,
+    cumsum_ptr,
+    num_experts: tl.constexpr,
+    block_size: tl.constexpr,
+    numel: tl.constexpr,
+    tokens_per_thread: tl.constexpr,
+):
+    pid = tl.program_id(0)
+    start_idx = tl.load(cumsum_ptr + pid)
+    end_idx = tl.load(cumsum_ptr + pid + 1)
+
+    for i in range(start_idx, end_idx, block_size):
+        tl.store(expert_ids_ptr + i // block_size, pid)
+
+    start_idx = pid * tokens_per_thread
+    off_t = pid * num_experts
+
+    for i in range(start_idx, tl.minimum(start_idx + tokens_per_thread,
+                                         numel)):
+        expert_id = tl.load(topk_ids_ptr + i)
+        token_cnt = tl.load(tokens_cnts_ptr + off_t + expert_id)
+        rank_post_pad = token_cnt + tl.load(cumsum_ptr + expert_id)
+        tl.store(sorted_token_ids_ptr + rank_post_pad, i)
+        tl.store(tokens_cnts_ptr + off_t + expert_id, token_cnt + 1)
+
+
+# Triton implementation based on:
+# https://github.com/sgl-project/sglang/commit/ba5112ff691d791a9e38c6c71f59324a5fcb49d0
+def moe_align_block_size_triton(
+    topk_ids: torch.Tensor,
+    num_experts: int,
+    block_size: int,
+    sorted_token_ids: torch.Tensor,
+    expert_ids: torch.Tensor,
+    num_tokens_post_pad: torch.Tensor,
+) -> None:
+    numel = topk_ids.numel()
+    grid = (num_experts, )
+    tokens_cnts = torch.zeros((num_experts + 1, num_experts),
+                              dtype=torch.int32,
+                              device=topk_ids.device)
+    cumsum = torch.zeros((num_experts + 1, ),
+                         dtype=torch.int32,
+                         device=topk_ids.device)
+    tokens_per_thread = ceil_div(numel, num_experts)
+
+    moe_align_block_size_stage1[grid](
+        topk_ids,
+        tokens_cnts,
+        num_experts,
+        numel,
+        tokens_per_thread,
+    )
+    moe_align_block_size_stage2[grid](
+        tokens_cnts,
+        num_experts,
+    )
+    moe_align_block_size_stage3[(1, )](
+        num_tokens_post_pad,
+        tokens_cnts,
+        cumsum,
+        num_experts,
+        block_size,
+    )
+    moe_align_block_size_stage4[grid](
+        topk_ids,
+        sorted_token_ids,
+        expert_ids,
+        tokens_cnts,
+        cumsum,
+        num_experts,
+        block_size,
+        numel,
+        tokens_per_thread,
+    )
+
+
 def moe_align_block_size(
         topk_ids: torch.Tensor, block_size: int,
         num_experts: int) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
@@ -457,8 +595,28 @@ def moe_align_block_size(
     num_tokens_post_pad = torch.empty((1),
                                       dtype=torch.int32,
                                       device=topk_ids.device)
-    ops.moe_align_block_size(topk_ids, num_experts, block_size, sorted_ids,
-                             expert_ids, num_tokens_post_pad)
+    if num_experts >= 224:
+        if envs.VLLM_ENABLE_MOE_ALIGN_BLOCK_SIZE_TRITON:
+            moe_align_block_size_triton(
+                topk_ids,
+                num_experts,
+                block_size,
+                sorted_ids,
+                expert_ids,
+                num_tokens_post_pad,
+            )
+        else:
+            ops.sgl_moe_align_block_size(
+                topk_ids,
+                num_experts,
+                block_size,
+                sorted_ids,
+                expert_ids,
+                num_tokens_post_pad,
+            )
+    else:
+        ops.moe_align_block_size(topk_ids, num_experts, block_size, sorted_ids,
+                                 expert_ids, num_tokens_post_pad)
     return sorted_ids, expert_ids, num_tokens_post_pad