Optimize MLA/GQA/MQA Triton decoding (#1138)

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

python/sglang/srt/layers/decode_attention.py

@@ -58,7 +58,6 @@ def _fwd_kernel_stage1(
     att_stride_h,
     kv_group_num: tl.constexpr,
     BLOCK_DMODEL: tl.constexpr,
-    BLOCK_DPE: tl.constexpr,
     BLOCK_N: tl.constexpr,
     logit_cap: tl.constexpr,
 ):
@@ -78,10 +77,6 @@ def _fwd_kernel_stage1(
 
     off_q = cur_batch * stride_qbs + cur_head * stride_qh + offs_d
 
-    if BLOCK_DPE > 0:
-        offs_dpe = BLOCK_DMODEL + tl.arange(0, BLOCK_DPE)
-        off_qpe = cur_batch * stride_qbs + cur_head * stride_qh + offs_dpe
-
     offs_n = start_n * BLOCK_N + tl.arange(0, BLOCK_N)
 
     block_stard_index = start_n * BLOCK_N
@@ -106,19 +101,6 @@ def _fwd_kernel_stage1(
             other=0.0,
         ).to(REDUCE_TRITON_TYPE)
         att_value = tl.sum(q[None, :] * k, 1)
-        if BLOCK_DPE > 0:
-            qpe = tl.load(Q + off_qpe + start_mark).to(REDUCE_TRITON_TYPE)
-            offs_buf_kpe = (
-                k_loc[:, None] * stride_buf_kbs
-                + cur_kv_head * stride_buf_kh
-                + offs_dpe[None, :]
-            )
-            kpe = tl.load(
-                K_Buffer + offs_buf_kpe,
-                mask=offs_n_new[:, None] < cur_batch_end_index,
-                other=0.0,
-            ).to(REDUCE_TRITON_TYPE)
-            att_value += tl.sum(qpe[None, :] * kpe, 1)
         att_value *= sm_scale
 
         if logit_cap > 0:
@@ -214,14 +196,7 @@ def _decode_att_m_fwd(
     # shape constraints
     Lq, Lk = q.shape[-1], k_buffer.shape[-1]
     assert Lq == Lk
-    assert Lk in {16, 32, 64, 128, 256, 576}
-
-    if Lk == 576:
-        BLOCK_DMODEL = 512
-        BLOCK_DPE = 64
-    else:
-        BLOCK_DMODEL = Lk
-        BLOCK_DPE = 0
+    assert Lk in {16, 32, 64, 128, 256}
 
     batch, head_num = B_req_idx.shape[0], q.shape[1]
 
@@ -249,8 +224,7 @@ def _decode_att_m_fwd(
         k_buffer.stride(1),
         att_out.stride(0),
         kv_group_num=kv_group_num,
-        BLOCK_DMODEL=BLOCK_DMODEL,
-        BLOCK_DPE=BLOCK_DPE,
+        BLOCK_DMODEL=Lk,
         BLOCK_N=BLOCK,
         logit_cap=logit_cap,
         num_warps=num_warps,
@@ -296,6 +270,293 @@ def _decode_softmax_reducev_fwd(
     )
 
 
+@triton.jit
+def _fwd_grouped_kernel_stage1(
+    Q,
+    K_Buffer,
+    sm_scale,
+    Req_to_tokens,
+    B_req_idx,
+    B_Start_Loc,
+    B_Seqlen,
+    Att_Out,
+    stride_req_to_tokens_b,
+    stride_qbs,
+    stride_qh,
+    stride_buf_kbs,
+    stride_buf_kh,
+    att_stride_h,
+    kv_group_num: tl.constexpr,
+    q_head_num: tl.constexpr,
+    BLOCK_DMODEL: tl.constexpr,
+    BLOCK_DPE: tl.constexpr,
+    BLOCK_N: tl.constexpr,
+    BLOCK_H: tl.constexpr,
+    logit_cap: tl.constexpr,
+):
+    cur_batch = tl.program_id(0)
+    cur_kv_head = tl.program_id(1)
+    start_n = tl.program_id(2)
+
+    cur_head = cur_kv_head * kv_group_num + tl.arange(0, BLOCK_H)
+    mask_h = cur_head < (cur_kv_head + 1) * kv_group_num
+    mask_h = mask_h & (cur_head < q_head_num)
+
+    offs_d = tl.arange(0, BLOCK_DMODEL)
+    cur_batch_seq_len = tl.load(B_Seqlen + cur_batch)
+    cur_batch_in_all_start_index = tl.load(B_Start_Loc + cur_batch)
+    cur_batch_req_idx = tl.load(B_req_idx + cur_batch)
+
+    cur_batch_start_index = 0
+    cur_batch_end_index = cur_batch_seq_len
+
+    offs_q = cur_batch * stride_qbs + cur_head[:, None] * stride_qh + offs_d[None, :]
+
+    if BLOCK_DPE > 0:
+        offs_dpe = BLOCK_DMODEL + tl.arange(0, BLOCK_DPE)
+        off_qpe = (
+            cur_batch * stride_qbs + cur_head[:, None] * stride_qh + offs_dpe[None, :]
+        )
+
+    offs_n = start_n * BLOCK_N + tl.arange(0, BLOCK_N)
+
+    block_stard_index = start_n * BLOCK_N
+    block_mask = tl.where(block_stard_index < cur_batch_seq_len, 1, 0)
+
+    for start_mark in range(0, block_mask, 1):
+        q = tl.load(Q + offs_q + start_mark, mask=mask_h[:, None]).to(
+            REDUCE_TRITON_TYPE
+        )
+        offs_n_new = cur_batch_start_index + offs_n
+        k_loc = tl.load(
+            Req_to_tokens + stride_req_to_tokens_b * cur_batch_req_idx + offs_n_new,
+            mask=offs_n_new < cur_batch_end_index,
+            other=0,
+        )
+        offs_buf_k = (
+            k_loc[None, :] * stride_buf_kbs
+            + cur_kv_head * stride_buf_kh
+            + offs_d[:, None]
+        )
+        k = tl.load(
+            K_Buffer + offs_buf_k,
+            mask=offs_n_new[None, :] < cur_batch_end_index,
+            other=0.0,
+        ).to(REDUCE_TRITON_TYPE)
+        qk = tl.dot(q, k)
+        if BLOCK_DPE > 0:
+            qpe = tl.load(Q + off_qpe + start_mark, mask=mask_h[:, None]).to(
+                REDUCE_TRITON_TYPE
+            )
+            offs_buf_kpe = (
+                k_loc[None, :] * stride_buf_kbs
+                + cur_kv_head * stride_buf_kh
+                + offs_dpe[:, None]
+            )
+            kpe = tl.load(
+                K_Buffer + offs_buf_kpe,
+                mask=offs_n_new[None, :] < cur_batch_end_index,
+                other=0.0,
+            ).to(REDUCE_TRITON_TYPE)
+            qk += tl.dot(qpe, kpe)
+        qk *= sm_scale
+
+        if logit_cap > 0:
+            qk = logit_cap * tanh(qk / logit_cap)
+
+        offs_o = cur_head[:, None] * att_stride_h + (
+            cur_batch_in_all_start_index + offs_n[None, :]
+        )
+
+        tl.store(
+            Att_Out + offs_o,
+            qk,
+            mask=mask_h[:, None] & (offs_n_new[None, :] < cur_batch_end_index),
+        )
+
+
+@triton.jit
+def _fwd_grouped_kernel_stage2(
+    Logics,
+    V_Buffer,
+    Out,
+    Req_to_tokens,
+    B_req_idx,
+    B_Start_Loc,
+    B_Seqlen,
+    stride_logic_h,
+    stride_buf_vbs,
+    stride_buf_vh,
+    stride_obs,
+    stride_oh,
+    stride_req_to_token_b,
+    kv_group_num: tl.constexpr,
+    q_head_num: tl.constexpr,
+    BLOCK_DMODEL: tl.constexpr,
+    BLOCK_N: tl.constexpr,
+    BLOCK_H: tl.constexpr,
+):
+    cur_batch = tl.program_id(0)
+    cur_kv_head = tl.program_id(1)
+
+    cur_head = cur_kv_head * kv_group_num + tl.arange(0, BLOCK_H)
+    mask_h = cur_head < (cur_kv_head + 1) * kv_group_num
+    mask_h = mask_h & (cur_head < q_head_num)
+
+    cur_batch_seq_len = tl.load(B_Seqlen + cur_batch)
+    cur_batch_start_loc = tl.load(B_Start_Loc + cur_batch)
+    cur_batch_req_idx = tl.load(B_req_idx + cur_batch)
+
+    offs_n = tl.arange(0, BLOCK_N)
+    offs_d = tl.arange(0, BLOCK_DMODEL)
+
+    offs_buf_v = cur_kv_head * stride_buf_vh + offs_d[None, :]
+    v_ptrs = V_Buffer + offs_buf_v
+
+    e_max = tl.zeros([BLOCK_H], dtype=tl.float32) - float("inf")
+    e_sum = tl.zeros([BLOCK_H], dtype=tl.float32)
+    acc = tl.zeros([BLOCK_H, BLOCK_DMODEL], dtype=tl.float32)
+
+    for start_n in range(0, cur_batch_seq_len, BLOCK_N):
+        start_n = tl.multiple_of(start_n, BLOCK_N)
+        v_index = tl.load(
+            Req_to_tokens
+            + cur_batch_req_idx * stride_req_to_token_b
+            + (start_n + offs_n),
+            mask=(start_n + offs_n) < cur_batch_seq_len,
+            other=0,
+        )
+
+        offs_qk = cur_head[:, None] * stride_logic_h + (
+            cur_batch_start_loc + start_n + offs_n[None, :]
+        )
+
+        qk = tl.load(
+            Logics + offs_qk,
+            mask=mask_h[:, None] & (start_n + offs_n[None, :] < cur_batch_seq_len),
+            other=float("-inf"),
+        )
+
+        n_e_max = tl.maximum(tl.max(qk, 1), e_max)
+        old_scale = tl.exp(e_max - n_e_max)
+        p = tl.exp(qk - n_e_max[:, None])
+        e_sum = e_sum * old_scale + tl.sum(p, 1)
+        v = tl.load(v_ptrs + v_index[:, None] * stride_buf_vbs)
+        p = p.to(v.dtype)
+        acc = acc * old_scale[:, None] + tl.dot(p, v)
+        e_max = n_e_max
+
+    acc = acc / e_sum[:, None]
+    off_o = cur_batch * stride_obs + cur_head[:, None] * stride_oh + offs_d[None, :]
+    out_ptrs = Out + off_o
+    tl.store(out_ptrs, acc, mask=mask_h[:, None])
+
+
+def _decode_grouped_att_m_fwd(
+    q,
+    k_buffer,
+    att_out,
+    Req_to_tokens,
+    B_req_idx,
+    B_Start_Loc,
+    B_Seqlen,
+    max_len_in_batch,
+    sm_scale,
+    logit_cap,
+):
+    BLOCK = 32
+    # shape constraints
+    Lq, Lk = q.shape[-1], k_buffer.shape[-1]
+    assert Lq == Lk
+    assert Lk in {16, 32, 64, 128, 256, 576}
+
+    if Lk == 576:
+        BLOCK_DMODEL = 512
+        BLOCK_DPE = 64
+    else:
+        BLOCK_DMODEL = Lk
+        BLOCK_DPE = 0
+
+    batch, head_num = B_req_idx.shape[0], q.shape[1]
+    kv_group_num = q.shape[1] // k_buffer.shape[1]
+
+    BLOCK_H = max(16, triton.next_power_of_2(kv_group_num))
+    grid = (
+        batch,
+        triton.cdiv(head_num, min(BLOCK_H, kv_group_num)),
+        triton.cdiv(max_len_in_batch, BLOCK),
+    )
+
+    num_warps = 4
+
+    _fwd_grouped_kernel_stage1[grid](
+        q,
+        k_buffer,
+        sm_scale,
+        Req_to_tokens,
+        B_req_idx,
+        B_Start_Loc,
+        B_Seqlen,
+        att_out,
+        Req_to_tokens.stride(0),
+        q.stride(0),
+        q.stride(1),
+        k_buffer.stride(0),
+        k_buffer.stride(1),
+        att_out.stride(0),
+        kv_group_num=kv_group_num,
+        q_head_num=head_num,
+        BLOCK_DMODEL=BLOCK_DMODEL,
+        BLOCK_DPE=BLOCK_DPE,
+        BLOCK_N=BLOCK,
+        BLOCK_H=BLOCK_H,
+        logit_cap=logit_cap,
+        num_warps=num_warps,
+        num_stages=1,
+    )
+
+
+def _decode_grouped_softmax_reducev_fwd(
+    logics,
+    v_buffer,
+    o,
+    req_to_tokens,
+    b_req_idx,
+    b_start_loc,
+    b_seq_len,
+):
+    BLOCK = 128
+    batch, head_num = b_seq_len.shape[0], logics.shape[0]
+    kv_group_num = logics.shape[0] // v_buffer.shape[1]
+    BLOCK_H = max(16, triton.next_power_of_2(kv_group_num))
+    grid = (batch, triton.cdiv(head_num, min(BLOCK_H, kv_group_num)), 1)
+
+    num_warps = 8
+
+    _fwd_grouped_kernel_stage2[grid](
+        logics,
+        v_buffer,
+        o,
+        req_to_tokens,
+        b_req_idx,
+        b_start_loc,
+        b_seq_len,
+        logics.stride(0),
+        v_buffer.stride(0),
+        v_buffer.stride(1),
+        o.stride(0),
+        o.stride(1),
+        req_to_tokens.stride(0),
+        kv_group_num=kv_group_num,
+        q_head_num=head_num,
+        BLOCK_DMODEL=v_buffer.shape[-1],
+        BLOCK_N=BLOCK,
+        BLOCK_H=BLOCK_H,
+        num_warps=num_warps,
+        num_stages=1,
+    )
+
+
 def decode_attention_fwd(
     q,
     k_buffer,
@@ -316,24 +577,51 @@ def decode_attention_fwd(
             (q.shape[-2], total_num_tokens), dtype=REDUCE_TORCH_TYPE, device="cuda"
         )
 
-    _decode_att_m_fwd(
-        q,
-        k_buffer,
-        att_m,
-        req_to_token,
-        b_req_idx,
-        b_start_loc,
-        b_seq_len,
-        max_len_in_batch,
-        sm_scale,
-        logit_cap,
-    )
-    _decode_softmax_reducev_fwd(
-        att_m,
-        v_buffer,
-        o,
-        req_to_token,
-        b_req_idx,
-        b_start_loc,
-        b_seq_len,
-    )
+    kv_group_num = q.shape[1] // v_buffer.shape[1]
+
+    if kv_group_num == 1:
+        # MHA
+        _decode_att_m_fwd(
+            q,
+            k_buffer,
+            att_m,
+            req_to_token,
+            b_req_idx,
+            b_start_loc,
+            b_seq_len,
+            max_len_in_batch,
+            sm_scale,
+            logit_cap,
+        )
+        _decode_softmax_reducev_fwd(
+            att_m,
+            v_buffer,
+            o,
+            req_to_token,
+            b_req_idx,
+            b_start_loc,
+            b_seq_len,
+        )
+    else:
+        # GQA/MQA/MLA
+        _decode_grouped_att_m_fwd(
+            q,
+            k_buffer,
+            att_m,
+            req_to_token,
+            b_req_idx,
+            b_start_loc,
+            b_seq_len,
+            max_len_in_batch,
+            sm_scale,
+            logit_cap,
+        )
+        _decode_grouped_softmax_reducev_fwd(
+            att_m,
+            v_buffer,
+            o,
+            req_to_token,
+            b_req_idx,
+            b_start_loc,
+            b_seq_len,
+        )