Triton MLA perf fixes (#33529)

Signed-off-by: Koushik Dutta <koushd@gmail.com>
Co-authored-by: root <root@ubuntu-nvidia.localdomain>
Co-authored-by: Cyrus Leung <tlleungac@connect.ust.hk>

vllm/v1/attention/backends/mla/triton_mla.py

@@ -14,6 +14,7 @@ from vllm.model_executor.layers.attention.mla_attention import (
     MLACommonMetadata,
 )
 from vllm.platforms.interface import DeviceCapability
+from vllm.triton_utils import triton
 from vllm.utils.torch_utils import is_quantized_kv_cache
 from vllm.v1.attention.backend import (
     AttentionLayer,
@@ -115,6 +116,8 @@ class TritonMLAImpl(MLACommonImpl[MLACommonMetadata]):
         if is_quantized_kv_cache(self.kv_cache_dtype):
             self.supports_quant_query_input = False
 
+        self._sm_count = torch.cuda.get_device_properties(0).multi_processor_count
+
     def _flash_attn_varlen_diff_headdims(
         self, q, k, v, return_softmax_lse=False, softmax_scale=None, **kwargs
     ):
@@ -149,7 +152,24 @@ class TritonMLAImpl(MLACommonImpl[MLACommonMetadata]):
         lse = torch.zeros(B, q_num_heads, dtype=q.dtype, device=q.device)
 
         # For batch invariance, use only 1 split to ensure deterministic reduction
-        num_kv_splits = 1 if envs.VLLM_BATCH_INVARIANT else 4
+        if envs.VLLM_BATCH_INVARIANT:
+            num_kv_splits = 1
+        else:
+            # Minimum work per split
+            # hardware dependent
+            min_work_per_split = 512
+
+            ideal_splits = max(1, attn_metadata.max_seq_len // min_work_per_split)
+
+            # use power of 2 to avoid excessive kernel instantiations
+            ideal_splits = triton.next_power_of_2(ideal_splits)
+
+            # Calculate SM-based maximum splits with occupancy multiplier
+            # 2-4x allows multiple blocks per SM for latency hiding
+            # hardware dependent
+            occupancy_multiplier = 2
+            max_splits = self._sm_count * occupancy_multiplier
+            num_kv_splits = min(ideal_splits, max_splits)
 
         # TODO(lucas) Allocate ahead of time
         attn_logits = torch.empty(
@@ -186,6 +206,7 @@ class TritonMLAImpl(MLACommonImpl[MLACommonMetadata]):
             PAGE_SIZE,
             k_scale=layer._k_scale,
             v_scale=layer._k_scale,
+            is_mla=True,
         )
 
         return o, lse

vllm/v1/attention/ops/triton_decode_attention.py

@@ -291,6 +291,7 @@ def _fwd_grouped_kernel_stage1(
     logit_cap: tl.constexpr,
     Lk: tl.constexpr,
     Lv: tl.constexpr,
+    IS_MLA: tl.constexpr = False,
 ):
     cur_batch = tl.program_id(0)
     cur_head_id = tl.program_id(1)
@@ -310,7 +311,12 @@ def _fwd_grouped_kernel_stage1(
     cur_batch_req_idx = cur_batch
 
     offs_q = cur_batch * stride_qbs + cur_head[:, None] * stride_qh + offs_d[None, :]
-    q = tl.load(Q + offs_q, mask=(mask_h[:, None]) & (mask_d[None, :]), other=0.0)
+    q = tl.load(
+        Q + offs_q,
+        mask=(mask_h[:, None]) & (mask_d[None, :]),
+        other=0.0,
+        cache_modifier=".ca",
+    )
 
     if BLOCK_DPE > 0:
         offs_dpe = BLOCK_DMODEL + tl.arange(0, BLOCK_DPE)
@@ -319,7 +325,10 @@ def _fwd_grouped_kernel_stage1(
             cur_batch * stride_qbs + cur_head[:, None] * stride_qh + offs_dpe[None, :]
         )
         qpe = tl.load(
-            Q + off_qpe, mask=(mask_h[:, None]) & (mask_dpe[None, :]), other=0.0
+            Q + off_qpe,
+            mask=(mask_h[:, None]) & (mask_dpe[None, :]),
+            other=0.0,
+            cache_modifier=".ca",
         )
 
     kv_len_per_split = tl.cdiv(cur_batch_seq_len, NUM_KV_SPLITS)
@@ -331,9 +340,14 @@ def _fwd_grouped_kernel_stage1(
     acc = tl.zeros([BLOCK_H, BLOCK_DV], dtype=tl.float32)
 
     if split_kv_end > split_kv_start:
+        base_offs_k = cur_kv_head * stride_buf_kh + offs_d[:, None]
+        base_offs_v = cur_kv_head * stride_buf_vh + offs_dv[None, :]
+        if BLOCK_DPE > 0:
+            base_offs_kpe = cur_kv_head * stride_buf_kh + offs_dpe[:, None]
+
         ks = tl.load(k_scale)
         vs = tl.load(v_scale)
-        for start_n in range(split_kv_start, split_kv_end, BLOCK_N):
+        for start_n in tl.range(split_kv_start, split_kv_end, BLOCK_N):
             offs_n = start_n + tl.arange(0, BLOCK_N)
             kv_page_number = tl.load(
                 Req_to_tokens
@@ -341,31 +355,29 @@ def _fwd_grouped_kernel_stage1(
                 + offs_n // PAGE_SIZE,
                 mask=offs_n < split_kv_end,
                 other=0,
+                cache_modifier=".ca",
             )
             kv_loc = kv_page_number * PAGE_SIZE + offs_n % PAGE_SIZE
-            offs_buf_k = (
-                kv_loc[None, :] * stride_buf_kbs
-                + cur_kv_head * stride_buf_kh
-                + offs_d[:, None]
-            )
+
+            # explicitly facilitate overlapping load/compute
+            offs_buf_k = kv_loc[None, :] * stride_buf_kbs + base_offs_k
             k = tl.load(
                 K_Buffer + offs_buf_k,
                 mask=(offs_n[None, :] < split_kv_end) & (mask_d[:, None]),
                 other=0.0,
+                cache_modifier=".cg",
             )
+
             if k.dtype.is_fp8():
                 k = (k.to(tl.float32) * ks).to(q.dtype)
             qk = tl.dot(q, k.to(q.dtype))
             if BLOCK_DPE > 0:
-                offs_buf_kpe = (
-                    kv_loc[None, :] * stride_buf_kbs
-                    + cur_kv_head * stride_buf_kh
-                    + offs_dpe[:, None]
-                )
+                offs_buf_kpe = kv_loc[None, :] * stride_buf_kbs + base_offs_kpe
                 kpe = tl.load(
                     K_Buffer + offs_buf_kpe,
                     mask=(offs_n[None, :] < split_kv_end) & (mask_dpe[:, None]),
                     other=0.0,
+                    cache_modifier=".cg",
                 )
                 if kpe.dtype.is_fp8():
                     kpe = (kpe.to(tl.float32) * ks).to(qpe.dtype)
@@ -379,18 +391,20 @@ def _fwd_grouped_kernel_stage1(
                 mask_h[:, None] & (offs_n[None, :] < split_kv_end), qk, float("-inf")
             )
 
-            offs_buf_v = (
-                kv_loc[:, None] * stride_buf_vbs
-                + cur_kv_head * stride_buf_vh
-                + offs_dv[None, :]
-            )
-            v = tl.load(
-                V_Buffer + offs_buf_v,
-                mask=(offs_n[:, None] < split_kv_end) & (mask_dv[None, :]),
-                other=0.0,
-            )
-            if v.dtype.is_fp8():
-                v = (v.to(tl.float32) * vs).to(q.dtype)
+            if not IS_MLA:
+                offs_buf_v = kv_loc[:, None] * stride_buf_vbs + base_offs_v
+                v = tl.load(
+                    V_Buffer + offs_buf_v,
+                    mask=(offs_n[:, None] < split_kv_end) & (mask_dv[None, :]),
+                    other=0.0,
+                )
+                if v.dtype.is_fp8():
+                    v = (v.to(tl.float32) * vs).to(q.dtype)
+            else:
+                # MLA uses a single c_kv.
+                # loading the same c_kv to interpret it as v is not necessary.
+                # transpose the existing c_kv (aka k) for the dot product.
+                v = tl.trans(k)
 
             n_e_max = tl.maximum(tl.max(qk, 1), e_max)
             re_scale = tl.exp(e_max - n_e_max)
@@ -441,7 +455,10 @@ def _decode_grouped_att_m_fwd(
     logit_cap,
     k_scale,
     v_scale,
+    is_mla=False,
 ):
+    # with is_mla there is only a single c_kv in smem.
+    # could increase BLOCK or num_stages.
     BLOCK = 32
     Lk = k_buffer.shape[-1]
     Lv = v_buffer.shape[-1]
@@ -514,6 +531,7 @@ def _decode_grouped_att_m_fwd(
         num_stages=num_stages,
         Lk=Lk,
         Lv=Lv,
+        IS_MLA=is_mla,
         **extra_kargs,
     )
 
@@ -673,6 +691,7 @@ def decode_attention_fwd_grouped(
     logit_cap=0.0,
     k_scale=None,
     v_scale=None,
+    is_mla=False,
 ):
     _decode_grouped_att_m_fwd(
         q,
@@ -687,6 +706,7 @@ def decode_attention_fwd_grouped(
         logit_cap,
         k_scale,
         v_scale,
+        is_mla=is_mla,
     )
     _decode_softmax_reducev_fwd(
         attn_logits, q, o, lse, v_buffer, b_seq_len, num_kv_splits
@@ -708,6 +728,7 @@ def decode_attention_fwd(
     logit_cap=0.0,
     k_scale=None,
     v_scale=None,
+    is_mla=False,
 ):
     assert num_kv_splits == attn_logits.shape[2]
 
@@ -753,4 +774,5 @@ def decode_attention_fwd(
             logit_cap,
             k_scale,
             v_scale,
+            is_mla=is_mla,
         )