Support multistream of MLA vector operations (#1135)

### What this PR does / why we need it?
Move all vector operations to a secondary stream, with the expected
overlaping being:
```
              | q_rmsnorm |                  | kv_norm_rope_cache |       | q_rope |
| matmul W_DQ | matmul W_DKV | index | index |    matmul W_UQ     | split | matmul W_KV_T |
```

Currently, the `IndexByTensor` operators introduced by computation of
`cos` and `sin` can't be offloaded to the secondary stream due to a
known bug of graph fusion optimization pass. So we instead keep it in
the main stream, only requires it be computed before `matmul W_UQ` to
avoid hindering later overlapping. The problem may be solved by later
optimization (#993), which hoists the computation of `cos` and `sin` up
to the first layer.

### Does this PR introduce _any_ user-facing change?
Controlled by `torchair_graph_config.enable_multistream_mla`, defaulted
to False.

### How was this patch tested?
Tested on 1x16 910 node, with tailored 2 layer DSKv2.

Signed-off-by: sdmyzlp <lrwei2@petalmail.com>

docs/source/user_guide/additional_config.md

@@ -39,6 +39,7 @@ The details of each config option are as follows:
 | Name | Type | Default | Description |
 | ---- | ---- | ------- | ----------- |
 | `enabled` | bool | `False` | Whether to enable torchair graph mode |
+| `enable_multistream_mla`| bool | `False` | Whether to put vector ops of MLA to another stream |
 | `enable_multistream_moe`| bool | `False` | Whether to enable multistream shared expert |
 | `enable_view_optimize` | bool | `True` | Whether to enable torchair view optimization |
 | `use_cached_graph` | bool | `False` | Whether to use cached graph |

tests/singlecard/test_ascend_config.py

@@ -59,6 +59,7 @@ def test_run_with_ascend_config():
             "graph_batch_sizes": [1, 2, 4, 8],
             "graph_batch_sizes_init": False,
             "enable_multistream_moe": True,
+            "enable_multistream_mla": True,
         },
         "ascend_scheduler_config": {
             "enabled": True,
@@ -79,6 +80,7 @@ def test_run_with_ascend_config():
             1, 2, 4, 8
         ]
         assert not ascend_config.torchair_graph_config.graph_batch_sizes_init
+        assert ascend_config.torchair_graph_config.enable_multistream_mla
         assert ascend_config.torchair_graph_config.enable_multistream_moe
         assert ascend_config.ascend_scheduler_config.enabled
         assert ascend_config.ascend_scheduler_config.enable_chunked_prefill

vllm_ascend/ascend_config.py

@@ -54,6 +54,8 @@ class TorchairGraphConfig:
             "graph_batch_sizes", [])
         self.graph_batch_sizes_init = torchair_graph_config.get(
             "graph_batch_sizes_init", False)
+        self.enable_multistream_mla = torchair_graph_config.get(
+            "enable_multistream_mla", False)
         self.enable_multistream_moe = torchair_graph_config.get(
             "enable_multistream_moe", False)
         self.enable_view_optimize = torchair_graph_config.get(

vllm_ascend/attention/mla_v1.py

@@ -19,6 +19,7 @@ from vllm_ascend.multistream.base import MSAttentionMetadataSplitConfig
 from vllm_ascend.multistream.context import get_multistream_comm_context
 from vllm_ascend.multistream.ms_split import model_input_split_v1_mla_attn
 from vllm_ascend.ops.attention import vanilla_chunked_prefill_mla
+from vllm_ascend.utils import npu_stream_switch, npu_wait_tensor
 
 if TYPE_CHECKING:
     from vllm.v1.core.sched.output import SchedulerOutput
@@ -481,6 +482,9 @@ class AscendMLAImpl(MLAAttentionImpl):
         ascend_config = get_ascend_config()
         self.torchair_graph_enabled = ascend_config.torchair_graph_config.enabled
         self.enable_kv_nz = ascend_config.torchair_graph_config.enable_kv_nz
+        self.enable_multistream_mla = \
+            ascend_config.torchair_graph_config.enable_multistream_mla
+
         # Adapt torch air graph mode with spec decoding.
         speculative_config = get_current_vllm_config().speculative_config
         if speculative_config is not None:
@@ -664,17 +668,20 @@ class AscendMLAImpl(MLAAttentionImpl):
         # npu_kv_rmsnorm_rope_cache needs [B, N, S, D]
         kv = kv.view(B, N, S, self.kv_lora_rank + self.qk_rope_head_dim)
         cache_mode = "PA_NZ" if self.enable_kv_nz else "PA"
-        k_pe, k_nope, _, _ = torch_npu.npu_kv_rmsnorm_rope_cache(
+        with npu_stream_switch("mla_secondary",
-            kv,
+                               0,
-            self.kv_a_layernorm.weight,
+                               enabled=self.enable_multistream_mla):
-            cos,
+            k_pe, k_nope, _, _ = torch_npu.npu_kv_rmsnorm_rope_cache(
-            sin,
+                kv,
-            slots.to(torch.int64),
+                self.kv_a_layernorm.weight,
-            kv_cache[1],
+                cos,
-            kv_cache[0],
+                sin,
-            epsilon=self.kv_a_layernorm.variance_epsilon,
+                slots.to(torch.int64),
-            cache_mode=cache_mode,
+                kv_cache[1],
-        )
+                kv_cache[0],
+                epsilon=self.kv_a_layernorm.variance_epsilon,
+                cache_mode=cache_mode,
+            )
         return k_pe, k_nope
 
     def exec_kv_prefill(
@@ -867,23 +874,38 @@ class AscendMLAImpl(MLAAttentionImpl):
         if has_decode:
             decode_k_nope = None
             assert attn_metadata.decode is not None
-            decode_ql_nope, decode_q_pe = \
-                self._q_proj_and_k_up_proj(decode_hs_or_q_c)
             if self.running_in_graph:
                 seq_len = self.rotary_emb.max_position_embeddings
                 cos = self.rotary_emb.cos_cached[:seq_len].to(
-                    dtype=decode_q_pe.dtype)
+                    dtype=decode_hs_or_q_c.dtype)
                 sin = self.rotary_emb.sin_cached[:seq_len].to(
-                    dtype=decode_q_pe.dtype)
+                    dtype=decode_hs_or_q_c.dtype)
                 cos = cos[attn_metadata.decode.input_positions]
                 sin = sin[attn_metadata.decode.input_positions]
                 cos = cos[:, None, None, :]
                 sin = sin[:, None, None, :]
-
+                # Without explicitly controlling the order, IndexByTensor operations
-                decode_q_pe = self.rope_single(decode_q_pe, cos, sin)
+                # would be placed after `matmul W_KV_T` hindering the overlapping of
+                # KvRmsNormRopeCache and SingleRope.
+                npu_wait_tensor(decode_hs_or_q_c,
+                                cos,
+                                enabled=self.enable_multistream_mla)
+                npu_wait_tensor(decode_hs_or_q_c,
+                                sin,
+                                enabled=self.enable_multistream_mla)
+            decode_ql_nope, decode_q_pe = \
+                self._q_proj_and_k_up_proj(decode_hs_or_q_c)
+            if self.running_in_graph:
                 decode_k_pe, decode_k_nope = self.exec_kv(
                     hidden_states_or_kv_c_normed, cos, sin, kv_cache,
                     attn_metadata.slot_mapping)
+                with npu_stream_switch("mla_secondary",
+                                       0,
+                                       enabled=self.enable_multistream_mla):
+                    npu_wait_tensor(decode_q_pe,
+                                    decode_k_pe,
+                                    enabled=self.enable_multistream_mla)
+                    decode_q_pe = self.rope_single(decode_q_pe, cos, sin)
             else:
                 decode_q_pe[...], decode_k_pe[...] = self.rotary_emb(
                     attn_metadata.decode.input_positions,

vllm_ascend/models/deepseek_v2.py

@@ -71,7 +71,8 @@ from vllm_ascend.distributed.parallel_state import get_ep_group
 from vllm_ascend.ops.fused_moe import AscendFusedMoE
 from vllm_ascend.quantization.quant_config import AscendLinearMethod
 from vllm_ascend.quantization.w8a8_dynamic import AscendW8A8DynamicLinearMethod
-from vllm_ascend.utils import dispose_tensor
+from vllm_ascend.utils import (dispose_tensor, npu_stream_switch,
+                               npu_wait_tensor)
 
 VLLM_ENABLE_MC2: bool = envs_ascend.VLLM_ENABLE_MC2
 
@@ -496,6 +497,8 @@ class CustomDeepseekV2MLAAttention(DeepseekV2MLAAttention):
 
         ascend_config = get_ascend_config()
         self.torchair_graph_enabled = ascend_config.torchair_graph_config.enabled
+        self.enable_multistream_mla = \
+            ascend_config.torchair_graph_config.enable_multistream_mla
 
     def forward(
             self,
@@ -505,7 +508,14 @@ class CustomDeepseekV2MLAAttention(DeepseekV2MLAAttention):
             attn_metadata: Optional[AttentionMetadata] = None) -> torch.Tensor:
         if self.q_lora_rank is not None:
             ckq = self.q_a_proj(hidden_states)[0]
-            hidden_states_or_q_c = self.q_a_layernorm(ckq)
+            use_multistream_mla = (self.enable_multistream_mla
+                                   and attn_metadata is not None
+                                   and attn_metadata.num_decodes > 0)
+            npu_wait_tensor(hidden_states, ckq, enabled=use_multistream_mla)
+            with npu_stream_switch("mla_secondary",
+                                   0,
+                                   enabled=use_multistream_mla):
+                hidden_states_or_q_c = self.q_a_layernorm(ckq)
         else:
             hidden_states_or_q_c = hidden_states
         if self.torchair_graph_enabled: