Offload vector operations of MLA to another stream

With the expected overlaping being:
```
| cos/sin |   | q_rmsnorm |  | kv_norm_rope_cache |       | q_rope |
| matmul W_DQ | matmul W_DKV |    matmul W_UQ     | split | matmul W_KV_T |
```
Controlled by `torchair_graph_config.enable_multistream_mla`, defaulted
to False.

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

Author: sdmyzlp

docs/source/user_guide/additional_config.md

@@ -38,6 +38,7 @@ 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

@@ -53,6 +53,8 @@ def __init__(self, torchair_graph_config):
             "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

@@ -17,6 +17,7 @@
 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
@@ -461,6 +462,8 @@ def __init__(
 
         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 _v_up_proj_and_o_proj(self, x):
         # Convert from (B, N, L) to (N, B, L)
@@ -626,17 +629,18 @@ def exec_kv(
         kv = self.kv_a_proj_with_mqa(hidden_states)[0]
         # 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)
-        k_pe, k_nope, _, _ = torch_npu.npu_kv_rmsnorm_rope_cache(
-            kv,
-            self.kv_a_layernorm.weight,
-            cos,
-            sin,
-            slots.to(torch.int64),
-            kv_cache[1],
-            kv_cache[0],
-            epsilon=self.kv_a_layernorm.variance_epsilon,
-            cache_mode="PA",
-        )
+        with npu_stream_switch("mla_secondary", 0, self.enable_multistream_mla):
+            k_pe, k_nope, _, _ = torch_npu.npu_kv_rmsnorm_rope_cache(
+                kv,
+                self.kv_a_layernorm.weight,
+                cos,
+                sin,
+                slots.to(torch.int64),
+                kv_cache[1],
+                kv_cache[0],
+                epsilon=self.kv_a_layernorm.variance_epsilon,
+                cache_mode="PA",
+            )
         return k_pe, k_nope
 
     def rope_single(
@@ -769,20 +773,25 @@ def forward(
             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)
-                sin = self.rotary_emb.sin_cached[:seq_len].to(
-                    dtype=decode_q_pe.dtype)
-                cos = cos[attn_metadata.decode.input_positions]
-                sin = sin[attn_metadata.decode.input_positions]
-                cos = cos[:, None, None, :]
-                sin = sin[:, None, None, :]
-
-                decode_q_pe = self.rope_single(decode_q_pe, cos, sin)
+                with npu_stream_switch("mla_secondary", 0,
+                                       self.enable_multistream_mla):
+                    seq_len = self.rotary_emb.max_position_embeddings
+                    cos = self.rotary_emb.cos_cached[:seq_len].to(
+                        dtype=decode_q_pe.dtype)
+                    sin = self.rotary_emb.sin_cached[:seq_len].to(
+                        dtype=decode_q_pe.dtype)
+                    cos = cos[attn_metadata.decode.input_positions]
+                    sin = sin[attn_metadata.decode.input_positions]
+                    cos = cos[:, None, None, :]
+                    sin = sin[:, None, None, :]
                 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,
+                                       self.enable_multistream_mla):
+                    npu_wait_tensor(decode_q_pe, decode_k_pe,
+                                    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

@@ -70,7 +70,8 @@
 from vllm_ascend.distributed.parallel_state import get_ep_group
 from vllm_ascend.ops.fused_moe import AscendFusedMoE
 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
 
@@ -488,6 +489,8 @@ def __init__(
 
         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,
@@ -497,7 +500,12 @@ def forward(
             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, use_multistream_mla)
+            with npu_stream_switch("mla_secondary", 0, 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: