[Feat][Graph] Support FULL_DECODE_ONLY mode for GQA/MHA models (#2128)

Note: This depends on [vLLM
#25161](https://github.com/vllm-project/vllm/pull/25161) and the
torch\_npu release from September 30.

### What this PR does / why we need it?
This pull request adds `FULL_DECODE_ONLY` mode for GQA/MHA models (MLA
models like DeepSeek V3/R1 are not included). Key improvements include:

* **Reduced dispatch latency:** By replaying the entire model execution
graph at once, we cut overhead compared with multiple smaller replays.
* **Stabilized multi-device performance:** Captureing the whole model as
one static graph also mitigates the dispatch fluctuations across
devices.
* **Stream/resource savings:** Consolidating graph captures frees up
streams, allowing more graphs to be captured.

**Known issues:**

1. `_npu_paged_attention` currently manages its own workspace in
`torch_npu`, which can deadlock when synchronizing during graph replay —
we’re working on a fix.

There may be other corner cases. This PR is the first in a planned
series; we’ll continue to iterate and address remaining issues in
follow-ups.

This is essentially a port of #1503 and #1677, but includes two major
changes:

1. Let `graph_dispatcher` decide the graph mode instead of hard-coding
it in the backend, which decouples Full Graph and Piecewise Graph and
could make it possible to remove dynamo.
2. Adapt to the new `attn_group` logic, but leave a small hack in
`update_graph_params`; multi-attention models may or may not be fully
supported yet.

### Does this PR introduce _any_ user-facing change?
```python
compilation_config={
    "cudagraph_mode": "FULL_DECODE_ONLY",
},
```

### How was this patch tested?
Tests included.


- vLLM version: v0.10.2
- vLLM main:
9607d5eb44

---------

Signed-off-by: Yizhou Liu <liu_yizhou@outlook.com>

vllm_ascend/utils.py

@@ -22,6 +22,7 @@ import functools
 import math
 import os
 from contextlib import contextmanager, nullcontext
+from dataclasses import dataclass
 from enum import Enum
 from threading import Lock
 from typing import TYPE_CHECKING, Any, List, Optional, Tuple, Union
@@ -634,3 +635,34 @@ def npu_stream_switch(target_stream: torch.npu.Stream,
         return nullcontext()
     assert target_stream is not None
     return torch.npu.stream(target_stream)
+
+
+@dataclass
+class GraphParams:
+    events: dict[int, list[torch.npu.ExternalEvent]]
+    workspaces: dict[int, torch.Tensor]
+    handles: dict[int, list[torch_npu._C._NPUTaskGroupHandle]]
+    attn_params: dict[int, list[tuple]]
+
+
+_graph_params: Optional[GraphParams] = None
+
+
+def set_graph_params(aclgraph_capture_sizes: set[int]):
+    global _graph_params
+    if _graph_params is not None:
+        raise ValueError("Graph parameters have already been set!")
+    _graph_params = GraphParams(
+        {size: []
+         for size in aclgraph_capture_sizes},
+        {size: None
+         for size in aclgraph_capture_sizes},
+        {size: []
+         for size in aclgraph_capture_sizes},
+        {size: []
+         for size in aclgraph_capture_sizes},
+    )
+
+
+def get_graph_params():
+    return _graph_params