support aclgraph (#426)

<!--  Thanks for sending a pull request!

BEFORE SUBMITTING, PLEASE READ
https://docs.vllm.ai/en/latest/contributing/overview.html

-->
### What this PR does / why we need it?
<!--
- Please clarify what changes you are proposing. The purpose of this
section is to outline the changes and how this PR fixes the issue.
If possible, please consider writing useful notes for better and faster
reviews in your PR.

- Please clarify why the changes are needed. For instance, the use case
and bug description.

- Fixes #
-->
This PR supports the access of vllm-acend to the piecewise_graph feature
provided by the v1 engine.

1. register unifiled_ascend_attention_with_output for piecewise_graph to
split graph.
2. support NPUGraph to accelerate kernel launch.

### Does this PR introduce _any_ user-facing change?
<!--
Note that it means *any* user-facing change including all aspects such
as API, interface or other behavior changes.
Documentation-only updates are not considered user-facing changes.
-->
support npugraph to default， Users can disenable the npugraph feature by
configuring enforce_eager.

This has corresponding requirements for the versions of torch_npu and
CANN, and they need to support graph capture.

### How was this patch tested?
<!--
CI passed with new added/existing test.
If it was tested in a way different from regular unit tests, please
clarify how you tested step by step, ideally copy and paste-able, so
that other reviewers can test and check, and descendants can verify in
the future.
If tests were not added, please describe why they were not added and/or
why it was difficult to add.
-->
it turn to default

---------

Signed-off-by: Bug Hunter Yan <yanpq@zju.edu.cn>
Signed-off-by: Yizhou Liu <liu_yizhou@outlook.com>
Co-authored-by: Yizhou Liu <liu_yizhou@outlook.com>

vllm_ascend/worker/model_runner_v1.py

@@ -19,7 +19,10 @@
 
 import gc
 import os
+import time
 import weakref
+from contextlib import contextmanager, nullcontext
+from dataclasses import dataclass
 from typing import TYPE_CHECKING, Dict, List, Optional, Union
 
 import numpy as np
@@ -28,7 +31,7 @@ import torch
 import torch.nn as nn
 from vllm.attention import AttentionType, get_attn_backend
 from vllm.attention.layer import Attention
-from vllm.config import VllmConfig
+from vllm.config import CompilationLevel, VllmConfig
 from vllm.distributed.parallel_state import get_pp_group
 from vllm.forward_context import set_forward_context
 from vllm.inputs import INPUT_REGISTRY
@@ -58,6 +61,43 @@ else:
     xgr = LazyLoader("xgr", globals(), "xgrammar")
 
 
+@dataclass
+class GraphCaptureContext:
+    stream: torch.npu.Stream
+
+
+@contextmanager
+def graph_capture(device: torch.device):
+    """
+    `graph_capture` is a context manager which should surround the code that
+    is capturing the NPU graph. Its main purpose is to ensure that the
+    some operations will be run after the graph is captured, before the graph
+    is replayed. It returns a `GraphCaptureContext` object which contains the
+    necessary data for the graph capture. Currently, it only contains the
+    stream that the graph capture is running on. This stream is set to the
+    current NPU stream when the context manager is entered and reset to the
+    default stream when the context manager is exited. This is to ensure that
+    the graph capture is running on a separate stream from the default stream,
+    in order to explicitly distinguish the kernels to capture
+    from other kernels possibly launched on background in the default stream.
+    """
+    graph_capture_context = GraphCaptureContext(
+        torch.npu.Stream(device=device))
+    stream = graph_capture_context.stream
+
+    # we use nullcontext now
+    maybe_ca_context = nullcontext()
+
+    # ensure all initialization operations complete before attempting to
+    # capture the graph on another stream
+    curr_stream = torch.npu.current_stream()
+    if curr_stream != stream:
+        stream.wait_stream(curr_stream)
+
+    with torch.npu.stream(stream), maybe_ca_context:
+        yield graph_capture_context
+
+
 class NPUModelRunner:
 
     def __init__(self, vllm_config: VllmConfig, device: torch.device):
@@ -229,6 +269,12 @@ class NPUModelRunner:
                                                 device="cpu")
         self.attn_mask = None
         self.attn_state = None
+        self.use_npu_graph = (self.vllm_config.compilation_config.level
+                              == CompilationLevel.PIECEWISE
+                              and not self.model_config.enforce_eager)
+        self.npugraph_batch_sizes = list(
+            reversed(
+                self.vllm_config.compilation_config.cudagraph_capture_sizes))
 
         # NOTE: Pre-construct a mask matrix to improve the efficiency of
         # attention mask construction during inference.
@@ -724,19 +770,19 @@ class NPUModelRunner:
         self.encoder_cache["tmp"] = dict(enumerate(dummy_encoder_outputs))
 
     @torch.inference_mode()
-    def _dummy_run(self) -> torch.Tensor:
+    def _dummy_run(self, num_tokens: int) -> torch.Tensor:
         model = self.model
         if self.is_multimodal_model:
             input_ids = None
-            inputs_embeds = self.inputs_embeds[:self.max_num_tokens]
+            inputs_embeds = self.inputs_embeds[:num_tokens]
         else:
-            input_ids = self.input_ids[:self.max_num_tokens]
+            input_ids = self.input_ids[:num_tokens]
             inputs_embeds = None
 
         if self.uses_mrope:
-            positions = self.mrope_positions[:, :self.max_num_tokens]
+            positions = self.mrope_positions[:, :num_tokens]
         else:
-            positions = self.input_positions_cpu[:self.max_num_tokens]
+            positions = self.positions[:num_tokens]
 
         if get_pp_group().is_first_rank:
             intermediate_tensors = None
@@ -744,17 +790,17 @@ class NPUModelRunner:
             if self.intermediate_tensors is None:
                 self.intermediate_tensors = (
                     self.model.make_empty_intermediate_tensors(
-                        batch_size=self.max_num_tokens,
+                        batch_size=num_tokens,
                         dtype=self.dtype,
                         device=self.device))
             intermediate_tensors = IntermediateTensors({
-                k: v[:self.max_num_tokens]
+                k: v[:num_tokens]
                 for k, v in self.intermediate_tensors.items()
             })
 
         with set_forward_context(None, self.vllm_config):
             hidden_states = model(input_ids=input_ids,
-                                  positions=positions.to(self.device),
+                                  positions=positions,
                                   intermediate_tensors=intermediate_tensors,
                                   inputs_embeds=inputs_embeds)
         return hidden_states
@@ -787,7 +833,7 @@ class NPUModelRunner:
         ]
 
         # Trigger compilation for general shape.
-        hidden_states = self._dummy_run()
+        hidden_states = self._dummy_run(self.max_num_tokens)
 
         if get_pp_group().is_last_rank:
             hidden_states = hidden_states[logit_indices]
@@ -892,3 +938,31 @@ class NPUModelRunner:
                     f"Unknown attention type: {attn_module.attn_type}")
 
         return kv_cache_spec
+
+    def capture_model(self) -> None:
+        if not self.use_npu_graph:
+            logger.warning(
+                "Skipping NPU graph capture. Please add "
+                "-O %s to use NPU graphs.", CompilationLevel.PIECEWISE)
+            return
+
+        start_time = time.perf_counter()
+        start_free_npu_memory = torch.npu.mem_get_info()[0]
+
+        # Trigger NPU graph capture for specific shapes.
+        # Capture the large shapes first so that the smaller shapes
+        # can reuse the memory pool allocated for the large shapes.
+        with graph_capture(device=self.device):
+            for num_tokens in reversed(self.npugraph_batch_sizes):
+                for _ in range(self.vllm_config.compilation_config.
+                               cudagraph_num_of_warmups):
+                    self._dummy_run(num_tokens)
+                self._dummy_run(num_tokens)
+
+        end_time = time.perf_counter()
+        end_free_npu_memory = torch.npu.mem_get_info()[0]
+        elapsed_time = end_time - start_time
+        npu_graph_size = start_free_npu_memory - end_free_npu_memory
+        # This usually takes 5~20 seconds.
+        logger.info("Graph capturing finished in %.0f secs, took %.2f GiB",
+                    elapsed_time, npu_graph_size / (1 << 30))