[Enhancement] Add padding for ACL Graph (#803)

### What this PR does / why we need it?
Add padding for ACL Graph and refactor graph batch size adjustments to
utils.py

---------

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

vllm_ascend/attention/attention_v1.py

@@ -104,6 +104,7 @@ class AscendAttentionState(Enum):
 
 @dataclass
 class AscendMetadata:
+    num_actual_tokens: int  # Number of tokens excluding padding.
     # (batch_size, max_blocks_per_seq).
     # Block addresses per sequence. (Seq id -> list of physical block)
     block_tables: torch.Tensor
@@ -125,7 +126,6 @@ class AscendMetadata:
     is_only_prefill: bool = False
     # Current state of this attention run.
     attn_state: AscendAttentionState = AscendAttentionState.ChunkedPrefill
-
     attn_mask: Optional[torch.Tensor] = None
 
 
@@ -149,7 +149,8 @@ class AscendAttentionMetadataBuilder:
         attn_mask = self.runner.attn_mask
         attn_state = self.runner.attn_state
 
-        attn_metadata = AscendMetadata(block_tables=block_table,
+        attn_metadata = AscendMetadata(num_actual_tokens=num_actual_tokens,
+                                       block_tables=block_table,
                                        query_lens=query_lens,
                                        seq_lens=seq_lens,
                                        max_query_len=max_query_len,
@@ -234,9 +235,9 @@ class AscendAttentionBackendImpl(AttentionImpl):
                 output=output,
                 layer_name=layer.layer_name)
         else:
-            num_tokens = query.shape[0]
             if attn_metadata is None:
                 return output.view(num_tokens, self.hidden_size)
+            num_actual_tokens = attn_metadata.num_actual_tokens
             assert layer._k_scale_float == 1.0 and layer._v_scale_float == 1.0
             attn_type = self.attn_type
             if attn_type != AttentionType.DECODER:
@@ -255,11 +256,12 @@ class AscendAttentionBackendImpl(AttentionImpl):
                 if self.key_cache is None:
                     self.key_cache, self.value_cache = kv_cache[0], kv_cache[1]
                 slots = attn_metadata.slot_mapping
-                torch_npu._npu_reshape_and_cache(key=key,
+                torch_npu._npu_reshape_and_cache(
-                                                 value=value,
+                    key=key[:num_actual_tokens],
-                                                 key_cache=self.key_cache,
+                    value=value[:num_actual_tokens],
-                                                 value_cache=self.value_cache,
+                    key_cache=self.key_cache,
-                                                 slot_indices=slots)
+                    value_cache=self.value_cache,
+                    slot_indices=slots)
 
             if hasattr(layer, 'quant_method'):
                 # TODO: Add attr (num_prefills, prefill_metadata, decode_metadata) to AscendMetadata

vllm_ascend/platform.py

@@ -25,6 +25,8 @@ from vllm.logger import logger
 from vllm.platforms import Platform, PlatformEnum
 from vllm.utils import supports_dynamo
 
+from vllm_ascend.utils import update_aclgraph_sizes
+
 CUSTOM_OP_ENABLED = False
 try:
     # register custom ops into torch_library here
@@ -144,6 +146,7 @@ class NPUPlatform(Platform):
             compilation_config.use_inductor = False
             compilation_config.splitting_ops.extend(
                 ["vllm.unified_ascend_attention_with_output"])
+            update_aclgraph_sizes(vllm_config)
 
         if vllm_config.additional_config is not None:
             enable_graph_mode = vllm_config.additional_config.get(

vllm_ascend/utils.py

@@ -16,12 +16,28 @@
 # This file is a part of the vllm-ascend project.
 # Adapted from vllm-project/vllm/vllm/worker/worker.py
 #
+
+import math
+from typing import TYPE_CHECKING
+
 import torch
 from packaging.version import InvalidVersion, Version
 from vllm.logger import logger
 
 import vllm_ascend.envs as envs
 
+if TYPE_CHECKING:
+    from vllm.config import VllmConfig
+else:
+    VllmConfig = None
+
+# NOTE: Currently, we can only capture 1920 graphs at most,
+# due to the limitation of ACL graph. This number is bounded by
+# the number of streams, which is 2048, we save 128 streams
+# as a buffer.
+# Maximum number of graphs that can be captured by ACL Graph
+MAX_CAPTURE_SIZE = 1920
+
 
 def try_register_lib(lib_name: str, lib_info: str = ""):
     import importlib
@@ -99,3 +115,55 @@ def vllm_version_is(target_vllm_version: str):
             "is installed probably. Set the environment variable VLLM_VERSION "
             "to control it by hand. And please make sure the vaule follows the "
             "format of x.y.z.")
+
+
+def update_aclgraph_sizes(vllm_config: VllmConfig) -> None:
+    """Update ACL graph capture sizes based on hardware limitations"""
+    # Store original configuration and temporarily clear it
+    compilation_config = vllm_config.compilation_config
+    original_sizes, compilation_config.cudagraph_capture_sizes = \
+        compilation_config.cudagraph_capture_sizes, None
+
+    # Calculate parallel configuration factor (increases with DP or TP)
+    # TODO(Yizhou): This is a temporary solution, need to be improved
+    # in the future, taking into account the other parallel configurations.
+    num_hidden_layers = vllm_config.model_config.hf_config.num_hidden_layers
+    parallel_config = vllm_config.parallel_config
+    parallel_factor = 1 + sum(size > 1 for size in [
+        parallel_config.data_parallel_size,
+        parallel_config.tensor_parallel_size
+    ])
+
+    # Calculate maximum supported batch sizes considering model architecture
+    max_num_batch_sizes = math.floor(MAX_CAPTURE_SIZE /
+                                     (num_hidden_layers + 1) / parallel_factor)
+    logger.info("Calculated maximum supported batch sizes for ACL graph: %s",
+                max_num_batch_sizes)
+
+    # If original sizes exceed maximum, sample a representative subset
+    if max_num_batch_sizes < len(original_sizes):
+        # Sample uniformly from original sizes
+        step = (len(original_sizes) - 1) / (max_num_batch_sizes - 1)
+        indices = [round(i * step) for i in range(max_num_batch_sizes)]
+
+        # Ensure first and last elements are preserved
+        indices[0], indices[-1] = 0, len(original_sizes) - 1
+
+        sampled_sizes = [original_sizes[i] for i in indices]
+        compilation_config.init_with_cudagraph_sizes(sampled_sizes)
+
+        logger.info(
+            "Adjusted ACL graph batch sizes for %s model (layers: %d): %d → %d sizes",
+            vllm_config.model_config.architectures[0],
+            num_hidden_layers,
+            len(original_sizes),
+            len(compilation_config.
+                cudagraph_capture_sizes  # type: ignore[arg-type]
+                ))
+    else:
+        # No adjustment needed
+        compilation_config.cudagraph_capture_sizes = original_sizes
+        logger.info(
+            "No adjustment needed for ACL graph batch sizes: %s model (layers: %d) with %d sizes",
+            vllm_config.model_config.architectures[0], num_hidden_layers,
+            len(original_sizes))

vllm_ascend/worker/model_runner_v1.py

@@ -18,7 +18,6 @@
 #
 
 import gc
-import math
 import os
 import time
 import weakref
@@ -293,9 +292,9 @@ class NPUModelRunner:
                                                 device="cpu")
         self.attn_mask = None
         self.attn_state = None
-        self.use_npu_graph = (self.vllm_config.compilation_config.level
+        self.use_aclgraph = (self.vllm_config.compilation_config.level
-                              == CompilationLevel.PIECEWISE
+                             == CompilationLevel.PIECEWISE
-                              and not self.model_config.enforce_eager)
+                             and not self.model_config.enforce_eager)
         self.aclgraph_batch_sizes = list(
             reversed(
                 self.vllm_config.compilation_config.cudagraph_capture_sizes))
@@ -508,6 +507,13 @@ class NPUModelRunner:
         assert total_num_scheduled_tokens > 0
         num_reqs = self.input_batch.num_reqs
         assert num_reqs > 0
+        if (self.use_aclgraph and
+                total_num_scheduled_tokens <= self.aclgraph_batch_sizes[-1]):
+            # Add padding to the batch size.
+            num_input_tokens = self.vllm_config.pad_for_cudagraph(
+                total_num_scheduled_tokens)
+        else:
+            num_input_tokens = total_num_scheduled_tokens
 
         modified_batch = self.attn_metadata_builder.reorder_batch(
             self.input_batch, scheduler_output)
@@ -546,7 +552,7 @@ class NPUModelRunner:
 
         self.positions[:total_num_scheduled_tokens].copy_(
             self.positions_cpu[:total_num_scheduled_tokens], non_blocking=True)
-        positions = self.positions[:total_num_scheduled_tokens]
+        positions = self.positions[:num_input_tokens]
         self.query_lens = torch.from_numpy(num_scheduled_tokens)
 
         self.seq_lens_np[:num_reqs] = (
@@ -605,7 +611,7 @@ class NPUModelRunner:
         # Copy the tensors to the NPU.
         self.input_ids[:total_num_scheduled_tokens].copy_(
             self.input_ids_cpu[:total_num_scheduled_tokens], non_blocking=True)
-        input_ids = self.input_ids[:total_num_scheduled_tokens]
+        input_ids = self.input_ids[:num_input_tokens]
 
         if self.enable_torchair_graph_mode and attn_metadata.attn_state == AscendAttentionState.DecodeOnly:
             padding = torch.zeros(graph_pad_size,
@@ -615,7 +621,9 @@ class NPUModelRunner:
             positions = torch.cat([positions, padding])
 
         # Run forward pass
-        with set_forward_context(attn_metadata, self.vllm_config):
+        with set_forward_context(attn_metadata,
+                                 self.vllm_config,
+                                 num_tokens=num_input_tokens):
             model_kwargs = {}
             if self.enable_torchair_graph_mode:
                 model_kwargs["kv_caches"] = self.kv_caches
@@ -1062,7 +1070,7 @@ class NPUModelRunner:
         return kv_cache_spec
 
     def capture_model(self) -> None:
-        if not self.use_npu_graph:
+        if not self.use_aclgraph:
             logger.warning(
                 "Skipping NPU graph capture. Please add "
                 "-O %s to use NPU graphs.", CompilationLevel.PIECEWISE)
@@ -1070,9 +1078,6 @@ class NPUModelRunner:
 
         start_time = time.perf_counter()
         start_free_npu_memory = torch.npu.mem_get_info()[0]
-        #  Since vllm aclgraph_batch_sizes is too large,
-        #  we need to adjust its length to proper size.
-        self.verify_adjust_aclgraph_batch_sizes()
 
         # Trigger ACL graph capture for specific shapes.
         # Capture the large shapes first so that the smaller shapes
@@ -1091,63 +1096,3 @@ class NPUModelRunner:
         # This usually takes 5~20 seconds.
         logger.info("Graph capturing finished in %.0f secs, took %.2f GiB",
                     elapsed_time, npu_graph_size / (1 << 30))
-
-    def verify_adjust_aclgraph_batch_sizes(self) -> None:
-        # Now, vllm-ascend support max capture size is 1920
-        max_capture_size = 1920
-        original_aclgraph_batch_sizes = self.aclgraph_batch_sizes
-        num_hidden_layers = self.vllm_config.model_config.hf_config.num_hidden_layers
-        max_support_len_aclgraph = self.get_max_support_len(
-            max_capture_size, num_hidden_layers)
-
-        if max_support_len_aclgraph < len(original_aclgraph_batch_sizes):
-            self.aclgraph_batch_sizes = self.sample_from_list(
-                max_support_len_aclgraph)
-
-            logger.info(
-                "Model:%s-num_hidden_layers:%d will adjust aclgraph_batch_sizes, pre-adjust-len: %s, post-adjust-len: %s",
-                self.vllm_config.model_config.architectures[0],
-                num_hidden_layers, len(original_aclgraph_batch_sizes),
-                len(self.aclgraph_batch_sizes))
-        else:
-            logger.info(
-                "Model:%s-num_hidden_layers:%d no need adjust aclgraph_batch_sizes, list_len: %s",
-                self.vllm_config.model_config.architectures[0],
-                num_hidden_layers, len(original_aclgraph_batch_sizes))
-
-    def get_max_support_len(self, max_capture_size, num_hidden_layers) -> int:
-        parallel_type_cnt = 0
-        dp_size = self.vllm_config.parallel_config.data_parallel_size
-        tp_size = self.vllm_config.parallel_config.tensor_parallel_size
-        if dp_size > 1:
-            parallel_type_cnt += 1
-        if tp_size > 1:
-            parallel_type_cnt += 1
-        max_support_len_aclgraph = math.floor(max_capture_size /
-                                              (num_hidden_layers + 1) /
-                                              (parallel_type_cnt + 1))
-        logger.info(
-            "max_capture_size:%s, dp_size:%s, tp_size:%s, parallel_type_cnt:%s, max_support_len_aclgraph: %s:",
-            max_capture_size,
-            dp_size,
-            tp_size,
-            parallel_type_cnt,
-            max_support_len_aclgraph,
-        )
-
-        return max_support_len_aclgraph
-
-    def sample_from_list(self, sample_len) -> list[int]:
-        # we use this function to sample a new list from old list by given length, and maintain uniformity, for example:
-        # original: [1 8 16 24 32 40 48 56 64]
-        # --> sample length = 3: [1 32 64]
-        # --> sample length = 5: [1 16 32 48 64]
-        original_len = len(self.aclgraph_batch_sizes)
-        step = (original_len - 1) / (sample_len - 1)
-        indices = [round(i * step) for i in range(sample_len)]
-        # Align first and last element of the original list and sub-list
-        indices[0] = 0
-        indices[-1] = original_len - 1
-        # Sample new list
-        new_list = [self.aclgraph_batch_sizes[i] for i in indices]
-        return new_list