[Misc] M2m 0724 continuation (vllm-project#12648)

### What this PR does / why we need it?

Adapt vllm-ascend to vLLM main commits up to July 24.

### Changes

| Files | Upstream vLLM Change | vllm-ascend Adaptation |
|---|---|---|
|
`vllm_ascend/worker/v2/attn_utils.py`<br>`vllm_ascend/worker/v2/spec_decode/dflash/speculator.py`<br>`vllm_ascend/worker/v2/spec_decode/dspark/speculator.py`
| [vllm#44492](vllm-project/vllm#44492)
([4ec199b6](vllm-project/vllm@4ec199b6)) —
Populate draft `seq_lens_cpu_upper_bound` for spec-decode attention
metadata. Upstream threads `seq_lens_cpu_upper_bound` through
`_multi_step_decode()` (new 5th positional arg) and
`_build_draft_attn_metadata()` (new `seq_lens_cpu_upper_bound` + `step`
kwargs). | `attn_utils.py`: Added `build_attn_metadata` with
`seq_lens_cpu_upper_bound` param (defaults to `seq_lens_cpu`) and
`build_attn_metadata_wrapper()` context manager for Ascend NPU attention
metadata building.<br>`dflash/speculator.py`: `propose()` caches
`self.input_batch` before `super().propose()`;
`build_draft_attn_metadatas()` passes actual `num_reqs`, padded counts,
`seq_lens_cpu_upper_bound`, and `step` to
`_build_draft_attn_metadata`.<br>`dspark/speculator.py`:
`build_draft_attn_metadatas()` passes `seq_lens_cpu_upper_bound` +
`step`; `propose()` caches `self.input_batch` (version-gated for main
only, since v0.25.1 never reaches this code path via `run_fullgraph`). |
| `vllm_ascend/worker/v2/spec_decode/dflash/aclgraph.py` |
[vllm#47914](vllm-project/vllm#47914)
([0d12618e](vllm-project/vllm@0d12618e)) —
`causal` moved from `DFlashCudaGraphManager.__init__` to `capture()`. |
`capture()` is version-gated: passes `causal` kwarg only on main branch;
v0.25.1 receives it via `__init__`. `run_fullgraph()` accesses
`self.speculator.input_batch.seq_lens_cpu_upper_bound` for building
draft attention metadata during full-graph replay, then updates Ascend
NPU full-graph params. |
| `vllm_ascend/core/recompute_scheduler.py` |
[vllm#47312](vllm-project/vllm#47312)
([12213c67](vllm-project/vllm@12213c67)) —
Handle grammar compilation failures. `finish_requests()` return type
changed from `list[tuple(str, int)]` to `list[Request]`. | Assertion in
`_finish_recomputed_request` is version-gated: on main branch extracts
`(request_id, client_index)` from `Request` objects; v0.25.1 retains the
direct tuple comparison. |
| `vllm_ascend/_310p/model_runner_310p.py` | General vLLM V1 API
evolution (model loading, KV cache allocation,
`determine_available_memory` flow). | New file. 310P hardware platform
model runner extending `NPUModelRunner`. Adapts 310P-specific ACL graph
capture/replay, splitfuse attention, NPU input batch, and FRACTAL_NZ
weight layout to the upstream V1 API. |
| `.github/vllm-main-verified.commit` | N/A (CI tracking) | Updated to
`d02df748bf9efd99022f1a062597dc3cb3808485` to record the verified
upstream vLLM commit. |

### Does this PR introduce _any_ user-facing change?

### How was this patch tested?


- vLLM version: v0.25.1
- vLLM main:
vllm-project/vllm@fe784ff

---------

Signed-off-by: Chen zhichao <chenzhichao33@h-partners.com>
Signed-off-by: hfadzxy <starmoon_zhang@163.com>
Co-authored-by: Chen zhichao <chenzhichao33@h-partners.com>
Co-authored-by: hfadzxy <starmoon_zhang@163.com>

Author: 3 people

.github/vllm-main-verified.commit

@@ -1 +1 @@
-fe784ff22e630a31fd798f392b01e0a75c18f047
+d02df748bf9efd99022f1a062597dc3cb3808485

vllm_ascend/_310p/model_runner_310p.py

@@ -54,7 +54,11 @@
 from vllm_ascend.spec_decode.utils import (
     update_num_computed_tokens_for_batch_change,
 )
-from vllm_ascend.utils import ACL_FORMAT_FRACTAL_NZ, is_rc_device, lmhead_tp_enable
+from vllm_ascend.utils import (
+    ACL_FORMAT_FRACTAL_NZ,
+    is_rc_device,
+    lmhead_tp_enable,
+)
 from vllm_ascend.worker.model_runner_v1 import NPUModelRunner
 
 _NGRAM_GRAPH_UNIFORM_DECODE_QUERY_LEN = 1

vllm_ascend/core/recompute_scheduler.py

@@ -143,7 +143,12 @@ def _finish_recomputed_request(
             request.request_id,
             RequestStatus.FINISHED_ABORTED,
         )
-        assert finished_reqs == [(request.request_id, request.client_index)]
+        if vllm_version_is("0.25.1"):
+            assert finished_reqs == [(request.request_id, request.client_index)]
+        else:
+            assert [(r.request_id, r.client_index) for r in finished_reqs] == [
+                (request.request_id, request.client_index)
+            ]
 
     def schedule(self, throttle_prefills: bool = False) -> RecomputeSchedulerOutput:
         self.current_step += 1

vllm_ascend/worker/v2/attn_utils.py

@@ -45,7 +45,7 @@
 from vllm_ascend.attention.utils import AscendCommonAttentionMetadata
 from vllm_ascend.core.kv_cache_interface import AscendMLAAttentionSpec
 from vllm_ascend.quantization.utils import enable_fa_quant
-from vllm_ascend.utils import calc_split_factor
+from vllm_ascend.utils import calc_split_factor, vllm_version_is
 
 _ATTENTION_MASK_BUILDER = None
 
@@ -109,6 +109,7 @@ def build_attn_metadata(
     dcp_local_seq_lens: torch.Tensor | None = None,
     # extra attributes for ascend npus.
     seq_lens_np: np.ndarray | None = None,
+    seq_lens_cpu_upper_bound: torch.Tensor | None = None,
     num_computed_tokens_cpu: torch.Tensor | None = None,
     positions: torch.Tensor | None = None,
     attn_state: Any | None = None,
@@ -127,6 +128,8 @@ def build_attn_metadata(
     if seq_lens_np is None:
         seq_lens_np = np.full(num_reqs, max_seq_len, dtype=np.int32)
     seq_lens_cpu = torch.from_numpy(seq_lens_np)[:num_reqs]
+    if not vllm_version_is("0.25.1") and seq_lens_cpu_upper_bound is None:
+        seq_lens_cpu_upper_bound = seq_lens_cpu
 
     attn_metadata: dict[str, Any] = {}
     kv_cache_groups = kv_cache_config.kv_cache_groups
@@ -145,7 +148,7 @@ def build_attn_metadata(
             query_start_loc=query_start_loc_gpu,
             query_start_loc_cpu=query_start_loc_cpu,
             seq_lens_cpu=seq_lens_cpu,
-            seq_lens_cpu_upper_bound=seq_lens_cpu,
+            seq_lens_cpu_upper_bound=seq_lens_cpu_upper_bound,
             seq_lens=seq_lens[:num_reqs],
             num_reqs=num_reqs,
             num_actual_tokens=num_tokens,

vllm_ascend/worker/v2/spec_decode/autoregressive/speculator.py

@@ -35,6 +35,7 @@
 from vllm.v1.worker.utils import AttentionGroup
 
 from vllm_ascend.attention.attention_v1 import AscendAttentionState
+from vllm_ascend.utils import vllm_version_is
 from vllm_ascend.worker.v2.attn_utils import build_attn_metadata_wrapper
 from vllm_ascend.worker.v2.input_batch import AscendInputBuffers
 
@@ -275,6 +276,7 @@ def _multi_step_decode(
         skip_attn: bool,
         batch_desc: BatchExecutionDescriptor,
         num_tokens_across_dp: torch.Tensor | None,
+        seq_lens_cpu_upper_bound: torch.Tensor | None = None,
     ) -> None:
         """Minimal override to handle the merged multi-step graph in FULL mode.
 
@@ -288,23 +290,29 @@ def _multi_step_decode(
             assert self.decode_cudagraph_manager is not None
             self.decode_cudagraph_manager.run_fullgraph(batch_desc)
             return
-        super()._multi_step_decode(num_reqs, skip_attn, batch_desc, num_tokens_across_dp)
+        if vllm_version_is("0.25.1"):
+            super()._multi_step_decode(num_reqs, skip_attn, batch_desc, num_tokens_across_dp)
+        else:
+            super()._multi_step_decode(num_reqs, skip_attn, batch_desc, num_tokens_across_dp, seq_lens_cpu_upper_bound)
 
     def _build_draft_attn_metadata(
         self,
         num_reqs: int,
         num_reqs_padded: int,
         num_tokens_padded: int,
+        seq_lens_cpu_upper_bound: torch.Tensor | None = None,
+        step: int = 1,
         num_query_per_req: int = 1,
         causal: bool = True,
     ) -> dict[str, Any] | None:
-        attn_metadata = super()._build_draft_attn_metadata(
-            num_reqs,
-            num_reqs_padded,
-            num_tokens_padded,
-            num_query_per_req,
-            causal,
-        )
+        if vllm_version_is("0.25.1"):
+            attn_metadata = super()._build_draft_attn_metadata(
+                num_reqs, num_reqs_padded, num_tokens_padded, num_query_per_req, causal
+            )
+        else:
+            attn_metadata = super()._build_draft_attn_metadata(
+                num_reqs, num_reqs_padded, num_tokens_padded, seq_lens_cpu_upper_bound, step, num_query_per_req, causal
+            )
         if attn_metadata is not None:
             # Ascend-specific: force DecodeOnly attention state for the draft model.
             for metadata in attn_metadata.values():

vllm_ascend/worker/v2/spec_decode/dflash/aclgraph.py

@@ -105,7 +105,10 @@ def run_fullgraph(self, desc: BatchExecutionDescriptor) -> torch.Tensor | tuple[
         """Override run_fullgraph to update full graph params in run_fullgraph."""
         num_tokens = desc.num_tokens
 
-        draft_attn_metadatas = self.speculator.build_draft_attn_metadatas(desc.num_reqs)
+        draft_attn_metadatas = self.speculator.build_draft_attn_metadatas(
+            desc.num_reqs,
+            self.speculator.input_batch.seq_lens_cpu_upper_bound,
+        )
 
         ret = super().run_fullgraph(desc)
 

vllm_ascend/worker/v2/spec_decode/dflash/speculator.py

@@ -16,6 +16,7 @@
     DFlashSpeculator,
 )
 
+from vllm_ascend.utils import vllm_version_is
 from vllm_ascend.worker.v2.attn_utils import build_attn_metadata_wrapper
 
 logger = logging.getLogger(__name__)
@@ -77,13 +78,15 @@ def set_attn(
 
     # NOTE: upstream vLLM named this to _build_draft_attn_metadatas;
     # keep the current name for now as upstream may change it again.
-    def build_draft_attn_metadatas(self, num_reqs_padded):
+    def build_draft_attn_metadatas(self, num_reqs_padded, seq_lens_cpu_upper_bound):
         num_tokens_padded = num_reqs_padded * self.num_query_per_req
         with build_attn_metadata_wrapper():
             attn_metadata = self._build_draft_attn_metadata(
                 num_reqs=num_reqs_padded,
                 num_reqs_padded=num_reqs_padded,
                 num_tokens_padded=num_tokens_padded,
+                seq_lens_cpu_upper_bound=seq_lens_cpu_upper_bound,
+                step=self.num_query_per_req,
                 causal=self._group_causal,
             )
         return [attn_metadata]
@@ -107,6 +110,11 @@ def propose(
         mm_inputs: tuple[list[torch.Tensor], torch.Tensor] | None = None,
         is_profile: bool = False,
     ) -> torch.Tensor:
+        # TODO: Remove the if not vllm_version_is guard when dropping v0.25.1.
+        # input_batch is cached here so that DFlashAclGraphManager can access
+        # seq_lens_cpu_upper_bound during full-graph replay.
+        if not vllm_version_is("0.25.1"):
+            self.input_batch = input_batch
         with build_attn_metadata_wrapper():
             return super().propose(
                 input_batch,

vllm_ascend/worker/v2/spec_decode/dspark/speculator.py

@@ -28,6 +28,7 @@
     DSparkSpeculator,
 )
 
+from vllm_ascend.utils import vllm_version_is
 from vllm_ascend.worker.v2.attn_utils import build_attn_metadata_wrapper
 
 
@@ -36,6 +37,11 @@ class AscendDSparkSpeculator(DSparkSpeculator):
 
     def __init__(self, vllm_config: VllmConfig, device: torch.device):
         super().__init__(vllm_config, device)
+        # TODO: Remove the if not vllm_version_is guard when dropping v0.25.1.
+        # input_batch is cached in propose() so that DFlashAclGraphManager
+        # can access seq_lens_cpu_upper_bound during full-graph replay.
+        if not vllm_version_is("0.25.1"):
+            self.input_batch: InputBatch | None = None
 
         # we need to update full graph params in run_fullgraph,
         # so create a stream to update full graph params.
@@ -82,13 +88,15 @@ def set_attn(
 
         self.attn_backends = attn_backends
 
-    def build_draft_attn_metadatas(self, num_reqs_padded):
+    def build_draft_attn_metadatas(self, num_reqs_padded, seq_lens_cpu_upper_bound):
         num_tokens_padded = num_reqs_padded * self.num_query_per_req
         with build_attn_metadata_wrapper():
             attn_metadata = self._build_draft_attn_metadata(
                 num_reqs=num_reqs_padded,
                 num_reqs_padded=num_reqs_padded,
                 num_tokens_padded=num_tokens_padded,
+                seq_lens_cpu_upper_bound=seq_lens_cpu_upper_bound,
+                step=self.num_query_per_req,
                 causal=self._group_causal,
             )
         return [attn_metadata]
@@ -112,6 +120,11 @@ def propose(
         mm_inputs: tuple[list[torch.Tensor], torch.Tensor] | None = None,
         is_profile: bool = False,
     ) -> torch.Tensor:
+        # TODO: Remove the if not vllm_version_is guard when dropping v0.25.1.
+        # input_batch is cached here so that DFlashAclGraphManager can access
+        # seq_lens_cpu_upper_bound during full-graph replay.
+        if not vllm_version_is("0.25.1"):
+            self.input_batch = input_batch
         with build_attn_metadata_wrapper():
             return super().propose(
                 input_batch,

vllm_ascend/worker/v2/spec_decode/eagle/aclgraph.py

@@ -26,6 +26,7 @@
     set_draft_graph_prefill_params,
     update_full_graph_params,
 )
+from vllm_ascend.utils import vllm_version_is
 from vllm_ascend.worker.v2.aclgraph_utils import collect_sorted_captured_token_sizes, model_capture_wrapper
 from vllm_ascend.worker.v2.utils import communicator_switch
 
@@ -111,11 +112,16 @@ def create_forward_fn(desc: BatchExecutionDescriptor, warmup: bool):
                     kv_cache_config,
                     skip_attn=(desc.cg_mode == CUDAGraphMode.PIECEWISE),
                 )
+                if vllm_version_is("0.25.1"):
+                    seq_lens_cpu_upper_bound = None
+                else:
+                    seq_lens_cpu_upper_bound = input_buffers.seq_lens_cpu[:num_reqs]
                 return lambda cg_mode: forward_fn(
                     num_reqs,
                     cg_mode == CUDAGraphMode.PIECEWISE,
                     BatchExecutionDescriptor(cg_mode=cg_mode, num_tokens=num_tokens, num_reqs=num_reqs),
                     num_tokens_across_dp,
+                    seq_lens_cpu_upper_bound,
                 )
 
             CudaGraphManager.capture(self, create_forward_fn, progress_bar_desc=progress_bar_desc)

vllm_ascend/worker/v2/spec_decode/mtp/speculator.py

@@ -23,9 +23,11 @@
 from contextlib import contextmanager
 from typing import Any
 
+import torch
 import vllm.v1.worker.gpu.spec_decode.speculator as _upstream_speculator
 from vllm.v1.worker.gpu.spec_decode.mtp.speculator import MTPSpeculator
 
+from vllm_ascend.utils import vllm_version_is
 from vllm_ascend.worker.v2.spec_decode.autoregressive.speculator import AscendAutoRegressiveSpeculator
 
 
@@ -76,6 +78,8 @@ def _build_draft_attn_metadata(
         num_reqs: int,
         num_reqs_padded: int,
         num_tokens_padded: int,
+        seq_lens_cpu_upper_bound: torch.Tensor | None = None,
+        step: int = 1,
         num_query_per_req: int = 1,
         causal: bool = True,
     ) -> dict[str, Any] | None:
@@ -84,6 +88,17 @@ def _build_draft_attn_metadata(
         # super() path does not forward them. Wrap build_attn_metadata to inject
         # positions[:num_tokens_padded] and reuse super() (no arg duplication).
         with build_position_wrapper(self.input_buffers.positions, num_tokens_padded):
-            return super()._build_draft_attn_metadata(
-                num_reqs, num_reqs_padded, num_tokens_padded, num_query_per_req, causal
-            )
+            if vllm_version_is("0.25.1"):
+                return super()._build_draft_attn_metadata(
+                    num_reqs, num_reqs_padded, num_tokens_padded, num_query_per_req, causal
+                )
+            else:
+                return super()._build_draft_attn_metadata(
+                    num_reqs,
+                    num_reqs_padded,
+                    num_tokens_padded,
+                    seq_lens_cpu_upper_bound,
+                    step,
+                    num_query_per_req,
+                    causal,
+                )