[Bugfix] Fix output tensor shape in vanilla_chunked_prefill and update import paths for model_loader (#773)

<!--  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 #
-->
Fix output tensor shape in vanilla_chunked_prefill function.

### 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.
-->
None.

### 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.
-->
Run offline inference on DeepSeek models.

---------

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

Author: yiz-liu

tests/utils.py

@@ -42,13 +42,19 @@
                               init_distributed_environment)
 from vllm.engine.arg_utils import AsyncEngineArgs
 from vllm.entrypoints.openai.cli_args import make_arg_parser
-from vllm.model_executor.model_loader.loader import get_model_loader
 from vllm.platforms import current_platform
 from vllm.transformers_utils.tokenizer import get_tokenizer
 from vllm.utils import FlexibleArgumentParser, GB_bytes, get_open_port
 
+from vllm_ascend.utils import vllm_version_is
+
 from .model_utils import TextTextLogprobs
 
+if vllm_version_is("0.8.5") or vllm_version_is("0.8.5.post1"):
+    from vllm.model_executor.model_loader.loader import get_model_loader  # type: ignore[import]  # isort: skip
+else:
+    from vllm.model_executor.model_loader import get_model_loader
+
 VLLM_PATH = Path(__file__).parent.parent
 """Path to root of the vLLM repository."""
 

vllm_ascend/ops/attention.py

@@ -131,6 +131,7 @@ def vanilla_chunked_prefill(
 
     attn_output = (attn_output[q_mask].view([-1, num_query_heads,
                                              head_dim]).to(output.dtype))
+    output = output.view_as(attn_output)
     output.copy_(attn_output)
     return attn_output
 

vllm_ascend/worker/model_runner.py

@@ -64,6 +64,8 @@
     _init_attn_metadata_from_tensor_dict,
     _init_sampling_metadata_from_tensor_dict)
 
+from vllm_ascend.utils import vllm_version_is
+
 if TYPE_CHECKING:
     from vllm.attention.backends.abstract import AttentionBackend
 
@@ -1007,7 +1009,10 @@ def save_sharded_state(
         pattern: Optional[str] = None,
         max_size: Optional[int] = None,
     ) -> None:
-        from vllm.model_executor.model_loader.loader import ShardedStateLoader
+        if vllm_version_is("0.8.5") or vllm_version_is("0.8.5.post1"):
+            from vllm.model_executor.model_loader.loader import ShardedStateLoader  # type: ignore[import]  # isort: skip  # noqa
+        else:
+            from vllm.model_executor.model_loader import ShardedStateLoader
         ShardedStateLoader.save_model(
             self.model,
             path,
@@ -1019,7 +1024,12 @@ def save_tensorized_model(
         self,
         tensorizer_config: TensorizerConfig,
     ) -> None:
-        from vllm.model_executor.model_loader.loader import TensorizerLoader
+        if vllm_version_is("0.8.5") or vllm_version_is("0.8.5.post1"):
+            from vllm.model_executor.model_loader.loader import \
+                TensorizerLoader  # type: ignore  # noqa
+        else:
+            from vllm.model_executor.model_loader import \
+                TensorizerLoader  # type: ignore  # noqa
         TensorizerLoader.save_model(
             self.model,
             tensorizer_config=tensorizer_config,