Support 310 series

[niejingwei]

.[2025-2-11 3:31:59] [WARNING] [ascend310p] te_rmsnorm_646750ff8f1f95942648e3ac22e8da92b00f3bba6f0542271074f58a2f021cb3 not found cann_version.h
......ERROR 02-11 11:32:58 engine.py:389] call aclnnPromptFlashAttentionV3 failed, detail:EZ1001: [PID: 312378] 2025-02-11-11:32:58.201.638 PromptFlashAttention LaunchAicore failed.
ERROR 02-11 11:32:58 engine.py:389]         TraceBack (most recent call last):
ERROR 02-11 11:32:58 engine.py:389]         Cannot find bin of op PromptFlashAttention, integral key 0/1/|bf16/ND/bf16/ND/bf16/ND/bf16/ND/.
ERROR 02-11 11:32:58 engine.py:389]         Cannot find binary for op PromptFlashAttention.
ERROR 02-11 11:32:58 engine.py:389]         Kernel GetWorkspace failed. opType: 5
ERROR 02-11 11:32:58 engine.py:389]         PromptFlashAttention LaunchAicore failed.
ERROR 02-11 11:32:58 engine.py:389]
ERROR 02-11 11:32:58 engine.py:389] [ERROR] 2025-02-11-11:32:58 (PID:312378, Device:0, RankID:-1) ERR01100 OPS call acl api failed
ERROR 02-11 11:32:58 engine.py:389] Traceback (most recent call last):
ERROR 02-11 11:32:58 engine.py:389]   File "/root/miniconda3/envs/vllm-ascend-cp310/lib/python3.10/site-packages/vllm/engine/multiprocessing/engine.py", line 380, in run_mp_engine
ERROR 02-11 11:32:58 engine.py:389]     engine = MQLLMEngine.from_engine_args(engine_args=engine_args,
ERROR 02-11 11:32:58 engine.py:389]   File "/root/miniconda3/envs/vllm-ascend-cp310/lib/python3.10/site-packages/vllm/engine/multiprocessing/engine.py", line 123, in from_engine_args
ERROR 02-11 11:32:58 engine.py:389]     return cls(ipc_path=ipc_path,
ERROR 02-11 11:32:58 engine.py:389]   File "/root/miniconda3/envs/vllm-ascend-cp310/lib/python3.10/site-packages/vllm/engine/multiprocessing/engine.py", line 75, in __init__
ERROR 02-11 11:32:58 engine.py:389]     self.engine = LLMEngine(*args, **kwargs)
ERROR 02-11 11:32:58 engine.py:389]   File "/root/miniconda3/envs/vllm-ascend-cp310/lib/python3.10/site-packages/vllm/engine/llm_engine.py", line 276, in __init__
ERROR 02-11 11:32:58 engine.py:389]     self._initialize_kv_caches()
ERROR 02-11 11:32:58 engine.py:389]   File "/root/miniconda3/envs/vllm-ascend-cp310/lib/python3.10/site-packages/vllm/engine/llm_engine.py", line 416, in _initialize_kv_caches
ERROR 02-11 11:32:58 engine.py:389]     self.model_executor.determine_num_available_blocks())
ERROR 02-11 11:32:58 engine.py:389]   File "/root/miniconda3/envs/vllm-ascend-cp310/lib/python3.10/site-packages/vllm/executor/executor_base.py", line 101, in determine_num_available_blocks
ERROR 02-11 11:32:58 engine.py:389]     results = self.collective_rpc("determine_num_available_blocks")

报以上错误
机器环境如下

[wangxiyuan]

vllm-ascend doesn't work on 310 currently. It'll be done in Q2. Thanks.

[Yikun]

For 310P supported:

According the support experience for other engine [1], it's not a difficult thing to have a initial suppport, but in order to meet the performance requirement, need more efforts and bandwidth.

We don't have a certain plans for it yet, but we welcome contributions, feel free to open PR and issue/RFC.

[1] ggml-org/llama.cpp#10216

[Yikun]

Related: #914