[refact] unified soc_version code (#4359)

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

Currently, there are two paths to judge the chip type in code,
`get_ascend_soc_version` use `get_soc_version` api in torch_npu, and
`is_310p` `use _build_info.__soc_version__`, which generate when
install. We need to unify the two paths.

We need to unify these codes based on the following points:

1. We need to ensure consistency in chip type judgment between compiling
and running states;
2. In compiling state, we need chip type to complete op's compilation,
but in running state, we only need device
type(910B/910_93/310P/910_95/etc) to make code branch judgement;
3. In compiling state, torch_npu may not have been installed yet, so we
can't use torch_npu's api.

Based on the above points, we have made the following changes:

1. When user set env `SOC_VERSION`, use it; when not set, query
soc_version by `npu-smi`;
2. generate device_type based on soc_version when compiling, and write
`__device_type__` instead of `__soc_version__` in `_build_info.py`;
3. In running state, use `__device_type__` to judge code branch.

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

When not set env `SOC_VERSION`, it will not be `ASCEND910B1` by default,
we will query soc_version by `npu-smi`. And env `SOC_VERSION` must be in
the list `soc_to_device` in `setup.py`.

- vLLM version: v0.11.0
- vLLM main:
2918c1b49c

Signed-off-by: zzzzwwjj <1183291235@qq.com>
This commit is contained in:
zzzzwwjj
2025-11-26 14:28:55 +08:00
committed by GitHub
parent a91e76cd84
commit 136ea9ff56
42 changed files with 361 additions and 243 deletions

View File

@@ -16,7 +16,7 @@ from unittest.mock import MagicMock, patch
import pytest
from vllm_ascend.ascend_forward_context import MoECommType
from vllm_ascend.utils import AscendSocVersion
from vllm_ascend.utils import AscendDeviceType
from vllm_ascend.worker.model_runner_v1 import NPUModelRunner
@@ -25,21 +25,21 @@ from vllm_ascend.worker.model_runner_v1 import NPUModelRunner
"soc_version, enable_expert_parallel, world_size, num_tokens, mc2_tokens_capacity, quant_type, expected_method",
[
# Case 1: Expert parallel is disabled, should always be 'allgather'
(AscendSocVersion.A2, False, 8, 100, 256, None, MoECommType.ALLGATHER),
(AscendSocVersion.A3, False, 16, 500, 256, None, MoECommType.ALLGATHER),
(AscendDeviceType._910B, False, 8, 100, 256, None, MoECommType.ALLGATHER),
(AscendDeviceType._910_93, False, 16, 500, 256, None, MoECommType.ALLGATHER),
# Case 2: A2 SOC with w4a8_dynamic -> use alltoall when not mc2
(AscendSocVersion.A2, True, 8, 100, 256, "w4a8_dynamic", MoECommType.ALLTOALL),
(AscendSocVersion.A2, True, 16, 257, 256, "w4a8_dynamic", MoECommType.ALLTOALL),
(AscendSocVersion.A2, True, 16, 100, 256, "w4a8_dynamic", MoECommType.MC2), # meets mc2 condition
(AscendDeviceType._910B, True, 8, 100, 256, "w4a8_dynamic", MoECommType.ALLTOALL),
(AscendDeviceType._910B, True, 16, 257, 256, "w4a8_dynamic", MoECommType.ALLTOALL),
(AscendDeviceType._910B, True, 16, 100, 256, "w4a8_dynamic", MoECommType.MC2), # meets mc2 condition
# Case 3: A2 SOC without w4a8_dynamic -> fallback to allgather
(AscendSocVersion.A2, True, 8, 100, 256, None, MoECommType.ALLGATHER),
(AscendSocVersion.A2, True, 16, 257, 256, None, MoECommType.ALLGATHER),
(AscendDeviceType._910B, True, 8, 100, 256, None, MoECommType.ALLGATHER),
(AscendDeviceType._910B, True, 16, 257, 256, None, MoECommType.ALLGATHER),
# Case 4: A3 SOC
(AscendSocVersion.A3, True, 8, 100, 256, None, MoECommType.MC2),
(AscendSocVersion.A3, True, 8, 257, 256, None, MoECommType.ALLTOALL),
(AscendDeviceType._910_93, True, 8, 100, 256, None, MoECommType.MC2),
(AscendDeviceType._910_93, True, 8, 257, 256, None, MoECommType.ALLTOALL),
])
# yapf: enable
def test_select_moe_comm_method(soc_version, enable_expert_parallel,
@@ -65,7 +65,7 @@ def test_select_moe_comm_method(soc_version, enable_expert_parallel,
mock_runner.vllm_config = mock_vllm_config
# Patch the helper functions
with patch('vllm_ascend.worker.model_runner_v1.get_ascend_soc_version',
with patch('vllm_ascend.worker.model_runner_v1.get_ascend_device_type',
return_value=soc_version), \
patch('vllm_ascend.worker.model_runner_v1.is_global_first_rank',
return_value=True), \
@@ -100,7 +100,7 @@ def test_select_moe_comm_method_unsupported_soc():
unsupported_soc = "UnsupportedSOC"
with patch('vllm_ascend.worker.model_runner_v1.get_ascend_soc_version',
with patch('vllm_ascend.worker.model_runner_v1.get_ascend_device_type',
return_value=unsupported_soc), \
patch('vllm_ascend.worker.model_runner_v1.is_global_first_rank',
return_value=True), \