EngineX/xc-llm-ascend
3
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity

[Feature][Quant] Reapply auto-detect quantization format and support remote model ID (#7111)

Browse Source 
### What this PR does / why we need it?
Reapply the auto-detect quantization format feature (originally in
#6645, reverted in #6873) and extend it to support remote model
identifiers (e.g., `org/model-name`).

Changes:
- Reapply auto-detection of quantization method from model files
(`quant_model_description.json` for ModelSlim, `config.json` for
compressed-tensors)
- Add `get_model_file()` utility to handle file retrieval from both
local paths and remote repos (HuggingFace Hub / ModelScope)
- Update `detect_quantization_method()` to accept remote repo IDs with
optional `revision` parameter
- Update `maybe_update_config()` to work with remote model identifiers
- Add platform-level `auto_detect_quantization` support
- Add unit tests and e2e tests for both local and remote model ID
scenarios

Closes #6836

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

Yes. When `--quantization` is not explicitly specified, vllm-ascend will
now automatically detect the quantization format from the model files
for both local directories and remote model IDs.

- vLLM version: v0.16.0
- vLLM main:
4034c3d32e
---------
Signed-off-by: SlightwindSec <slightwindsec@gmail.com>
This commit is contained in:
Cao Yi
2026-03-13 22:53:25 +08:00
committed by GitHub
parent 6852a2e267
commit 5ec610e832
7 changed files with 658 additions and 12 deletions
Download Patch File Download Diff File Expand all files Collapse all files

37
tests/e2e/singlecard/test_quantization.py
View File

@@ -49,6 +49,43 @@ def test_qwen3_w8a8_quant():
name_1="vllm_quant_w8a8_outputs",
)
# fmt: off
def test_qwen3_w8a8_quant_auto_detect():
"""Test that ModelSlim quantization is auto-detected without --quantization.
Uses the same W8A8 model as test_qwen3_w8a8_quant but omits the
quantization parameter, verifying that the auto-detection in
maybe_auto_detect_quantization() picks up quant_model_description.json
and produces identical results.
"""
max_tokens = 5
example_prompts = [
"vLLM is a high-throughput and memory-efficient inference and serving engine for LLMs."
]
vllm_target_outputs = [([
85, 4086, 44, 374, 264, 1550, 42747, 628, 323, 4938, 72816, 44378, 323,
13480, 4712, 369, 444, 10994, 82, 13, 1084, 374, 6188, 311, 387
], 'vLLM is a high-throughput and memory-efficient inference and serving engine for LLMs. It is designed to be'
)]
# fmt: on
with VllmRunner(
"vllm-ascend/Qwen3-0.6B-W8A8",
max_model_len=8192,
gpu_memory_utilization=0.7,
cudagraph_capture_sizes=[1, 2, 4, 8],
) as vllm_model:
vllm_quant_auto_detect_outputs = vllm_model.generate_greedy(
example_prompts, max_tokens)
check_outputs_equal(
outputs_0_lst=vllm_target_outputs,
outputs_1_lst=vllm_quant_auto_detect_outputs,
name_0="vllm_target_outputs",
name_1="vllm_quant_auto_detect_outputs",
)
# fmt: off
def test_qwen3_dense_w8a16():
max_tokens = 5

91
tests/ut/quantization/test_modelslim_config.py
View File

@@ -1,3 +1,6 @@
import json
import os
import tempfile
from unittest.mock import MagicMock, patch
from vllm.model_executor.layers.fused_moe import FusedMoE
@@ -7,6 +10,7 @@ from vllm.model_executor.layers.linear import LinearBase
from tests.ut.base import TestBase
from vllm_ascend.ops.linear import AscendUnquantizedLinearMethod
from vllm_ascend.quantization.modelslim_config import (
MODELSLIM_CONFIG_FILENAME,
AscendModelSlimConfig,
)
from vllm_ascend.utils import ASCEND_QUANTIZATION_METHOD
@@ -53,7 +57,7 @@ class TestAscendModelSlimConfig(TestBase):
def test_get_config_filenames(self):
filenames = AscendModelSlimConfig.get_config_filenames()
self.assertEqual(filenames, ["quant_model_description.json"])
self.assertEqual(filenames, [])
def test_from_config(self):
config = AscendModelSlimConfig.from_config(self.sample_config)
@@ -161,5 +165,90 @@ class TestAscendModelSlimConfig(TestBase):
with self.assertRaises(ValueError):
config.is_layer_skipped_ascend("fused_layer", fused_mapping)
def test_init_with_none_config(self):
config = AscendModelSlimConfig(None)
self.assertEqual(config.quant_description, {})
def test_init_with_default_config(self):
config = AscendModelSlimConfig()
self.assertEqual(config.quant_description, {})
def test_maybe_update_config_already_populated(self):
# When quant_description is already populated, should be a no-op
self.assertTrue(len(self.ascend_config.quant_description) > 0)
self.ascend_config.maybe_update_config("/some/model/path")
# quant_description should remain unchanged
self.assertEqual(self.ascend_config.quant_description,
self.sample_config)
def test_maybe_update_config_loads_from_file(self):
config = AscendModelSlimConfig()
self.assertEqual(config.quant_description, {})
quant_data = {"layer1.weight": "INT8", "layer2.weight": "FLOAT"}
with tempfile.TemporaryDirectory() as tmpdir:
config_path = os.path.join(tmpdir, MODELSLIM_CONFIG_FILENAME)
with open(config_path, "w") as f:
json.dump(quant_data, f)
config.maybe_update_config(tmpdir)
self.assertEqual(config.quant_description, quant_data)
def test_maybe_update_config_raises_when_file_missing(self):
config = AscendModelSlimConfig()
with tempfile.TemporaryDirectory() as tmpdir:
with self.assertRaises(ValueError) as ctx:
config.maybe_update_config(tmpdir)
error_msg = str(ctx.exception)
self.assertIn("ModelSlim Quantization Config Not Found", error_msg)
self.assertIn(MODELSLIM_CONFIG_FILENAME, error_msg)
def test_maybe_update_config_raises_with_json_files_listed(self):
config = AscendModelSlimConfig()
with tempfile.TemporaryDirectory() as tmpdir:
# Create a dummy json file that is NOT the config file
dummy_path = os.path.join(tmpdir, "config.json")
with open(dummy_path, "w") as f:
json.dump({"dummy": True}, f)
with self.assertRaises(ValueError) as ctx:
config.maybe_update_config(tmpdir)
error_msg = str(ctx.exception)
self.assertIn("config.json", error_msg)
def test_maybe_update_config_non_directory_raises(self):
config = AscendModelSlimConfig()
with self.assertRaises(ValueError) as ctx:
config.maybe_update_config("not_a_real_directory_path")
error_msg = str(ctx.exception)
self.assertIn("ModelSlim Quantization Config Not Found", error_msg)
def test_apply_extra_quant_adaptations_shared_head(self):
config = AscendModelSlimConfig()
config.quant_description = {
"model.layers.0.shared_head.weight": "INT8",
}
config._apply_extra_quant_adaptations()
self.assertIn("model.layers.0.weight", config.quant_description)
self.assertEqual(config.quant_description["model.layers.0.weight"],
"INT8")
def test_apply_extra_quant_adaptations_weight_packed(self):
config = AscendModelSlimConfig()
config.quant_description = {
"model.layers.0.weight_packed": "INT8",
}
config._apply_extra_quant_adaptations()
self.assertIn("model.layers.0.weight", config.quant_description)
self.assertEqual(config.quant_description["model.layers.0.weight"],
"INT8")
def test_get_scaled_act_names(self):
self.assertEqual(self.ascend_config.get_scaled_act_names(), [])

192
tests/ut/quantization/test_quant_utils.py Normal file
View File

@@ -0,0 +1,192 @@
import json
import logging
import os
import tempfile
from pathlib import Path
from unittest.mock import MagicMock, patch
from tests.ut.base import TestBase
from vllm_ascend.quantization.modelslim_config import MODELSLIM_CONFIG_FILENAME
from vllm_ascend.quantization.utils import (
detect_quantization_method,
maybe_auto_detect_quantization,
)
from vllm_ascend.utils import ASCEND_QUANTIZATION_METHOD, COMPRESSED_TENSORS_METHOD
class TestDetectQuantizationMethod(TestBase):
def test_returns_none_for_non_existent_path(self):
result = detect_quantization_method("/non/existent/path")
self.assertIsNone(result)
def test_detects_modelslim(self):
with tempfile.TemporaryDirectory() as tmpdir:
config_path = os.path.join(tmpdir, MODELSLIM_CONFIG_FILENAME)
with open(config_path, "w") as f:
json.dump({"layer.weight": "INT8"}, f)
result = detect_quantization_method(tmpdir)
self.assertEqual(result, ASCEND_QUANTIZATION_METHOD)
def test_detects_compressed_tensors(self):
with tempfile.TemporaryDirectory() as tmpdir:
config_path = os.path.join(tmpdir, "config.json")
with open(config_path, "w") as f:
json.dump({
"quantization_config": {
"quant_method": "compressed-tensors"
}
}, f)
result = detect_quantization_method(tmpdir)
self.assertEqual(result, COMPRESSED_TENSORS_METHOD)
def test_returns_none_for_no_quant(self):
with tempfile.TemporaryDirectory() as tmpdir:
result = detect_quantization_method(tmpdir)
self.assertIsNone(result)
def test_returns_none_for_non_compressed_tensors_quant_method(self):
with tempfile.TemporaryDirectory() as tmpdir:
config_path = os.path.join(tmpdir, "config.json")
with open(config_path, "w") as f:
json.dump({
"quantization_config": {
"quant_method": "gptq"
}
}, f)
result = detect_quantization_method(tmpdir)
self.assertIsNone(result)
def test_returns_none_for_config_without_quant_config(self):
with tempfile.TemporaryDirectory() as tmpdir:
config_path = os.path.join(tmpdir, "config.json")
with open(config_path, "w") as f:
json.dump({"model_type": "llama"}, f)
result = detect_quantization_method(tmpdir)
self.assertIsNone(result)
def test_returns_none_for_malformed_config_json(self):
with tempfile.TemporaryDirectory() as tmpdir:
config_path = os.path.join(tmpdir, "config.json")
with open(config_path, "w") as f:
f.write("not valid json{{{")
result = detect_quantization_method(tmpdir)
self.assertIsNone(result)
def test_modelslim_takes_priority_over_compressed_tensors(self):
"""When both ModelSlim config and compressed-tensors config exist,
ModelSlim should take priority."""
with tempfile.TemporaryDirectory() as tmpdir:
modelslim_path = os.path.join(tmpdir, MODELSLIM_CONFIG_FILENAME)
with open(modelslim_path, "w") as f:
json.dump({"layer.weight": "INT8"}, f)
config_path = os.path.join(tmpdir, "config.json")
with open(config_path, "w") as f:
json.dump({
"quantization_config": {
"quant_method": "compressed-tensors"
}
}, f)
result = detect_quantization_method(tmpdir)
self.assertEqual(result, ASCEND_QUANTIZATION_METHOD)
class TestMaybeAutoDetectQuantization(TestBase):
def _make_vllm_config(self, model_path="/fake/model",
quantization=None, revision=None):
vllm_config = MagicMock()
vllm_config.model_config.model = model_path
vllm_config.model_config.quantization = quantization
vllm_config.model_config.revision = revision
return vllm_config
@patch("vllm_ascend.quantization.utils.detect_quantization_method",
return_value=None)
def test_no_detection_does_nothing(self, mock_detect):
vllm_config = self._make_vllm_config()
maybe_auto_detect_quantization(vllm_config)
self.assertIsNone(vllm_config.model_config.quantization)
@patch("vllm_ascend.quantization.utils.detect_quantization_method",
return_value=ASCEND_QUANTIZATION_METHOD)
def test_user_specified_same_method_no_change(self, mock_detect):
vllm_config = self._make_vllm_config(
quantization=ASCEND_QUANTIZATION_METHOD)
maybe_auto_detect_quantization(vllm_config)
self.assertEqual(vllm_config.model_config.quantization,
ASCEND_QUANTIZATION_METHOD)
@patch("vllm.config.VllmConfig._get_quantization_config",
return_value=MagicMock())
@patch("vllm_ascend.quantization.utils.detect_quantization_method",
return_value=ASCEND_QUANTIZATION_METHOD)
def test_auto_detect_sets_quantization_and_logs_info(
self, mock_detect, mock_get_quant_config):
"""When no --quantization is specified but ModelSlim config is found,
the method should auto-set quantization and emit an INFO log."""
vllm_config = self._make_vllm_config(
model_path="/fake/quant_model", quantization=None)
with self.assertLogs("vllm_ascend.quantization.utils",
level=logging.INFO) as cm:
maybe_auto_detect_quantization(vllm_config)
self.assertEqual(vllm_config.model_config.quantization,
ASCEND_QUANTIZATION_METHOD)
log_output = "\n".join(cm.output)
self.assertIn("Auto-detected quantization method", log_output)
self.assertIn(ASCEND_QUANTIZATION_METHOD, log_output)
self.assertIn("/fake/quant_model", log_output)
@patch("vllm_ascend.quantization.utils.detect_quantization_method",
return_value=ASCEND_QUANTIZATION_METHOD)
def test_user_mismatch_logs_warning(self, mock_detect):
"""When user specifies a different method than auto-detected,
a WARNING should be emitted and user's choice should be respected."""
vllm_config = self._make_vllm_config(
model_path="/fake/quant_model",
quantization=COMPRESSED_TENSORS_METHOD)
with self.assertLogs("vllm_ascend.quantization.utils",
level=logging.WARNING) as cm:
maybe_auto_detect_quantization(vllm_config)
self.assertEqual(vllm_config.model_config.quantization,
COMPRESSED_TENSORS_METHOD)
log_output = "\n".join(cm.output)
self.assertIn("Auto-detected quantization method", log_output)
self.assertIn(ASCEND_QUANTIZATION_METHOD, log_output)
self.assertIn(COMPRESSED_TENSORS_METHOD, log_output)
@patch("vllm_ascend.quantization.utils.detect_quantization_method",
return_value=None)
def test_no_detection_emits_no_log(self, mock_detect):
"""When no quantization is detected, no log should be emitted."""
vllm_config = self._make_vllm_config(quantization=None)
logger_name = "vllm_ascend.quantization.utils"
with self.assertRaises(AssertionError):
# assertLogs raises AssertionError when no logs are emitted
with self.assertLogs(logger_name, level=logging.DEBUG):
maybe_auto_detect_quantization(vllm_config)
self.assertIsNone(vllm_config.model_config.quantization)
@patch("vllm.config.VllmConfig._get_quantization_config",
return_value=MagicMock())
@patch("vllm_ascend.quantization.utils.detect_quantization_method",
return_value=ASCEND_QUANTIZATION_METHOD)
def test_passes_revision_to_detect(self, mock_detect, mock_get_quant):
"""Verify that model revision is forwarded to detect_quantization_method."""
vllm_config = self._make_vllm_config(
model_path="org/model-name", revision="v1.0", quantization=None)
maybe_auto_detect_quantization(vllm_config)
mock_detect.assert_called_once_with("org/model-name", revision="v1.0")

24
tests/ut/test_platform.py
View File

@@ -125,13 +125,14 @@ class TestNPUPlatform(TestBase):
self.assertIsNone(self.platform.inference_mode())
mock_inference_mode.assert_called_once()
@patch("vllm_ascend.quantization.utils.maybe_auto_detect_quantization")
@patch("vllm_ascend.ascend_config.init_ascend_config")
@patch("vllm_ascend.utils.update_aclgraph_sizes")
@patch("vllm_ascend.utils.get_ascend_device_type", return_value=AscendDeviceType.A3)
@patch("os.environ", {})
@patch("vllm_ascend.core.recompute_scheduler.RecomputeSchedulerConfig.initialize_from_config")
def test_check_and_update_config_basic_config_update(
self, mock_init_recompute, mock_soc_version, mock_update_acl, mock_init_ascend
self, mock_init_recompute, mock_soc_version, mock_update_acl, mock_init_ascend, mock_auto_detect
):
mock_init_ascend.return_value = TestNPUPlatform.mock_vllm_ascend_config()
vllm_config = TestNPUPlatform.mock_vllm_config()
@@ -155,11 +156,12 @@ class TestNPUPlatform(TestBase):
mock_init_ascend.assert_called_once_with(vllm_config)
@patch("vllm_ascend.quantization.utils.maybe_auto_detect_quantization")
@patch("vllm_ascend.utils.get_ascend_device_type", return_value=AscendDeviceType.A3)
@patch("vllm_ascend.ascend_config.init_ascend_config")
@patch("vllm_ascend.core.recompute_scheduler.RecomputeSchedulerConfig.initialize_from_config")
def test_check_and_update_config_no_model_config_warning(
self, mock_init_recompute, mock_init_ascend, mock_soc_version
self, mock_init_recompute, mock_init_ascend, mock_soc_version, mock_auto_detect
):
mock_init_ascend.return_value = TestNPUPlatform.mock_vllm_ascend_config()
vllm_config = TestNPUPlatform.mock_vllm_config()
@@ -181,10 +183,11 @@ class TestNPUPlatform(TestBase):
self.assertTrue("Model config is missing" in cm.output[0])
@patch("vllm_ascend.quantization.utils.maybe_auto_detect_quantization")
@patch("vllm_ascend.utils.get_ascend_device_type", return_value=AscendDeviceType.A3)
@patch("vllm_ascend.ascend_config.init_ascend_config")
@patch("vllm_ascend.core.recompute_scheduler.RecomputeSchedulerConfig.initialize_from_config")
def test_check_and_update_config_enforce_eager_mode(self, mock_init_recompute, mock_init_ascend, mock_soc_version):
def test_check_and_update_config_enforce_eager_mode(self, mock_init_recompute, mock_init_ascend, mock_soc_version, mock_auto_detect):
mock_init_ascend.return_value = TestNPUPlatform.mock_vllm_ascend_config()
vllm_config = TestNPUPlatform.mock_vllm_config()
vllm_config.model_config.enforce_eager = True
@@ -215,11 +218,12 @@ class TestNPUPlatform(TestBase):
CUDAGraphMode.NONE,
)
@patch("vllm_ascend.quantization.utils.maybe_auto_detect_quantization")
@patch("vllm_ascend.utils.get_ascend_device_type", return_value=AscendDeviceType.A3)
@patch("vllm_ascend.ascend_config.init_ascend_config")
@patch("vllm_ascend.core.recompute_scheduler.RecomputeSchedulerConfig.initialize_from_config")
def test_check_and_update_config_unsupported_compilation_level(
self, mock_init_recompute, mock_init_ascend, mock_soc_version
self, mock_init_recompute, mock_init_ascend, mock_soc_version, mock_auto_detect
):
mock_init_ascend.return_value = TestNPUPlatform.mock_vllm_ascend_config()
vllm_config = TestNPUPlatform.mock_vllm_config()
@@ -253,9 +257,10 @@ class TestNPUPlatform(TestBase):
)
@pytest.mark.skip("Revert me when vllm support setting cudagraph_mode on oot platform")
@patch("vllm_ascend.quantization.utils.maybe_auto_detect_quantization")
@patch("vllm_ascend.utils.get_ascend_device_type", return_value=AscendDeviceType.A3)
@patch("vllm_ascend.ascend_config.init_ascend_config")
def test_check_and_update_config_unsupported_cudagraph_mode(self, mock_init_ascend, mock_soc_version):
def test_check_and_update_config_unsupported_cudagraph_mode(self, mock_init_ascend, mock_soc_version, mock_auto_detect):
mock_init_ascend.return_value = TestNPUPlatform.mock_vllm_ascend_config()
vllm_config = TestNPUPlatform.mock_vllm_config()
vllm_config.model_config.enforce_eager = False
@@ -277,11 +282,12 @@ class TestNPUPlatform(TestBase):
CUDAGraphMode.NONE,
)
@patch("vllm_ascend.quantization.utils.maybe_auto_detect_quantization")
@patch("vllm_ascend.utils.get_ascend_device_type", return_value=AscendDeviceType.A3)
@patch("vllm_ascend.ascend_config.init_ascend_config")
@patch("vllm_ascend.core.recompute_scheduler.RecomputeSchedulerConfig.initialize_from_config")
def test_check_and_update_config_cache_config_block_size(
self, mock_init_recompute, mock_init_ascend, mock_soc_version
self, mock_init_recompute, mock_init_ascend, mock_soc_version, mock_auto_detect
):
mock_init_ascend.return_value = TestNPUPlatform.mock_vllm_ascend_config()
vllm_config = TestNPUPlatform.mock_vllm_config()
@@ -301,11 +307,12 @@ class TestNPUPlatform(TestBase):
self.assertEqual(vllm_config.cache_config.block_size, 128)
@patch("vllm_ascend.quantization.utils.maybe_auto_detect_quantization")
@patch("vllm_ascend.utils.get_ascend_device_type", return_value=AscendDeviceType.A3)
@patch("vllm_ascend.ascend_config.init_ascend_config")
@patch("vllm_ascend.core.recompute_scheduler.RecomputeSchedulerConfig.initialize_from_config")
def test_check_and_update_config_v1_worker_class_selection(
self, mock_init_recompute, mock_init_ascend, mock_soc_version
self, mock_init_recompute, mock_init_ascend, mock_soc_version, mock_auto_detect
):
mock_init_ascend.return_value = TestNPUPlatform.mock_vllm_ascend_config()
vllm_config = TestNPUPlatform.mock_vllm_config()
@@ -336,10 +343,11 @@ class TestNPUPlatform(TestBase):
"vllm_ascend.xlite.xlite_worker.XliteWorker",
)
@patch("vllm_ascend.quantization.utils.maybe_auto_detect_quantization")
@patch("vllm_ascend.ascend_config.init_ascend_config")
@patch("vllm_ascend.utils.get_ascend_device_type", return_value=AscendDeviceType._310P)
@patch("vllm_ascend.core.recompute_scheduler.RecomputeSchedulerConfig.initialize_from_config")
def test_check_and_update_config_310p_no_custom_ops(self, mock_init_recompute, mock_soc_version, mock_init_ascend):
def test_check_and_update_config_310p_no_custom_ops(self, mock_init_recompute, mock_soc_version, mock_init_ascend, mock_auto_detect):
mock_init_ascend.return_value = TestNPUPlatform.mock_vllm_ascend_config()
vllm_config = TestNPUPlatform.mock_vllm_config()
vllm_config.compilation_config.custom_ops = []