biondizzle/vllm
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

[Core][Model] Terratorch backend integration (#23513)

Browse Source 
Signed-off-by: Michele Gazzetti <michele.gazzetti1@ibm.com>
Signed-off-by: Christian Pinto <christian.pinto@ibm.com>
Co-authored-by: Christian Pinto <christian.pinto@ibm.com>
Co-authored-by: Cyrus Leung <tlleungac@connect.ust.hk>
This commit is contained in:
mgazz
2025-09-04 08:22:41 +01:00
committed by GitHub
parent e7fc70016f
commit 51d5e9be7d
23 changed files with 305 additions and 208 deletions
Download Patch File Download Diff File Expand all files Collapse all files

3
tests/models/multimodal/generation/vlm_utils/core.py
View File

@@ -69,6 +69,9 @@ def run_test(
vllm_runner_kwargs_["tokenizer_mode"] = model_info.tokenizer_mode
if model_info.hf_overrides:
vllm_runner_kwargs_["hf_overrides"] = model_info.hf_overrides
if model_info.skip_tokenizer_init:
vllm_runner_kwargs_[
"skip_tokenizer_init"] = model_info.skip_tokenizer_init
if vllm_runner_kwargs:
vllm_runner_kwargs_.update(vllm_runner_kwargs)

2
tests/models/multimodal/pooling/test_prithvi_mae.py
View File

@@ -46,7 +46,7 @@ def _run_test(
vllm_model.encode(prompt)
MODELS = ["christian-pinto/Prithvi-EO-2.0-300M-TL-VLLM"]
MODELS = ["mgazz/Prithvi-EO-2.0-300M-TL-Sen1Floods11"]
@pytest.mark.core_model

4
tests/models/multimodal/processing/test_common.py
View File

@@ -66,7 +66,9 @@ def _test_processing_correctness(
hf_overrides=model_info.hf_overrides,
# Ensure that the cache can fit all of the data
mm_processor_cache_gb=2048,
)
skip_tokenizer_init=model_info.skip_tokenizer_init,
enforce_eager=model_info.enforce_eager,
dtype=model_info.dtype)
model_cls = MULTIMODAL_REGISTRY._get_model_cls(model_config)
factories = MULTIMODAL_REGISTRY._processor_factories[model_cls]

4
tests/models/multimodal/processing/test_tensor_schema.py
View File

@@ -196,7 +196,9 @@ def test_model_tensor_schema(model_arch: str, model_id: str):
revision=model_info.revision,
trust_remote_code=model_info.trust_remote_code,
hf_overrides=hf_overrides_fn,
)
skip_tokenizer_init=model_info.skip_tokenizer_init,
enforce_eager=model_info.enforce_eager,
dtype=model_info.dtype)
model_cls = MULTIMODAL_REGISTRY._get_model_cls(model_config)
factories = MULTIMODAL_REGISTRY._processor_factories[model_cls]

4
tests/models/multimodal/test_mapping.py
View File

@@ -59,7 +59,9 @@ def test_hf_model_weights_mapper(model_arch: str):
revision=model_info.revision,
trust_remote_code=model_info.trust_remote_code,
hf_overrides=model_info.hf_overrides,
)
skip_tokenizer_init=model_info.skip_tokenizer_init,
enforce_eager=model_info.enforce_eager,
dtype=model_info.dtype)
model_cls = MULTIMODAL_REGISTRY._get_model_cls(model_config)
original_weights = create_repo_dummy_weights(model_id)

40
tests/models/registry.py
View File

@@ -6,10 +6,11 @@ from dataclasses import dataclass, field
from typing import Any, Literal, Optional
import pytest
import torch
from packaging.version import Version
from transformers import __version__ as TRANSFORMERS_VERSION
from vllm.config import TokenizerMode
from vllm.config import ModelDType, TokenizerMode
@dataclass(frozen=True)
@@ -47,6 +48,23 @@ class _HfExamplesInfo:
The reason for the minimum/maximum version requirement.
"""
skip_tokenizer_init: bool = False
"""
If true, skip initialization of tokenizer and detokenizer.
"""
dtype: ModelDType = "auto"
"""
The data type for the model weights and activations.
"""
enforce_eager: bool = False
"""
Whether to enforce eager execution. If True, we will
disable CUDA graph and always execute the model in eager mode.
If False, we will use CUDA graph and eager execution in hybrid.
"""
is_available_online: bool = True
"""
Set this to ``False`` if the name of this architecture no longer exists on
@@ -76,6 +94,9 @@ class _HfExamplesInfo:
If not specified, the default revision will be used.
"""
max_num_seqs: Optional[int] = None
"""Maximum number of sequences to be processed in a single iteration."""
def check_transformers_version(
self,
*,
@@ -361,8 +382,21 @@ _EMBEDDING_EXAMPLE_MODELS = {
"Phi3VForCausalLM": _HfExamplesInfo("TIGER-Lab/VLM2Vec-Full",
trust_remote_code=True),
"Qwen2VLForConditionalGeneration": _HfExamplesInfo("MrLight/dse-qwen2-2b-mrl-v1"), # noqa: E501
"PrithviGeoSpatialMAE": _HfExamplesInfo("ibm-nasa-geospatial/Prithvi-EO-2.0-300M-TL-Sen1Floods11", # noqa: E501
is_available_online=False), # noqa: E501
"PrithviGeoSpatialMAE": _HfExamplesInfo("mgazz/Prithvi-EO-2.0-300M-TL-Sen1Floods11", # noqa: E501
dtype=torch.float16,
enforce_eager=True,
skip_tokenizer_init=True,
# This is to avoid the model
# going OOM in CI
max_num_seqs=32,
),
"Terratorch": _HfExamplesInfo("mgazz/Prithvi-EO-2.0-300M-TL-Sen1Floods11",
dtype=torch.float16,
enforce_eager=True,
skip_tokenizer_init=True,
# This is to avoid the model going OOM in CI
max_num_seqs=32,
),
}
_SEQUENCE_CLASSIFICATION_EXAMPLE_MODELS = {

5
tests/models/test_initialization.py
View File

@@ -73,6 +73,9 @@ def can_initialize(model_arch: str, monkeypatch: pytest.MonkeyPatch,
tokenizer=model_info.tokenizer,
tokenizer_mode=model_info.tokenizer_mode,
revision=model_info.revision,
enforce_eager=model_info.enforce_eager,
skip_tokenizer_init=model_info.skip_tokenizer_init,
dtype=model_info.dtype,
speculative_config={
"model": model_info.speculative_model,
"num_speculative_tokens": 1,
@@ -85,7 +88,7 @@ def can_initialize(model_arch: str, monkeypatch: pytest.MonkeyPatch,
model_impl=ModelImpl.TRANSFORMERS
if model_arch in _TRANSFORMERS_BACKEND_MODELS else ModelImpl.VLLM,
hf_overrides=hf_overrides_fn,
)
max_num_seqs=model_info.max_num_seqs)
@pytest.mark.parametrize("model_arch", HF_EXAMPLE_MODELS.get_supported_archs())

45
tests/models/test_terratorch.py Normal file
View File

@@ -0,0 +1,45 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
import pytest
import torch
from tests.conftest import VllmRunner
from vllm.utils import set_default_torch_num_threads
@pytest.mark.parametrize(
"model",
[
"mgazz/Prithvi-EO-2.0-300M-TL-Sen1Floods11",
"mgazz/Prithvi_v2_eo_300_tl_unet_agb"
],
)
def test_inference(
vllm_runner: type[VllmRunner],
model: str,
) -> None:
pixel_values = torch.full((6, 512, 512), 1.0, dtype=torch.float16)
location_coords = torch.full((1, 2), 1.0, dtype=torch.float16)
prompt = dict(prompt_token_ids=[1],
multi_modal_data=dict(pixel_values=pixel_values,
location_coords=location_coords))
with (
set_default_torch_num_threads(1),
vllm_runner(
model,
runner="pooling",
dtype=torch.float16,
enforce_eager=True,
skip_tokenizer_init=True,
# Limit the maximum number of sequences to avoid the
# test going OOM during the warmup run
max_num_seqs=32,
) as vllm_model,
):
vllm_output = vllm_model.llm.encode(prompt)
assert torch.equal(
torch.isnan(vllm_output[0].outputs.data).any(),
torch.tensor(False))

2
tests/models/utils.py
View File

@@ -294,6 +294,8 @@ def build_model_context(
limit_mm_per_prompt=limit_mm_per_prompt,
mm_processor_cache_gb=mm_processor_cache_gb,
hf_overrides=model_info.hf_overrides,
skip_tokenizer_init=model_info.skip_tokenizer_init,
enforce_eager=model_info.enforce_eager,
**model_config_kwargs,
)
return InputContext(model_config)