Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

support more models in openvino export #628

Merged
merged 5 commits into from
Apr 8, 2024
Merged
Show file tree
Hide file tree
Changes from 3 commits
Commits
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
15 changes: 14 additions & 1 deletion optimum/exporters/openvino/__main__.py
Original file line number Diff line number Diff line change
Expand Up @@ -202,7 +202,6 @@ def main_export(
quantization_config = getattr(config, "quantization_config", None)
do_gptq_patching = quantization_config and quantization_config["quant_method"] == "gptq"
model_type = config.model_type.replace("_", "-")

if model_type not in TasksManager._SUPPORTED_MODEL_TYPE:
custom_architecture = True
elif task not in TasksManager.get_supported_tasks_for_model_type(
Expand All @@ -220,6 +219,20 @@ def main_export(
)
if is_transformers_version(">=", "4.36") and model_type in SDPA_ARCHS_ONNX_EXPORT_NOT_SUPPORTED:
loading_kwargs["attn_implementation"] = "eager"
# there are some difference between remote and in library representation of past key values for some models,
# for avoiding confusion we disable remote code for them
if (
trust_remote_code
and model_type in {"falcon", "mpt", "phi"}
and ("with-past" in task or original_task == "auto")
and not custom_export_configs
):
logger.warning(
"Model type `{model_type}` export for task `{task}` is not supported for loading with `trust_remote_code=True`"
"using default export configuration, `trust_remote_code` will be disabled. "
"Please provide custon export config if you want load model with remote code."
)
trust_remote_code = False

# Patch the modules to export of GPTQ models w/o GPU
if do_gptq_patching:
Expand Down
2 changes: 1 addition & 1 deletion optimum/exporters/openvino/convert.py
Original file line number Diff line number Diff line change
Expand Up @@ -345,7 +345,7 @@ def ts_patched_forward(*args, **kwargs):
input_dict = dict(zip(keys, tuple_input))
kwargs[input_name] = input_dict
outputs = patched_forward(*args, **kwargs)
return tuple(outputs.values())
return tuple([value if not isinstance(value, list) else tuple(value) for value in outputs.values()])

patcher.patched_forward = ts_patched_forward

Expand Down
20 changes: 19 additions & 1 deletion optimum/exporters/openvino/model_configs.py
Original file line number Diff line number Diff line change
Expand Up @@ -74,7 +74,7 @@ def init_model_configs():


@register_in_tasks_manager("baichuan", *["text-generation", "text-generation-with-past"], library_name="transformers")
class BaichaunOpenVINOConfig(TextDecoderOnnxConfig):
class BaichaunOpenVINOConfig(TextDecoderWithPositionIdsOnnxConfig):
DEFAULT_ONNX_OPSET = 13
NORMALIZED_CONFIG_CLASS = NormalizedTextConfig.with_args(
num_layers="num_hidden_layers", num_attention_heads="num_attention_heads", hidden_size="hidden_size"
Expand Down Expand Up @@ -400,3 +400,21 @@ class Starcoder2OpenVINOConfig(TextDecoderWithPositionIdsOnnxConfig):
DUMMY_INPUT_GENERATOR_CLASSES = (DummyTextInputGenerator, MistralDummyPastKeyValuesGenerator)
DUMMY_PKV_GENERATOR_CLASS = MistralDummyPastKeyValuesGenerator
NORMALIZED_CONFIG_CLASS = NormalizedTextConfig


@register_in_tasks_manager("internlm2", *["text-generation", "text-generation-with-past"], library_name="transformers")
class InternLM2OpenVINOConfig(TextDecoderWithPositionIdsOnnxConfig):
DEFAULT_ONNX_OPSET = 14

DUMMY_INPUT_GENERATOR_CLASSES = (DummyTextInputGenerator, MistralDummyPastKeyValuesGenerator)
DUMMY_PKV_GENERATOR_CLASS = MistralDummyPastKeyValuesGenerator
NORMALIZED_CONFIG_CLASS = NormalizedTextConfig


@register_in_tasks_manager("orion", *["text-generation", "text-generation-with-past"], library_name="transformers")
class OrionOpenVINOConfig(TextDecoderWithPositionIdsOnnxConfig):
DEFAULT_ONNX_OPSET = 14

DUMMY_INPUT_GENERATOR_CLASSES = (DummyTextInputGenerator, MistralDummyPastKeyValuesGenerator)
DUMMY_PKV_GENERATOR_CLASS = MistralDummyPastKeyValuesGenerator
NORMALIZED_CONFIG_CLASS = NormalizedTextConfig
2 changes: 1 addition & 1 deletion optimum/exporters/openvino/model_patcher.py
Original file line number Diff line number Diff line change
Expand Up @@ -513,5 +513,5 @@ def __init__(
):
super().__init__(config, model, model_kwargs)
# model has first inference buffers initialization
if self._model.lm_head.first_flag:
if hasattr(self._model.lm_head, "first_flag"):
self._model(torch.ones((1, 10), dtype=torch.int64), torch.ones((1, 10), dtype=torch.int64))
4 changes: 3 additions & 1 deletion tests/openvino/test_modeling.py
Original file line number Diff line number Diff line change
Expand Up @@ -524,10 +524,12 @@ class OVModelForCausalLMIntegrationTest(unittest.TestCase):
"stablelm",
"starcoder2",
"phi",
"internlm2",
"orion",
)
GENERATION_LENGTH = 100
IS_SUPPORT_STATEFUL = is_openvino_version(">=", "2023.3")
REMOTE_CODE_MODELS = ("chatglm", "minicpm", "baichuan2", "jais", "qwen")
REMOTE_CODE_MODELS = ("chatglm", "minicpm", "baichuan2", "jais", "qwen", "internlm2", "olmo", "orion")

@parameterized.expand(SUPPORTED_ARCHITECTURES)
def test_compare_to_transformers(self, model_arch):
Expand Down
3 changes: 3 additions & 0 deletions tests/openvino/utils_tests.py
Original file line number Diff line number Diff line change
Expand Up @@ -50,6 +50,7 @@
"gptj": "hf-internal-testing/tiny-random-GPTJModel",
"hubert": "hf-internal-testing/tiny-random-HubertModel",
"ibert": "hf-internal-testing/tiny-random-ibert",
"internlm2": "katuni4ka/tiny-random-internlm2",
"levit": "hf-internal-testing/tiny-random-LevitModel",
"longt5": "hf-internal-testing/tiny-random-longt5",
"llama": "fxmarty/tiny-llama-fast-tokenizer",
Expand All @@ -69,6 +70,8 @@
"mpt": "hf-internal-testing/tiny-random-MptForCausalLM",
"mt5": "stas/mt5-tiny-random",
"nystromformer": "hf-internal-testing/tiny-random-NystromformerModel",
"olmo": "katuni4ka/tiny-random-olmo",
"orion": "katuni4ka/tiny-random-orion",
"pegasus": "hf-internal-testing/tiny-random-pegasus",
"pix2struct": "fxmarty/pix2struct-tiny-random",
"phi": "echarlaix/tiny-random-PhiForCausalLM",
Expand Down
Loading