Skip to content

add deepseek v3 transformers integration #1445

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.

Sign up for GitHub

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

Draft
wants to merge 8 commits into
base: main
Choose a base branch
from
Draft

add deepseek v3 transformers integration #1445

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
8 commits
Select commit Hold shift + click to select a range
0717b05
add deepseek v3 transformers integration
echarlaix Sep 12, 2025
cce99e3
style
echarlaix Sep 12, 2025
6fb5d75
merge main
echarlaix Sep 19, 2025
d254284
modify tests min transformers version
echarlaix Sep 19, 2025
f959504
add tiny model
echarlaix Sep 22, 2025
856128b
add test sdpa
echarlaix Sep 22, 2025
185284a
merge main in branch
echarlaix Oct 27, 2025
4bf1788
was removed by accident
echarlaix Oct 27, 2025
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
20 changes: 17 additions & 3 deletions optimum/exporters/openvino/model_configs.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -3659,9 +3659,6 @@ class M2M100OpenVINOConfig(BartOpenVINOConfig):
pass


@register_in_tasks_manager(
"deepseek_v3", *["text-generation", "text-generation-with-past"], library_name="transformers"
)
@register_in_tasks_manager(
"deepseek_v2", *["text-generation", "text-generation-with-past"], library_name="transformers"
)
Expand All @@ -3671,6 +3668,15 @@ class DeepseekOpenVINOConfig(MiniCPM3OpenVINOConfig):
_MODEL_PATCHER = DeepseekPatcher


@register_in_tasks_manager(
"deepseek_v3", *["text-generation", "text-generation-with-past"], library_name="transformers"
)
class DeepseekVOpenVINOConfig(LlamaOpenVINOConfig):
Copy link
Collaborator Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

before we can merge : need to fix rope_interleave for export https://github.com/huggingface/transformers/blob/v4.55.4/src/transformers/models/deepseek_v3/modeling_deepseek_v3.py#L402

MIN_TRANSFORMERS_VERSION = "4.51.0"
DUMMY_INPUT_GENERATOR_CLASSES = (DummyTextInputGenerator, OVMiniCPM3DummyPastKeyValuesGenerator)
DUMMY_PKV_GENERATOR_CLASS = OVMiniCPM3DummyPastKeyValuesGenerator


@register_in_tasks_manager("got_ocr2", *["image-to-text", "image-text-to-text"], library_name="transformers")
class GotOCR2OpenVINOConfig(BaseVLMOpenVINOConfig):
MIN_TRANSFORMERS_VERSION = "4.49.0"
Expand Down Expand Up @@ -4251,6 +4257,14 @@ def generate_dummy_inputs(self, framework: str = "pt", **kwargs):
return dummy_inputs


@register_in_tasks_manager("ernie4_5", *["text-generation", "text-generation-with-past"], library_name="transformers")
class ErnieOpenVINOConfig(LlamaOpenVINOConfig):
MIN_TRANSFORMERS_VERSION = "4.54.0"
DUMMY_INPUT_GENERATOR_CLASSES = (DummyTextInputGenerator, GemmaDummyPastKeyValuesGenerator)
DUMMY_PKV_GENERATOR_CLASS = GemmaDummyPastKeyValuesGenerator
NORMALIZED_CONFIG_CLASS = NormalizedTextConfig


@register_in_tasks_manager(
"gpt2",
*[
Expand Down
18 changes: 4 additions & 14 deletions tests/openvino/test_decoder.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -93,9 +93,6 @@ class OVModelForCausalLMIntegrationTest(unittest.TestCase):

if is_transformers_version(">=", "4.46.0"):
SUPPORTED_ARCHITECTURES += ("glm", "mistral-nemo", "minicpm3", "phi3-moe")
# openvino 2025.0 required for disabling check_trace
if is_openvino_version(">=", "2025.0"):
SUPPORTED_ARCHITECTURES += ("deepseek",)
Copy link
Collaborator Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

actually remote code deepseek_v3 : katuni4ka/tiny-random-deepseek-v3


# gptq and awq install disabled for windows test environment
if platform.system() != "Windows":
Expand All @@ -109,7 +106,7 @@ class OVModelForCausalLMIntegrationTest(unittest.TestCase):
SUPPORTED_ARCHITECTURES += ("gemma3_text",)

if is_transformers_version(">=", "4.51.0"):
SUPPORTED_ARCHITECTURES += ("qwen3", "qwen3_moe")
SUPPORTED_ARCHITECTURES += ("qwen3", "qwen3_moe", "deepseek_v3")

if is_transformers_version(">=", "4.51.3"):
SUPPORTED_ARCHITECTURES += ("glm4",)
Expand All @@ -120,6 +117,7 @@ class OVModelForCausalLMIntegrationTest(unittest.TestCase):
if is_transformers_version(">=", "4.54.0"):
# remote code models differs after transformers v4.54
SUPPORTED_ARCHITECTURES = tuple(set(SUPPORTED_ARCHITECTURES) - {"minicpm", "minicpm3", "arctic", "deepseek"})
SUPPORTED_ARCHITECTURES += ("ernie4_5",)

if is_transformers_version(">=", "4.55.0"):
SUPPORTED_ARCHITECTURES += ("gpt_oss", "gpt_oss_mxfp4")
Expand All @@ -144,7 +142,6 @@ class OVModelForCausalLMIntegrationTest(unittest.TestCase):
"exaone",
"decilm",
"minicpm3",
"deepseek",
)

EXPECTED_NUM_SDPA = {
Expand All @@ -158,6 +155,8 @@ class OVModelForCausalLMIntegrationTest(unittest.TestCase):
"chatglm": 2,
"codegen": 5,
"codegen2": 2,
"deepseek_v3": 2,
"ernie4_5": 4,
"gpt2": 5,
"gptj": 5,
"gpt_neo": 4,
Expand Down Expand Up @@ -202,7 +201,6 @@ class OVModelForCausalLMIntegrationTest(unittest.TestCase):
"mistral-nemo": 8,
"minicpm3": 6,
"phi3-moe": 2,
"deepseek": 2,
"opt_gptq": 12,
"mixtral_awq": 2,
"gemma3_text": 2,
Expand Down Expand Up @@ -324,10 +322,6 @@ def test_compare_to_transformers(self, model_arch):

ov_outputs = ov_model.generate(**tokens, generation_config=gen_config)

# TODO: add back once https://huggingface.co/katuni4ka/tiny-random-minicpm3/discussions/1 merged (for all models) as current mdoeling incompatible with transformers >= v4.49
if model_arch in {"deepseek"} and is_transformers_version(">=", "4.49"):
self.skipTest("Incompatible modeling code")

additional_inputs = {}
# gemma2 does not support dynamic cache, it is unfair to compare dynamic cache result vs hybrid cache,
# align cache representation in torch model
Expand Down Expand Up @@ -569,10 +563,6 @@ def test_beam_search(self, model_arch):
if model_arch in ["qwen", "chatglm", "chatglm4"]:
return

# TODO: add back once https://huggingface.co/katuni4ka/tiny-random-minicpm3/discussions/1 merged (for all models) as current mdoeling incompatible with transformers >= v4.49
if model_arch in {"deepseek"} and is_transformers_version(">=", "4.49"):
self.skipTest("Incompatible modeling code")

tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=model_arch in self.REMOTE_CODE_MODELS)
if model_arch == "persimmon":
tokenizer.pad_token_id = tokenizer.bos_token_id
Expand Down
3 changes: 2 additions & 1 deletion tests/openvino/utils_tests.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -66,7 +66,7 @@
"deberta": "optimum-intel-internal-testing/tiny-random-deberta",
"deberta-v2": "optimum-intel-internal-testing/tiny-random-DebertaV2Model",
"decilm": "optimum-intel-internal-testing/tiny-random-decilm",
"deepseek": "optimum-intel-internal-testing/tiny-random-deepseek-v3",
"deepseek_v3": "hf-internal-testing/tiny-random-DeepseekV3ForCausalLM",
"deit": "optimum-intel-internal-testing/tiny-random-DeiTModel",
"convnext": "optimum-intel-internal-testing/tiny-random-convnext",
"convnextv2": "optimum-intel-internal-testing/tiny-random-ConvNextV2Model",
Expand All @@ -75,6 +75,7 @@
"donut-swin": "optimum-intel-internal-testing/tiny-random-DonutSwinModel",
"detr": "optimum-intel-internal-testing/tiny-random-DetrModel",
"electra": "optimum-intel-internal-testing/tiny-random-electra",
"ernie4_5": "optimum-internal-testing/tiny-random-Ernie4_5ForCausalLM",
"esm": "optimum-intel-internal-testing/tiny-random-EsmModel",
"exaone": "optimum-intel-internal-testing/tiny-random-exaone",
"gemma": "optimum-intel-internal-testing/tiny-random-GemmaForCausalLM",
Expand Down
Loading