[OpenVINO] support ai-sage/GigaChat3-10B-A1.8B-bf16

docs/source/openvino/models.mdx

@@ -61,6 +61,7 @@ Here is the list of the supported architectures :
 - Falcon
 - Falcon-Mamba
 - Flaubert
+- GigaChat3
 - GLM-4
 - GLM-Edge
 - GPT-2

optimum/exporters/openvino/convert.py

@@ -721,7 +721,27 @@ def export_from_model(
         generation_config = getattr(model, "generation_config", None)
         if generation_config is not None:
             try:
+                # Preserve the original `transformers_version` from the source model's generation_config.json.
+                # Starting in transformers 4.50, _prepare_generation_config() applies model-default generation
+                # parameters (do_sample, temperature, top_p, …) when the user-provided GenerationConfig uses
+                # the global default value for those fields AND the stored `transformers_version` is >= 4.50.
+                # Exporting bumps the version to the current transformers release, which causes user-supplied
+                # params (e.g. do_sample=False) to be silently overridden by the model defaults at inference
+                # time. Preserving the original version keeps the OV model consistent with the PT original.
+                orig_transformers_version = getattr(generation_config, "transformers_version", None)
                 generation_config.save_pretrained(output)
+                if orig_transformers_version is not None:
+                    import json as _json
+                    from pathlib import Path as _Path
+
+                    gen_cfg_path = _Path(output) / "generation_config.json"
+                    if gen_cfg_path.exists():
+                        with open(gen_cfg_path, "r", encoding="utf-8") as _f:
+                            _cfg = _json.load(_f)
+                        if _cfg.get("transformers_version") != orig_transformers_version:
+                            _cfg["transformers_version"] = orig_transformers_version
+                            with open(gen_cfg_path, "w", encoding="utf-8") as _f:
+                                _json.dump(_cfg, _f, indent=2)
             except Exception as exception:
                 logger.warning(
                     f"The generation config will not be saved, saving failed with following error:\n{exception}"

optimum/exporters/openvino/model_configs.py

@@ -4081,8 +4081,8 @@ class M2M100OpenVINOConfig(BartOpenVINOConfig):
 )
 @register_in_tasks_manager("deepseek", *["text-generation", "text-generation-with-past"], library_name="transformers")
 class DeepseekOpenVINOConfig(MiniCPM3OpenVINOConfig):
-    MIN_TRANSFORMERS_VERSION = "4.46.0"
-    MAX_TRANSFORMERS_VERSION = "4.53.3"
+    MIN_TRANSFORMERS_VERSION = "4.53.0"
+    MAX_TRANSFORMERS_VERSION = None
     _MODEL_PATCHER = DeepseekPatcher
 
 

tests/openvino/test_decoder.py

@@ -119,13 +119,13 @@ class OVModelForCausalLMIntegrationTest(unittest.TestCase):
     if is_transformers_version(">=", "4.46.0"):
         SUPPORTED_ARCHITECTURES += ("glm", "mistral-nemo", "phimoe")
 
-        if is_transformers_version("<", "4.54.0"):
-            SUPPORTED_ARCHITECTURES += ("deepseek",)
-
         # gptq and awq install disabled for windows test environment
         if platform.system() != "Windows" and is_transformers_version("<", "4.56.0"):
             SUPPORTED_ARCHITECTURES += ("opt_gptq", "mixtral_awq")
 
+    if is_transformers_version(">=", "4.53.0"):
+        SUPPORTED_ARCHITECTURES += ("deepseek", "gigachat3")
+
     if is_transformers_version(">", "4.47"):
         SUPPORTED_ARCHITECTURES += ("olmo2",)
 
@@ -230,6 +230,7 @@ class OVModelForCausalLMIntegrationTest(unittest.TestCase):
         "minicpm3": 6,
         "phimoe": 2,
         "deepseek": 2,
+        "gigachat3": 2,
         "opt_gptq": 12,
         "mixtral_awq": 2,
         "gemma3_text": 2,
@@ -290,11 +291,8 @@ def test_find_untested_architectures(self):
 
         if "llama4_text" in supported_architectures:
             supported_architectures.remove("llama4_text")
-        if is_transformers_version(">=", str(DeepseekOpenVINOConfig.MAX_TRANSFORMERS_VERSION)):
-            if "deepseek_v2" in supported_architectures:
-                supported_architectures.remove("deepseek_v2")
-            if "deepseek_v3" in supported_architectures:
-                supported_architectures.remove("deepseek_v3")
+        if is_transformers_version(">=", str(DeepseekOpenVINOConfig.MIN_TRANSFORMERS_VERSION)):
+            supported_architectures -= {"deepseek_v2", "deepseek_v3"}
         if is_transformers_version("<", str(BitnetOpenVINOConfig.MIN_TRANSFORMERS_VERSION)):
             supported_architectures -= {"bitnet"}
         if is_transformers_version("<", str(LFM2OpenVINOConfig.MIN_TRANSFORMERS_VERSION)):
@@ -397,6 +395,12 @@ def test_compare_to_transformers(self, model_arch):
             return
 
         tokens = tokenizer(["Today is a nice day and I am longer", "This is me"], return_tensors="pt", padding=True)
+
+        # Gigachat3 tokenizer add token_type_ids which DeepSeekV3
+        # and similar models do not accept in generate(); strip it so both OV and PT calls succeed.
+        if model_arch in ["gigachat3"]:
+            tokens.pop("token_type_ids", None)
+
         ov_model.generation_config.eos_token_id = None
         transformers_model.generation_config.eos_token_id = None
         ov_model.config.eos_token_id = None
@@ -413,10 +417,6 @@ 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 modeling 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
@@ -666,10 +666,6 @@ def test_beam_search(self, model_arch):
         if model_arch in ["lfm2", "granitemoehybrid"]:
             return
 
-        # TODO: add back once https://huggingface.co/katuni4ka/tiny-random-minicpm3/discussions/1 merged (for all models) as current modeling 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 REMOTE_CODE_MODELS)
         if model_arch == "persimmon":
             tokenizer.pad_token_id = tokenizer.bos_token_id

tests/openvino/test_export.py

@@ -98,6 +98,9 @@ class ExportModelTest(unittest.TestCase):
     if is_transformers_version(">=", "4.48.0"):
         SUPPORTED_ARCHITECTURES.update({"cohere2": OVModelForCausalLM})
 
+    if is_transformers_version(">=", "4.53.0"):
+        SUPPORTED_ARCHITECTURES.update({"deepseek": OVModelForCausalLM, "gigachat3": OVModelForCausalLM})
+
     if is_transformers_version(">=", "4.49"):
         SUPPORTED_ARCHITECTURES.update({"zamba2": OVModelForCausalLM})
 

tests/openvino/test_exporters_cli.py

@@ -159,6 +159,13 @@ class OVCLIExportTestCase(unittest.TestCase):
             ]
         )
 
+    if is_transformers_version(">=", "4.53.0"):
+        SUPPORTED_ARCHITECTURES.extend(
+            [
+                ("text-generation-with-past", "gigachat3"),
+            ]
+        )
+
     if is_transformers_version(">=", "4.57.0"):
         SUPPORTED_ARCHITECTURES.extend(
             [
@@ -198,6 +205,7 @@ class OVCLIExportTestCase(unittest.TestCase):
         "exaone4": 2,
         "bitnet": 2,
         "granitemoehybrid": 2,
+        "gigachat3": 2,
     }
 
     TOKENIZER_CHAT_TEMPLATE_TESTS_MODELS = {

tests/openvino/utils_tests.py

@@ -70,6 +70,7 @@
     "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",
+    "gigachat3": "optimum-intel-internal-testing/tiny-random-gigachat3",
     "deit": "optimum-intel-internal-testing/tiny-random-DeiTModel",
     "convnext": "optimum-intel-internal-testing/tiny-random-convnext",
     "convnextv2": "optimum-intel-internal-testing/tiny-random-ConvNextV2Model",
@@ -372,6 +373,7 @@
     "hunyuan_v1_dense": {"model": 32},
     "qwen3_eagle3": {"model": 20},
     "qwen3_next": {"model": 100},
+    "gigachat3": {"model": 58},
 }
 
 TEST_IMAGE_URL = "http://images.cocodataset.org/val2017/000000039769.jpg"
@@ -397,7 +399,6 @@
     "exaone4",
     "decilm",
     "minicpm3",
-    "deepseek",
     "qwen3_eagle3",
 )
 
@@ -552,8 +553,10 @@ def get_supported_model_for_library(library_name):
         if supported_model_type[model_type].get("openvino"):
             export_config = next(iter(supported_model_type[model_type]["openvino"].values()))
 
-            min_transformers = str(getattr(export_config.func, "MIN_TRANSFORMERS_VERSION", "0"))
-            max_transformers = str(getattr(export_config.func, "MAX_TRANSFORMERS_VERSION", "999"))
+            raw_min = getattr(export_config.func, "MIN_TRANSFORMERS_VERSION", None)
+            raw_max = getattr(export_config.func, "MAX_TRANSFORMERS_VERSION", None)
+            min_transformers = str(raw_min) if raw_min is not None else "0"
+            max_transformers = str(raw_max) if raw_max is not None else "999"
 
             if is_transformers_version(">=", min_transformers) and is_transformers_version("<=", max_transformers):
                 valid_model.add(model_type)