ctranslate2 4.6.3__cp310-cp310-win_amd64.whl → 4.7.0__cp310-cp310-win_amd64.whl

ctranslate2/__init__.py

@@ -21,6 +21,8 @@ if sys.platform == "win32":
     add_dll_directory = getattr(os, "add_dll_directory", None)
     if add_dll_directory is not None:
         add_dll_directory(package_dir)
+        add_dll_directory(f"{package_dir}/../_rocm_sdk_core/bin")
+        add_dll_directory(f"{package_dir}/../_rocm_sdk_libraries_custom/bin")
 
     for library in glob.glob(os.path.join(package_dir, "*.dll")):
         ctypes.CDLL(library)

ctranslate2/converters/eole_ct2.py

@@ -3,7 +3,7 @@ import argparse
 from eole.config.run import PredictConfig
 from eole.constants import PositionEncodingType
 from eole.inputters.inputter import vocabs_to_dict
-from eole.models.model import BaseModel
+from eole.models.model import get_model_class
 
 from ctranslate2.converters import utils
 from ctranslate2.converters.converter import Converter
@@ -164,7 +164,8 @@ class EoleConverter(Converter):
 
         config = PredictConfig(model_path=self._model_path, src="dummy")
 
-        vocabs, model, model_config = BaseModel.load_test_model(config)
+        model_class = get_model_class(config.model)
+        model, vocabs, model_config = model_class.for_inference(config)
         vocabs_dict = vocabs_to_dict(vocabs)
 
         config.model = model_config

ctranslate2/converters/transformers.py

@@ -253,6 +253,30 @@ class ModelLoader(abc.ABC):
             "No activation smoothing logic is defined for this model"
         )
 
+    def get_rotary_params(self, config, default_rope_theta):
+        rope_scaling = getattr(config, "rope_scaling", None)
+        if rope_scaling:
+            rope_type = rope_scaling.get("type") or rope_scaling.get("rope_type")
+
+            if rope_type == "default":
+                rotary_scaling_type = None
+            else:
+                rotary_scaling_type = _SUPPORTED_ROPE_SCALING.get(rope_type)
+                if rotary_scaling_type is None:
+                    raise NotImplementedError(
+                        "RoPE scaling type '%s' is not yet implemented. "
+                        "The following RoPE scaling types are currently supported: %s"
+                        % (rope_type, ", ".join(_SUPPORTED_ROPE_SCALING.keys()))
+                    )
+            rotary_scaling_factor = rope_scaling.get("factor", 1)
+            rope_theta = rope_scaling.get("rope_theta", default_rope_theta)
+        else:
+            rotary_scaling_type = None
+            rotary_scaling_factor = 1
+            rope_theta = getattr(config, "rope_theta", default_rope_theta)
+
+        return rotary_scaling_type, rotary_scaling_factor, rope_theta
+
 
 @register_loader("BartConfig")
 class BartLoader(ModelLoader):
@@ -463,7 +487,7 @@ class M2M100Loader(BartLoader):
         if tokens[-1] == tokenizer.unk_token:
             tokens.insert(tokenizer.unk_token_id, tokens.pop())
 
-        for token in tokenizer.additional_special_tokens:
+        for token in tokenizer.special_tokens_map.get("additional_special_tokens", []):
             if token not in tokens:
                 tokens.append(token)
 
@@ -488,7 +512,7 @@ class MBartLoader(BartLoader):
         config.unk_token = tokenizer.unk_token
 
         # MBart-25 passes the language code as the decoder start token.
-        if model.config.tokenizer_class in ("MBartTokenizer", None):
+        if getattr(model.config, "tokenizer_class", None) in ("MBartTokenizer", None):
             config.decoder_start_token = None
         else:
             config.decoder_start_token = tokenizer.eos_token
@@ -928,12 +952,14 @@ class WhisperLoader(BartLoader):
             "<|nocaptions|>",
             "<|notimestamps|>",
         ]
+
+        additional_tokens = getattr(tokenizer, "additional_special_tokens", [])
+        if not additional_tokens:
+            return []
+
         return [
-            token_id
-            for token_id, token in zip(
-                tokenizer.additional_special_tokens_ids,
-                tokenizer.additional_special_tokens,
-            )
+            tokenizer.convert_tokens_to_ids(token)
+            for token in additional_tokens
             if token not in non_lang_special_tokens
         ]
 
@@ -1674,21 +1700,9 @@ class LlamaLoader(ModelLoader):
         if num_heads_kv == num_heads:
             num_heads_kv = None
 
-        rope_scaling = getattr(model.config, "rope_scaling", None)
-        if rope_scaling:
-            rope_type = rope_scaling.get("type") or rope_scaling["rope_type"]
-            rotary_scaling_type = _SUPPORTED_ROPE_SCALING.get(rope_type)
-            rotary_scaling_factor = rope_scaling["factor"]
-
-            if rotary_scaling_type is None:
-                raise NotImplementedError(
-                    "RoPE scaling type '%s' is not yet implemented. "
-                    "The following RoPE scaling types are currently supported: %s"
-                    % (rope_scaling["type"], ", ".join(_SUPPORTED_ROPE_SCALING.keys()))
-                )
-        else:
-            rotary_scaling_type = None
-            rotary_scaling_factor = 1
+        rotary_scaling_type, rotary_scaling_factor, rope_theta = self.get_rotary_params(
+            model.config, 10_000
+        )
 
         quantization_config = getattr(model.config, "quantization_config", None)
         if quantization_config:
@@ -1722,7 +1736,7 @@ class LlamaLoader(ModelLoader):
             rotary_interleave=False,
             rotary_scaling_type=rotary_scaling_type,
             rotary_scaling_factor=rotary_scaling_factor,
-            rotary_base=getattr(model.config, "rope_theta", 10000),
+            rotary_base=rope_theta,
             num_heads_kv=num_heads_kv,
             quant_type=quant_type,
             quant_group_size=quant_group_size,
@@ -1733,6 +1747,7 @@ class LlamaLoader(ModelLoader):
         self.set_linear(spec.decoder.projection, model.lm_head)
 
         # set extra RoPE parameters for Llama-3.1
+        rope_scaling = getattr(model.config, "rope_scaling", None)
         if rotary_scaling_type == attention_spec.RotaryScalingType.Llama3:
             for layer in spec.decoder.layer:
                 layer.self_attention.rotary_low_freq_factor = rope_scaling[
@@ -1859,8 +1874,12 @@ class Gemma3Loader(ModelLoader):
                     "Quantization type '%s' is not yet implemented."
                     % quantization_config.quant_method
                 )
+            quant_group_size = quantization_config.group_size
+            quant_bits = quantization_config.bits
         else:
             quant_type = common_spec.Quantization.CT2
+            quant_group_size = None
+            quant_bits = None
 
         # Create base spec using from_config
         spec = transformer_spec.TransformerDecoderModelSpec.from_config(
@@ -1881,6 +1900,9 @@ class Gemma3Loader(ModelLoader):
             head_dim=head_dim,
             sliding_window=sliding_window,  # Default to local sliding window
             pre_post_layer_norm=True,
+            quant_type=quant_type,
+            quant_group_size=quant_group_size,
+            quant_bits=quant_bits,
             qk_norm=True,
         )
 
@@ -1933,7 +1955,8 @@ class Gemma3Loader(ModelLoader):
             config.eos_token = tokenizer.eos_token
 
     def set_layer_norm(self, spec, layer_norm):
-        spec.gamma = layer_norm.weight + 1.0
+        spec.gamma = layer_norm.weight
+        spec.layer_norm_use_residual = True
 
     def set_decoder(self, spec, module, quant_type=common_spec.Quantization.CT2):
         spec.scale_embeddings = True
@@ -2022,20 +2045,9 @@ class MistralLoader(ModelLoader):
 
         sliding_window = getattr(model.config, "sliding_window", 0)
 
-        rope_scaling = getattr(model.config, "rope_scaling", None)
-        if rope_scaling:
-            rotary_scaling_type = _SUPPORTED_ROPE_SCALING.get(rope_scaling["type"])
-            rotary_scaling_factor = rope_scaling["factor"]
-
-            if rotary_scaling_type is None:
-                raise NotImplementedError(
-                    "RoPE scaling type '%s' is not yet implemented. "
-                    "The following RoPE scaling types are currently supported: %s"
-                    % (rope_scaling["type"], ", ".join(_SUPPORTED_ROPE_SCALING.keys()))
-                )
-        else:
-            rotary_scaling_type = None
-            rotary_scaling_factor = 1
+        rotary_scaling_type, rotary_scaling_factor, rope_theta = self.get_rotary_params(
+            model.config, 10_000
+        )
 
         quantization_config = getattr(model.config, "quantization_config", None)
         if quantization_config:
@@ -2068,7 +2080,7 @@ class MistralLoader(ModelLoader):
             rotary_interleave=False,
             rotary_scaling_type=rotary_scaling_type,
             rotary_scaling_factor=rotary_scaling_factor,
-            rotary_base=getattr(model.config, "rope_theta", 10000),
+            rotary_base=rope_theta,
             num_heads_kv=num_heads_kv,
             sliding_window=sliding_window,
             quant_type=quant_type,
@@ -2167,21 +2179,9 @@ class Qwen2Loader(ModelLoader):
         if num_heads_kv == num_heads:
             num_heads_kv = None
 
-        rope_scaling = getattr(model.config, "rope_scaling", None)
-        if rope_scaling:
-            rope_type = rope_scaling.get("type") or rope_scaling["rope_type"]
-            rotary_scaling_type = _SUPPORTED_ROPE_SCALING.get(rope_type)
-            rotary_scaling_factor = rope_scaling["factor"]
-
-            if rotary_scaling_type is None:
-                raise NotImplementedError(
-                    "RoPE scaling type '%s' is not yet implemented. "
-                    "The following RoPE scaling types are currently supported: %s"
-                    % (rope_scaling["type"], ", ".join(_SUPPORTED_ROPE_SCALING.keys()))
-                )
-        else:
-            rotary_scaling_type = None
-            rotary_scaling_factor = 1
+        rotary_scaling_type, rotary_scaling_factor, rope_theta = self.get_rotary_params(
+            model.config, 10_000
+        )
 
         # Check for AWQ quantization config
         quantization_config = getattr(model.config, "quantization_config", None)
@@ -2216,7 +2216,7 @@ class Qwen2Loader(ModelLoader):
             rotary_interleave=False,
             rotary_scaling_type=rotary_scaling_type,
             rotary_scaling_factor=rotary_scaling_factor,
-            rotary_base=getattr(model.config, "rope_theta", 10000),
+            rotary_base=rope_theta,
             num_heads_kv=num_heads_kv,
             quant_type=quant_type,
             quant_group_size=quant_group_size,
@@ -2323,21 +2323,9 @@ class Qwen3Loader(ModelLoader):
         if num_heads_kv == num_heads:
             num_heads_kv = None
 
-        rope_scaling = getattr(model.config, "rope_scaling", None)
-        if rope_scaling:
-            rope_type = rope_scaling.get("type") or rope_scaling["rope_type"]
-            rotary_scaling_type = _SUPPORTED_ROPE_SCALING.get(rope_type)
-            rotary_scaling_factor = rope_scaling["factor"]
-            if rotary_scaling_type is None:
-                raise NotImplementedError(
-                    "RoPE scaling type '%s' is not yet implemented. "
-                    "The following RoPE scaling types are currently supported: %s"
-                    % (rope_scaling["type"], ", ".join(_SUPPORTED_ROPE_SCALING.keys()))
-                )
-        else:
-            rotary_scaling_type = None
-            rotary_scaling_factor = 1
-
+        rotary_scaling_type, rotary_scaling_factor, rope_theta = self.get_rotary_params(
+            model.config, 1_000_000
+        )
         # Check for AWQ quantization config
         quantization_config = getattr(model.config, "quantization_config", None)
         if quantization_config:
@@ -2371,7 +2359,7 @@ class Qwen3Loader(ModelLoader):
             rotary_interleave=False,
             rotary_scaling_type=rotary_scaling_type,
             rotary_scaling_factor=rotary_scaling_factor,
-            rotary_base=getattr(model.config, "rope_theta", 10000),
+            rotary_base=rope_theta,
             num_heads_kv=num_heads_kv,
             head_dim=head_dim,
             qk_norm=True,

ctranslate2/specs/transformer_spec.py

@@ -275,7 +275,7 @@ class TransformerDecoderSpec(model_spec.LayerSpec):
             self.project_in = common_spec.LinearSpec()
             self.project_out = common_spec.LinearSpec()
 
-        if quant_type is not None:
+        if quant_type:
             self._config["quantization_type"] = quant_type
             self._config["quantization_bits"] = quant_bits
             self._config["quantization_group_size"] = quant_group_size

ctranslate2/version.py

@@ -1,3 +1,3 @@
 """Version information."""
 
-__version__ = "4.6.3"
+__version__ = "4.7.0"

{ctranslate2-4.6.3.dist-info → ctranslate2-4.7.0.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: ctranslate2
-Version: 4.6.3
+Version: 4.7.0
 Summary: Fast inference engine for Transformer models
 Home-page: https://opennmt.net
 Author: OpenNMT
@@ -100,6 +100,8 @@ generator.generate_batch(start_tokens)
 
 See the [documentation](https://opennmt.net/CTranslate2) for more information and examples.
 
+If you have an AMD ROCm GPU, we provide specific Python wheels on the [releases page](https://github.com/OpenNMT/CTranslate2/releases/).
+
 ## Benchmarks
 
 We translate the En->De test set *newstest2014* with multiple models:

{ctranslate2-4.6.3.dist-info → ctranslate2-4.7.0.dist-info}/RECORD

@@ -1,33 +1,33 @@
-ctranslate2/__init__.py,sha256=CGqShDaFxQ-u-aCtVq99T4HKuBdMB8b49l2KSxnQb8M,1735
-ctranslate2/_ext.cp310-win_amd64.pyd,sha256=QGC8xmnpa1hqFejEqMgPSvmgE7fO39ZpALFhT_h720s,699392
-ctranslate2/ctranslate2.dll,sha256=9TRGHvoyNSzXe9eEc3jKJa5-21-AeSENgp3DXvhCZ4M,58590720
+ctranslate2/__init__.py,sha256=LZy5gF-9vTRdcERSnTSP_RrCPDks9UDU7uzxw1-d0aU,1881
+ctranslate2/_ext.cp310-win_amd64.pyd,sha256=OzJjLkqT_URKa4cLjY5UzFzQ4J4m3fsONPKKz-GmB14,699392
+ctranslate2/ctranslate2.dll,sha256=umrDDC_rg_IbXg2MOi-8jNasZZdjb9b_Io8CWS5_M_U,59823104
 ctranslate2/cudnn64_9.dll,sha256=ntvN_3OwrwcOsWCyzmbln-ygSqAXNR2O7cxejhSZZ9I,266288
 ctranslate2/extensions.py,sha256=kDNt0H9KvfNCc3PrRGzfkj9Fkvna84i2O5Y-rav6UkU,21940
 ctranslate2/libiomp5md.dll,sha256=mCIzNmsK_NoeD1WgsTQJfjW3eWE_VN22nmhebNBrdV8,1614192
 ctranslate2/logging.py,sha256=P9evHdxuMx_iHvwJjEASEq-j5062H64Pl5-fJjxEuHk,1221
-ctranslate2/version.py,sha256=TboXlbA67GNmSOm1v2u_U8AKgYh5iminMMLTvi3Xho4,53
+ctranslate2/version.py,sha256=cWqiIzEeUIcvUfq82ZopTbW1pRWqZkZOW7b6pks8tz8,53
 ctranslate2/converters/__init__.py,sha256=ufYjcXf2sK4fiXAUU6tIJyWmNuLjKFf_KH3GWLXe4ls,507
 ctranslate2/converters/converter.py,sha256=Qkb8NGLLmgqMT6HZkFq61zwbxyq3NlWcaxLZ6Ap-YOQ,3601
-ctranslate2/converters/eole_ct2.py,sha256=RUcDJH_2AUt0jDs5oAqccE6tQPbO9LQ6JmVriC1DTy8,12564
+ctranslate2/converters/eole_ct2.py,sha256=sRXvPark9V-4umXpMxPuJVQekMLstyNZ7xNjyAFthvg,12623
 ctranslate2/converters/fairseq.py,sha256=2vlBk4AVCHwXxKkwPHVmcjyfo1dAV0_DJS1i6q-44NE,12822
 ctranslate2/converters/marian.py,sha256=1_7P3EbIDPOdyJbtb_Lp-LCBPBb9A8E9OhzoyFwTb64,11274
 ctranslate2/converters/openai_gpt2.py,sha256=1rXKM2ZURZHWRv4XZ135fPkVWpM4rTG-q7VR7OD6d-A,3304
 ctranslate2/converters/opennmt_py.py,sha256=zex4TbHiiJMy0tkqQg39oNjxmSZKf8dnRLH3iQ1H4z0,13227
 ctranslate2/converters/opennmt_tf.py,sha256=uBRp2wz5xriSQcA_c0S0ekY7ws6RpRX_0EKeMRdM7-s,16222
 ctranslate2/converters/opus_mt.py,sha256=5KbPaTiBhhorPzMpTugIfIJ8SgcqHfJUbJrWKBN-Djs,1254
-ctranslate2/converters/transformers.py,sha256=VRal3vKSQrAOvcNPwewjVMtgvWskz0KD5bdIrpNrZNA,142380
+ctranslate2/converters/transformers.py,sha256=41E9rMH6Qm77OIfswMVn7esp_NPZn3ZimiLTA6Be_50,141519
 ctranslate2/converters/utils.py,sha256=w7NG39lx-9dOdL57OqKVTdC__opkuP8RACg1TLlUJwM,3817
 ctranslate2/models/__init__.py,sha256=53p98uemtuvVPz8xK7_LbOhBiUJJu-c-NdmOHJgdXus,497
 ctranslate2/specs/__init__.py,sha256=9GabtSyczznYqiqUS6XvULi8pQ3_3RNRogXobGP0G80,653
 ctranslate2/specs/attention_spec.py,sha256=FnaSiQREWQw_cURgsCb9_aIpGOCxyVGTCpIOdd-08v8,3492
 ctranslate2/specs/common_spec.py,sha256=freTDhQMy5PYofBrij4_FDgrKokMYApWSPIpASZIlJc,1608
 ctranslate2/specs/model_spec.py,sha256=atCAYzDEIzyJ1TCayFGZVutHqSWa1ww-vbZ0OiIJqh8,25736
-ctranslate2/specs/transformer_spec.py,sha256=-GJ0oSjI3ns-Ei_-xXIM_P2GaZxt5Z-g03zJ0m_4ciU,34317
+ctranslate2/specs/transformer_spec.py,sha256=s6mY6MMHneraXrWua_531Xjb5MVEJZCUTemUERO11GI,34305
 ctranslate2/specs/wav2vec2_spec.py,sha256=NITsuOuf2F5bU1-aXit8-WEtWV9fH2Eq7A7857UyYho,2106
 ctranslate2/specs/wav2vec2bert_spec.py,sha256=UgtsJWC9mMgJ7bn4T_xg1uXK0rqA4-9tT2KMGVgPKnw,3529
 ctranslate2/specs/whisper_spec.py,sha256=_vm1sc5yOowOJ4iyvcxMXrgt-UcLJrZT8OtPscUXcQQ,2447
-ctranslate2-4.6.3.dist-info/METADATA,sha256=awoc6t4JSxpv51lmfAG28ZG91FhGQ8DHspyLzLqLo_Q,10839
-ctranslate2-4.6.3.dist-info/WHEEL,sha256=KUuBC6lxAbHCKilKua8R9W_TM71_-9Sg5uEP3uDWcoU,101
-ctranslate2-4.6.3.dist-info/entry_points.txt,sha256=ZHkojut_TmVRHl0bJIGm2b9wqr98GAJqxN9rlJtQshs,466
-ctranslate2-4.6.3.dist-info/top_level.txt,sha256=1hUaWzcFIuSo2BAIUHFA3Osgsu6S1giq0y6Rosv8HOQ,12
-ctranslate2-4.6.3.dist-info/RECORD,,
+ctranslate2-4.7.0.dist-info/METADATA,sha256=Vm9SM5sybdzcJHc6HBek2PgP6nbuDiEHWQFZuJjWDvc,10979
+ctranslate2-4.7.0.dist-info/WHEEL,sha256=lVtJYX4SZwMxwg8oP4kB_UdF4VQRXLlqu7hUy_2nnAE,102
+ctranslate2-4.7.0.dist-info/entry_points.txt,sha256=ZHkojut_TmVRHl0bJIGm2b9wqr98GAJqxN9rlJtQshs,466
+ctranslate2-4.7.0.dist-info/top_level.txt,sha256=1hUaWzcFIuSo2BAIUHFA3Osgsu6S1giq0y6Rosv8HOQ,12
+ctranslate2-4.7.0.dist-info/RECORD,,

{ctranslate2-4.6.3.dist-info → ctranslate2-4.7.0.dist-info}/WHEEL

@@ -1,5 +1,5 @@
 Wheel-Version: 1.0
-Generator: setuptools (80.9.0)
+Generator: setuptools (80.10.2)
 Root-Is-Purelib: false
 Tag: cp310-cp310-win_amd64