Skip to content

CLIP (Contrastive Language–Image Pre-training)

CLIP (Contrastive Language–Image Pre-training) is a neural network architecture that learns visual concepts from natural language supervision. It is trained on a large dataset of image-text pairs to create a unified vision-language model that can understand both images and text in a shared semantic space.

CLIP consists of two main components: 1. A vision encoder (Vision Transformer) that processes images into visual features 2. A text encoder (Transformer) that processes text into textual features

The model is trained using contrastive learning, where it learns to maximize the cosine similarity between the embeddings of matching image-text pairs while minimizing it for non-matching pairs. This allows CLIP to perform zero-shot classification by comparing image embeddings with text embeddings of potential labels.

CLIP was introduced in the paper "Learning Transferable Visual Models From Natural Language Supervision" and has shown remarkable zero-shot generalization capabilities across a wide range of visual classification tasks. The CLIP model combines a Vision Transformer and a Text Transformer to learn joint representations of images and text. It is trained to maximize the similarity between matching image-text pairs while minimizing similarity between non-matching pairs.

Flash / Splash Attention

CLIP supports hardware-accelerated attention via Tokamax. Pass an attention_fn at construction time:

Backend Hardware Notes
"mosaic" NVIDIA H100 (SM90) / B100 (SM100) Pallas Mosaic GPU kernel
"triton" Any NVIDIA GPU Pallas Triton kernel
"cudnn" NVIDIA GPU Via JAX-NN / cuDNN
"mosaic_tpu" TPU v5 / v7 Splash attention (block-sparse)
"xla_chunked" GPU / TPU Flash-style chunked XLA
"xla" Any Standard XLA fallback 
import jimm

# GPU: try H100 Mosaic kernel, fall back to Triton, then XLA
model = jimm.CLIP.from_pretrained("openai/clip-vit-large-patch14",
                                   attention_fn=jimm.make_tokamax_attention(["mosaic", "triton", "xla"]))

# TPU: try Splash attention, fall back to chunked XLA
model = jimm.CLIP.from_pretrained("openai/clip-vit-large-patch14",
                                   attention_fn=jimm.make_tokamax_attention(["mosaic_tpu", "xla_chunked"]))

You can also apply different kernels to each encoder via vision_attention_fn and text_attention_fn.

Note: Flash/Splash attention does not provide a speedup at typical CLIP context lengths (256 image tokens, 77 text tokens). The primary benefit is memory reduction at longer sequence lengths.

FSDP / Explicit Sharding

CLIP supports JAX explicit sharding (FSDP-style) out of the box via CLIPSharding. Large weight matrices are sharded on the contracting (in_features) dimension so that activations carry only the batch-axis sharding, avoiding duplicate-axis conflicts.

from jax.experimental import mesh_utils
from jax.sharding import AxisType, Mesh, NamedSharding, PartitionSpec as P
import jax

n_devices = jax.device_count()
mesh = Mesh(
    mesh_utils.create_device_mesh((1, n_devices)),
    ("data", "fsdp"),
    axis_types=(AxisType.Explicit, AxisType.Explicit),
)
jax.set_mesh(mesh)

model = jimm.CLIP.from_pretrained("openai/clip-vit-large-patch14")
# model params are automatically sharded across fsdp axis

CLIPSharding specs represent per-layer shapes. The Transformer stack prepends None for the scan axis to the Variable metadata after nnx.vmap, so the optimizer (e.g. nnx.Optimizer with AdamW) receives the correct stacked spec and initialises its state without any manual fixups.

To disable sharding, pass sharding=jimm.common.sharding.NoSharding().

jimm.models.clip.CLIPVisionModel

Bases: Module

Source code in src/jimm/models/clip/clip_model.py 
 17
 18
 19
 20
 21
 22
 23
 24
 25
 26
 27
 28
 29
 30
 31
 32
 33
 34
 35
 36
 37
 38
 39
 40
 41
 42
 43
 44
 45
 46
 47
 48
 49
 50
 51
 52
 53
 54
 55
 56
 57
 58
 59
 60
 61
 62
 63
 64
 65
 66
 67
 68
 69
 70
 71
 72
 73
 74
 75
 76
 77
 78
 79
 80
 81
 82
 83
 84
 85
 86
 87
 88
 89
 90
 91
 92
 93
 94
 95
 96
 97
 98
 99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
class CLIPVisionModel(nnx.Module):
    def __init__(
        self,
        image_resolution: int,
        vision_layers: int,
        vision_hidden_size: int,
        vision_patch_size: int,
        projection_dim: int,
        use_gradient_checkpointing: bool = False,
        attention_fn: Callable[..., Any] | None = None,
        rngs: rnglib.Rngs | None = None,
        dtype: DTypeLike = jnp.float32,
        param_dtype: DTypeLike = jnp.float32,
        sharding: ShardingSpec = CLIPSharding,
    ):
        """Initialize the Vision Encoder with projection.

        Args:
            image_resolution (int): The resolution of the input images.
            vision_layers (int): The number of layers in the vision transformer.
            vision_hidden_size (int): The hidden dimension size of the vision transformer.
            vision_patch_size (int): The patch size of the vision transformer.
            projection_dim (int): The output dimension after projection.
            use_gradient_checkpointing (bool, optional): Whether to use gradient checkpointing. Defaults to False.
            attention_fn (Callable[..., Any] | None, optional): Custom attention function (e.g. jimm.tokamax_attention or jimm.make_tokamax_attention("mosaic_tpu")). Defaults to None.
            rngs (rnglib.Rngs | None, optional): The random number generator state. If None, initializes to nnx.Rngs(0).
            dtype (DTypeLike, optional): The data type for computations. Defaults to jnp.float32.
            param_dtype (DTypeLike, optional): The data type for parameters. Defaults to jnp.float32.
            sharding (ShardingSpec, optional): Sharding specification for parameters. Defaults to CLIPSharding.
        """
        if rngs is None:
            rngs = nnx.Rngs(0)
        self.vision_layers = vision_layers
        self.vision_hidden_size = vision_hidden_size
        self.vision_patch_size = vision_patch_size
        self.projection_dim = projection_dim
        self.dtype = dtype

        vision_heads = vision_hidden_size // 64

        self.encoder = VisionTransformerBase(
            img_size=image_resolution,
            patch_size=vision_patch_size,
            in_channels=3,
            hidden_size=vision_hidden_size,
            num_layers=vision_layers,
            num_heads=vision_heads,
            mlp_dim=vision_hidden_size * 4,
            use_pre_norm=True,
            use_patch_bias=False,
            use_quick_gelu=True,
            use_gradient_checkpointing=use_gradient_checkpointing,
            attention_fn=attention_fn,
            pooling_type="CLS",
            layernorm_epsilon=1e-5,
            rngs=rngs,
            dtype=dtype,
            param_dtype=param_dtype,
            sharding=sharding,
        )
        self.visual_projection = nnx.Linear(
            vision_hidden_size,
            projection_dim,
            use_bias=False,
            dtype=dtype,
            param_dtype=param_dtype,
            rngs=rngs,
            kernel_init=nnx.with_partitioning(
                nnx.initializers.xavier_uniform(),
                sharding.proj_kernel,
            ),
        )

    def __call__(self, image: Float[Array, "batch height width channels"], do_projection: bool = True) -> Float[Array, "batch vision_hidden_size_or_projection_dim"]:
        """Encode images into embeddings.

        Args:
            image (Float[Array, "batch height width channels"]): Batch of input images.
            do_projection (bool): Whether to apply the visual projection layer. Defaults to True.

        Returns:
            Float[Array, "batch vision_hidden_size_or_projection_dim"]: Image embeddings.
            Shape depends on do_projection: vision_hidden_size if False, projection_dim if True.
        """
        features = self.encoder(image)
        if do_projection:
            out_shard = named_sharding_like(features, P(sharding_of(features).spec[0], sharding_of(self.visual_projection.kernel[...]).spec[-1]))
            return self.visual_projection(features, out_sharding=out_shard)
        return features

    @classmethod
    def from_pretrained(
        cls,
        model_name_or_path: str,
        use_pytorch: bool = False,
        rngs: rnglib.Rngs | None = None,
        dtype: DTypeLike = jnp.float32,
        param_dtype: DTypeLike = jnp.float32,
        sharding: ShardingSpec = CLIPSharding,
        use_gradient_checkpointing: bool = False,
        attention_fn: Callable[..., Any] | None = None,
    ) -> "CLIPVisionModel":
        """Load a pretrained vision encoder from a CLIP checkpoint.

        Args:
            model_name_or_path (str): Path to local weights or HuggingFace model ID.
            use_pytorch (bool): Whether to load from PyTorch weights. Defaults to False.
            rngs (rnglib.Rngs | None): Random number generator keys. If None, initializes to nnx.Rngs(0).
            dtype (DTypeLike): Data type for computations. Defaults to jnp.float32.
            param_dtype (DTypeLike): Data type for parameters. Defaults to jnp.float32.
            sharding (ShardingSpec): Sharding specification for parameters. Defaults to CLIPSharding.
            use_gradient_checkpointing (bool): Whether to use gradient checkpointing. Defaults to False.
            attention_fn (Callable[..., Any] | None, optional): Custom attention function (e.g. jimm.tokamax_attention or jimm.make_tokamax_attention("mosaic_tpu")). Defaults to None.

        Returns:
            CLIPVisionModel: Pretrained CLIP vision model
        """
        from .params import load_vision_from_pretrained

        return load_vision_from_pretrained(cls, model_name_or_path, use_pytorch, rngs, dtype, param_dtype, sharding, use_gradient_checkpointing, attention_fn)

    @classmethod
    def from_config(
        cls,
        config: dict[str, Any],
        *,
        rngs: rnglib.Rngs | None = None,
        dtype: DTypeLike = jnp.float32,
        param_dtype: DTypeLike = jnp.float32,
        sharding: ShardingSpec = CLIPSharding,
        use_gradient_checkpointing: bool = False,
        attention_fn: Callable[..., Any] | None = None,
    ) -> "CLIPVisionModel":
        """Create model from HuggingFace-compatible config dict.

        Args:
            config: Configuration with "vision_config" and "text_config" keys.
            rngs: Random number generator state.
            dtype: Data type for computations.
            param_dtype: Data type for parameters.
            sharding: Sharding specification for parameters.
            use_gradient_checkpointing: Enable gradient checkpointing.
            attention_fn: Custom attention function (e.g. jimm.tokamax_attention or jimm.make_tokamax_attention("mosaic_tpu")). Defaults to None.

        Returns:
            CLIPVisionModel with random weights.
        """
        if rngs is None:
            rngs = nnx.Rngs(0)
        vision_config = config["vision_config"]
        text_config = config["text_config"]

        return cls(
            image_resolution=vision_config["image_size"],
            vision_layers=vision_config["num_hidden_layers"],
            vision_hidden_size=vision_config["hidden_size"],
            vision_patch_size=vision_config["patch_size"],
            projection_dim=text_config["hidden_size"],
            use_gradient_checkpointing=use_gradient_checkpointing,
            attention_fn=attention_fn,
            rngs=rngs,
            dtype=dtype,
            param_dtype=param_dtype,
            sharding=sharding,
        )

    def save_pretrained(self, save_directory: str) -> None:
        """Save model weights and config in HuggingFace format.

        Args:
            save_directory (str): Directory path where the model will be saved.
        """
        from .params import save_vision_pretrained

        save_vision_pretrained(self, save_directory)

__call__(image, do_projection=True)

Encode images into embeddings.

Parameters:

Name Type Description Default
image Float[Array, 'batch height width channels']

Batch of input images.

 required
do_projection bool

Whether to apply the visual projection layer. Defaults to True.

 True

Returns:

Type Description
Float[Array, 'batch vision_hidden_size_or_projection_dim']

Float[Array, "batch vision_hidden_size_or_projection_dim"]: Image embeddings.
Float[Array, 'batch vision_hidden_size_or_projection_dim']

Shape depends on do_projection: vision_hidden_size if False, projection_dim if True.
Source code in src/jimm/models/clip/clip_model.py 
 90
 91
 92
 93
 94
 95
 96
 97
 98
 99
100
101
102
103
104
105
def __call__(self, image: Float[Array, "batch height width channels"], do_projection: bool = True) -> Float[Array, "batch vision_hidden_size_or_projection_dim"]:
    """Encode images into embeddings.

    Args:
        image (Float[Array, "batch height width channels"]): Batch of input images.
        do_projection (bool): Whether to apply the visual projection layer. Defaults to True.

    Returns:
        Float[Array, "batch vision_hidden_size_or_projection_dim"]: Image embeddings.
        Shape depends on do_projection: vision_hidden_size if False, projection_dim if True.
    """
    features = self.encoder(image)
    if do_projection:
        out_shard = named_sharding_like(features, P(sharding_of(features).spec[0], sharding_of(self.visual_projection.kernel[...]).spec[-1]))
        return self.visual_projection(features, out_sharding=out_shard)
    return features

__init__(image_resolution, vision_layers, vision_hidden_size, vision_patch_size, projection_dim, use_gradient_checkpointing=False, attention_fn=None, rngs=None, dtype=jnp.float32, param_dtype=jnp.float32, sharding=CLIPSharding)

Initialize the Vision Encoder with projection.

Parameters:

Name Type Description Default
image_resolution int

The resolution of the input images.

 required
vision_layers int

The number of layers in the vision transformer.

 required
vision_hidden_size int

The hidden dimension size of the vision transformer.

 required
vision_patch_size int

The patch size of the vision transformer.

 required
projection_dim int

The output dimension after projection.

 required
use_gradient_checkpointing bool

Whether to use gradient checkpointing. Defaults to False.

 False
attention_fn Callable[..., Any] | None

Custom attention function (e.g. jimm.tokamax_attention or jimm.make_tokamax_attention("mosaic_tpu")). Defaults to None.

 None
rngs Rngs | None

The random number generator state. If None, initializes to nnx.Rngs(0).

 None
dtype DTypeLike

The data type for computations. Defaults to jnp.float32.

 float32
param_dtype DTypeLike

The data type for parameters. Defaults to jnp.float32.

 float32
sharding ShardingSpec

Sharding specification for parameters. Defaults to CLIPSharding.

 CLIPSharding
Source code in src/jimm/models/clip/clip_model.py 
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
def __init__(
    self,
    image_resolution: int,
    vision_layers: int,
    vision_hidden_size: int,
    vision_patch_size: int,
    projection_dim: int,
    use_gradient_checkpointing: bool = False,
    attention_fn: Callable[..., Any] | None = None,
    rngs: rnglib.Rngs | None = None,
    dtype: DTypeLike = jnp.float32,
    param_dtype: DTypeLike = jnp.float32,
    sharding: ShardingSpec = CLIPSharding,
):
    """Initialize the Vision Encoder with projection.

    Args:
        image_resolution (int): The resolution of the input images.
        vision_layers (int): The number of layers in the vision transformer.
        vision_hidden_size (int): The hidden dimension size of the vision transformer.
        vision_patch_size (int): The patch size of the vision transformer.
        projection_dim (int): The output dimension after projection.
        use_gradient_checkpointing (bool, optional): Whether to use gradient checkpointing. Defaults to False.
        attention_fn (Callable[..., Any] | None, optional): Custom attention function (e.g. jimm.tokamax_attention or jimm.make_tokamax_attention("mosaic_tpu")). Defaults to None.
        rngs (rnglib.Rngs | None, optional): The random number generator state. If None, initializes to nnx.Rngs(0).
        dtype (DTypeLike, optional): The data type for computations. Defaults to jnp.float32.
        param_dtype (DTypeLike, optional): The data type for parameters. Defaults to jnp.float32.
        sharding (ShardingSpec, optional): Sharding specification for parameters. Defaults to CLIPSharding.
    """
    if rngs is None:
        rngs = nnx.Rngs(0)
    self.vision_layers = vision_layers
    self.vision_hidden_size = vision_hidden_size
    self.vision_patch_size = vision_patch_size
    self.projection_dim = projection_dim
    self.dtype = dtype

    vision_heads = vision_hidden_size // 64

    self.encoder = VisionTransformerBase(
        img_size=image_resolution,
        patch_size=vision_patch_size,
        in_channels=3,
        hidden_size=vision_hidden_size,
        num_layers=vision_layers,
        num_heads=vision_heads,
        mlp_dim=vision_hidden_size * 4,
        use_pre_norm=True,
        use_patch_bias=False,
        use_quick_gelu=True,
        use_gradient_checkpointing=use_gradient_checkpointing,
        attention_fn=attention_fn,
        pooling_type="CLS",
        layernorm_epsilon=1e-5,
        rngs=rngs,
        dtype=dtype,
        param_dtype=param_dtype,
        sharding=sharding,
    )
    self.visual_projection = nnx.Linear(
        vision_hidden_size,
        projection_dim,
        use_bias=False,
        dtype=dtype,
        param_dtype=param_dtype,
        rngs=rngs,
        kernel_init=nnx.with_partitioning(
            nnx.initializers.xavier_uniform(),
            sharding.proj_kernel,
        ),
    )

from_config(config, *, rngs=None, dtype=jnp.float32, param_dtype=jnp.float32, sharding=CLIPSharding, use_gradient_checkpointing=False, attention_fn=None) classmethod

Create model from HuggingFace-compatible config dict.

Parameters:

Name Type Description Default
config dict[str, Any]

Configuration with "vision_config" and "text_config" keys.

 required
rngs Rngs | None

Random number generator state.

 None
dtype DTypeLike

Data type for computations.

 float32
param_dtype DTypeLike

Data type for parameters.

 float32
sharding ShardingSpec

Sharding specification for parameters.

 CLIPSharding
use_gradient_checkpointing bool

Enable gradient checkpointing.

 False
attention_fn Callable[..., Any] | None

Custom attention function (e.g. jimm.tokamax_attention or jimm.make_tokamax_attention("mosaic_tpu")). Defaults to None.

 None

Returns:

Type Description
CLIPVisionModel

CLIPVisionModel with random weights.
Source code in src/jimm/models/clip/clip_model.py 
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
@classmethod
def from_config(
    cls,
    config: dict[str, Any],
    *,
    rngs: rnglib.Rngs | None = None,
    dtype: DTypeLike = jnp.float32,
    param_dtype: DTypeLike = jnp.float32,
    sharding: ShardingSpec = CLIPSharding,
    use_gradient_checkpointing: bool = False,
    attention_fn: Callable[..., Any] | None = None,
) -> "CLIPVisionModel":
    """Create model from HuggingFace-compatible config dict.

    Args:
        config: Configuration with "vision_config" and "text_config" keys.
        rngs: Random number generator state.
        dtype: Data type for computations.
        param_dtype: Data type for parameters.
        sharding: Sharding specification for parameters.
        use_gradient_checkpointing: Enable gradient checkpointing.
        attention_fn: Custom attention function (e.g. jimm.tokamax_attention or jimm.make_tokamax_attention("mosaic_tpu")). Defaults to None.

    Returns:
        CLIPVisionModel with random weights.
    """
    if rngs is None:
        rngs = nnx.Rngs(0)
    vision_config = config["vision_config"]
    text_config = config["text_config"]

    return cls(
        image_resolution=vision_config["image_size"],
        vision_layers=vision_config["num_hidden_layers"],
        vision_hidden_size=vision_config["hidden_size"],
        vision_patch_size=vision_config["patch_size"],
        projection_dim=text_config["hidden_size"],
        use_gradient_checkpointing=use_gradient_checkpointing,
        attention_fn=attention_fn,
        rngs=rngs,
        dtype=dtype,
        param_dtype=param_dtype,
        sharding=sharding,
    )

from_pretrained(model_name_or_path, use_pytorch=False, rngs=None, dtype=jnp.float32, param_dtype=jnp.float32, sharding=CLIPSharding, use_gradient_checkpointing=False, attention_fn=None) classmethod

Load a pretrained vision encoder from a CLIP checkpoint.

Parameters:

Name Type Description Default
model_name_or_path str

Path to local weights or HuggingFace model ID.

 required
use_pytorch bool

Whether to load from PyTorch weights. Defaults to False.

 False
rngs Rngs | None

Random number generator keys. If None, initializes to nnx.Rngs(0).

 None
dtype DTypeLike

Data type for computations. Defaults to jnp.float32.

 float32
param_dtype DTypeLike

Data type for parameters. Defaults to jnp.float32.

 float32
sharding ShardingSpec

Sharding specification for parameters. Defaults to CLIPSharding.

 CLIPSharding
use_gradient_checkpointing bool

Whether to use gradient checkpointing. Defaults to False.

 False
attention_fn Callable[..., Any] | None

Custom attention function (e.g. jimm.tokamax_attention or jimm.make_tokamax_attention("mosaic_tpu")). Defaults to None.

 None

Returns:

Name Type Description
CLIPVisionModel CLIPVisionModel

Pretrained CLIP vision model
Source code in src/jimm/models/clip/clip_model.py 
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
@classmethod
def from_pretrained(
    cls,
    model_name_or_path: str,
    use_pytorch: bool = False,
    rngs: rnglib.Rngs | None = None,
    dtype: DTypeLike = jnp.float32,
    param_dtype: DTypeLike = jnp.float32,
    sharding: ShardingSpec = CLIPSharding,
    use_gradient_checkpointing: bool = False,
    attention_fn: Callable[..., Any] | None = None,
) -> "CLIPVisionModel":
    """Load a pretrained vision encoder from a CLIP checkpoint.

    Args:
        model_name_or_path (str): Path to local weights or HuggingFace model ID.
        use_pytorch (bool): Whether to load from PyTorch weights. Defaults to False.
        rngs (rnglib.Rngs | None): Random number generator keys. If None, initializes to nnx.Rngs(0).
        dtype (DTypeLike): Data type for computations. Defaults to jnp.float32.
        param_dtype (DTypeLike): Data type for parameters. Defaults to jnp.float32.
        sharding (ShardingSpec): Sharding specification for parameters. Defaults to CLIPSharding.
        use_gradient_checkpointing (bool): Whether to use gradient checkpointing. Defaults to False.
        attention_fn (Callable[..., Any] | None, optional): Custom attention function (e.g. jimm.tokamax_attention or jimm.make_tokamax_attention("mosaic_tpu")). Defaults to None.

    Returns:
        CLIPVisionModel: Pretrained CLIP vision model
    """
    from .params import load_vision_from_pretrained

    return load_vision_from_pretrained(cls, model_name_or_path, use_pytorch, rngs, dtype, param_dtype, sharding, use_gradient_checkpointing, attention_fn)

save_pretrained(save_directory)

Save model weights and config in HuggingFace format.

Parameters:

Name Type Description Default
save_directory str

Directory path where the model will be saved.

 required
Source code in src/jimm/models/clip/clip_model.py 
183
184
185
186
187
188
189
190
191
def save_pretrained(self, save_directory: str) -> None:
    """Save model weights and config in HuggingFace format.

    Args:
        save_directory (str): Directory path where the model will be saved.
    """
    from .params import save_vision_pretrained

    save_vision_pretrained(self, save_directory)

jimm.models.clip.CLIPTextModel

Bases: Module

Source code in src/jimm/models/clip/clip_model.py 
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
class CLIPTextModel(nnx.Module):
    def __init__(
        self,
        context_length: int,
        vocab_size: int,
        text_hidden_size: int,
        num_text_heads: int,
        num_text_layers: int,
        use_gradient_checkpointing: bool = False,
        attention_fn: Callable[..., Any] | None = None,
        rngs: rnglib.Rngs | None = None,
        dtype: DTypeLike = jnp.float32,
        param_dtype: DTypeLike = jnp.float32,
        sharding: ShardingSpec = CLIPSharding,
    ):
        """Initialize CLIP text encoder.

        Args:
            context_length (int): Maximum sequence length.
            vocab_size (int): Size of vocabulary.
            text_hidden_size (int): Hidden dimension size of the text transformer.
            num_text_heads (int): Number of attention heads in the text transformer.
            num_text_layers (int): Number of transformer layers in the text transformer.
            use_gradient_checkpointing (bool): Enable gradient checkpointing.
            attention_fn (Callable[..., Any] | None, optional): Custom attention function (e.g. jimm.tokamax_attention or jimm.make_tokamax_attention("mosaic_tpu")). Defaults to None.
            rngs (rnglib.Rngs | None): RNG state. If None, initializes to nnx.Rngs(0).
            dtype (DTypeLike): Computation dtype.
            param_dtype (DTypeLike): Parameter dtype.
            sharding (ShardingSpec): Sharding specification for parameters.
        """
        if rngs is None:
            rngs = nnx.Rngs(0)
        self.context_length = context_length
        self.vocab_size = vocab_size
        self.text_hidden_size = text_hidden_size
        self.num_text_heads = num_text_heads
        self.num_text_layers = num_text_layers
        self.dtype = dtype

        self.token_embedding = nnx.Embed(
            num_embeddings=vocab_size,
            features=text_hidden_size,
            dtype=dtype,
            param_dtype=param_dtype,
            rngs=rngs,
            embedding_init=nnx.with_partitioning(
                nnx.initializers.xavier_uniform(),
                sharding.embed,
            ),
        )
        self.positional_embedding = nnx.Param(
            nnx.with_partitioning(
                nnx.initializers.truncated_normal(stddev=0.02),
                sharding.text_pos_embed,
            )(rngs.params(), (context_length, text_hidden_size))
        )

        attn_mask = jnp.tril(jnp.ones((context_length, context_length), dtype=dtype))
        self.transformer = Transformer(
            hidden_size=text_hidden_size,
            mlp_dim=text_hidden_size * 4,
            num_layers=num_text_layers,
            num_heads=num_text_heads,
            dropout_rate=0.0,
            attn_mask=attn_mask,
            use_quick_gelu=True,
            use_gradient_checkpointing=use_gradient_checkpointing,
            attention_fn=attention_fn,
            rngs=rngs,
            dtype=dtype,
            param_dtype=param_dtype,
            sharding=sharding,
        )

        self.ln_final = nnx.LayerNorm(
            text_hidden_size,
            epsilon=1e-5,
            dtype=dtype,
            param_dtype=param_dtype,
            rngs=rngs,
            scale_init=nnx.with_partitioning(
                nnx.initializers.ones_init(),
                sharding.layernorm,
            ),
            bias_init=nnx.with_partitioning(
                nnx.initializers.zeros_init(),
                sharding.layernorm,
            ),
        )

        self.text_projection = nnx.Linear(
            text_hidden_size,
            text_hidden_size,
            use_bias=False,
            dtype=dtype,
            param_dtype=param_dtype,
            rngs=rngs,
            kernel_init=nnx.with_partitioning(
                nnx.initializers.xavier_uniform(),
                sharding.proj_kernel,
            ),
        )

    def __call__(self, text: Int[Array, "batch context_length"], do_projection: bool = True) -> Float[Array, "batch text_hidden_size"]:
        """Encode text tokens into embeddings.

        Args:
            text (Int[Array, "batch context_length"]): Token sequences.
            do_projection (bool): Apply text projection layer.

        Returns:
            Float[Array, "batch text_hidden_size"]: Text embeddings.
        """
        seq_len = text.shape[1]
        text_sharding = sharding_of(text)
        embed_sharding = named_sharding_like(text, P(*text_sharding.spec, None))
        x = self.token_embedding.embedding[...].at[text].get(out_sharding=embed_sharding)
        pos_embed = jnp.broadcast_to(self.positional_embedding[...][:seq_len], x.shape)
        x = x + reshard_like(pos_embed, x)
        x = self.transformer(x)
        x = self.ln_final(x)

        eot_mask = jax.nn.one_hot(jnp.argmax(text, axis=-1), x.shape[1])
        x_spec = sharding_of(x).spec
        pooled_sharding = named_sharding_like(x, P(x_spec[0], x_spec[2]))
        x = jnp.einsum("bsh,bs->bh", x, eot_mask, out_sharding=pooled_sharding)

        if do_projection:
            out_shard = named_sharding_like(x, P(sharding_of(x).spec[0], sharding_of(self.text_projection.kernel[...]).spec[-1]))
            x = self.text_projection(x, out_sharding=out_shard)
        return x

    @classmethod
    def from_pretrained(
        cls,
        model_name_or_path: str,
        use_pytorch: bool = False,
        rngs: rnglib.Rngs | None = None,
        dtype: DTypeLike = jnp.float32,
        param_dtype: DTypeLike = jnp.float32,
        sharding: ShardingSpec = CLIPSharding,
        use_gradient_checkpointing: bool = False,
        attention_fn: Callable[..., Any] | None = None,
    ) -> "CLIPTextModel":
        """Load pretrained text encoder from CLIP checkpoint.

        Args:
            model_name_or_path (str): Local path or HuggingFace model ID.
            use_pytorch (bool): Load from PyTorch weights.
            rngs (rnglib.Rngs | None): RNG state. If None, initializes to nnx.Rngs(0).
            dtype (DTypeLike): Computation dtype.
            param_dtype (DTypeLike): Parameter dtype.
            sharding (ShardingSpec): Sharding specification for parameters.
            use_gradient_checkpointing (bool): Enable gradient checkpointing.
            attention_fn (Callable[..., Any] | None, optional): Custom attention function (e.g. jimm.tokamax_attention or jimm.make_tokamax_attention("mosaic_tpu")). Defaults to None.

        Returns:
            CLIPTextModel: Pretrained text model.
        """
        from .params import load_text_from_pretrained

        return load_text_from_pretrained(cls, model_name_or_path, use_pytorch, rngs, dtype, param_dtype, sharding, use_gradient_checkpointing, attention_fn)

    @classmethod
    def from_config(
        cls,
        config: dict[str, Any],
        *,
        rngs: rnglib.Rngs | None = None,
        dtype: DTypeLike = jnp.float32,
        param_dtype: DTypeLike = jnp.float32,
        sharding: ShardingSpec = CLIPSharding,
        use_gradient_checkpointing: bool = False,
        attention_fn: Callable[..., Any] | None = None,
    ) -> "CLIPTextModel":
        """Create model from HuggingFace-compatible config dict.

        Args:
            config: Configuration with "text_config" key.
            rngs: Random number generator state.
            dtype: Data type for computations.
            param_dtype: Data type for parameters.
            sharding: Sharding specification for parameters.
            use_gradient_checkpointing: Enable gradient checkpointing.
            attention_fn: Custom attention function (e.g. jimm.tokamax_attention or jimm.make_tokamax_attention("mosaic_tpu")). Defaults to None.

        Returns:
            CLIPTextModel with random weights.
        """
        if rngs is None:
            rngs = nnx.Rngs(0)
        text_config = config["text_config"]

        return cls(
            context_length=text_config["max_position_embeddings"],
            vocab_size=text_config["vocab_size"],
            text_hidden_size=text_config["hidden_size"],
            num_text_heads=text_config["num_attention_heads"],
            num_text_layers=text_config["num_hidden_layers"],
            use_gradient_checkpointing=use_gradient_checkpointing,
            attention_fn=attention_fn,
            rngs=rngs,
            dtype=dtype,
            param_dtype=param_dtype,
            sharding=sharding,
        )

    def save_pretrained(self, save_directory: str) -> None:
        """Save model weights and config in HuggingFace format.

        Args:
            save_directory (str): Directory path where the model will be saved.
        """
        from .params import save_text_pretrained

        save_text_pretrained(self, save_directory)

__call__(text, do_projection=True)

Encode text tokens into embeddings.

Parameters:

Name Type Description Default
text Int[Array, 'batch context_length']

Token sequences.

 required
do_projection bool

Apply text projection layer.

 True

Returns:

Type Description
Float[Array, 'batch text_hidden_size']

Float[Array, "batch text_hidden_size"]: Text embeddings.
Source code in src/jimm/models/clip/clip_model.py 
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
def __call__(self, text: Int[Array, "batch context_length"], do_projection: bool = True) -> Float[Array, "batch text_hidden_size"]:
    """Encode text tokens into embeddings.

    Args:
        text (Int[Array, "batch context_length"]): Token sequences.
        do_projection (bool): Apply text projection layer.

    Returns:
        Float[Array, "batch text_hidden_size"]: Text embeddings.
    """
    seq_len = text.shape[1]
    text_sharding = sharding_of(text)
    embed_sharding = named_sharding_like(text, P(*text_sharding.spec, None))
    x = self.token_embedding.embedding[...].at[text].get(out_sharding=embed_sharding)
    pos_embed = jnp.broadcast_to(self.positional_embedding[...][:seq_len], x.shape)
    x = x + reshard_like(pos_embed, x)
    x = self.transformer(x)
    x = self.ln_final(x)

    eot_mask = jax.nn.one_hot(jnp.argmax(text, axis=-1), x.shape[1])
    x_spec = sharding_of(x).spec
    pooled_sharding = named_sharding_like(x, P(x_spec[0], x_spec[2]))
    x = jnp.einsum("bsh,bs->bh", x, eot_mask, out_sharding=pooled_sharding)

    if do_projection:
        out_shard = named_sharding_like(x, P(sharding_of(x).spec[0], sharding_of(self.text_projection.kernel[...]).spec[-1]))
        x = self.text_projection(x, out_sharding=out_shard)
    return x

__init__(context_length, vocab_size, text_hidden_size, num_text_heads, num_text_layers, use_gradient_checkpointing=False, attention_fn=None, rngs=None, dtype=jnp.float32, param_dtype=jnp.float32, sharding=CLIPSharding)

Initialize CLIP text encoder.

Parameters:

Name Type Description Default
context_length int

Maximum sequence length.

 required
vocab_size int

Size of vocabulary.

 required
text_hidden_size int

Hidden dimension size of the text transformer.

 required
num_text_heads int

Number of attention heads in the text transformer.

 required
num_text_layers int

Number of transformer layers in the text transformer.

 required
use_gradient_checkpointing bool

Enable gradient checkpointing.

 False
attention_fn Callable[..., Any] | None

Custom attention function (e.g. jimm.tokamax_attention or jimm.make_tokamax_attention("mosaic_tpu")). Defaults to None.

 None
rngs Rngs | None

RNG state. If None, initializes to nnx.Rngs(0).

 None
dtype DTypeLike

Computation dtype.

 float32
param_dtype DTypeLike

Parameter dtype.

 float32
sharding ShardingSpec

Sharding specification for parameters.

 CLIPSharding
Source code in src/jimm/models/clip/clip_model.py 
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
def __init__(
    self,
    context_length: int,
    vocab_size: int,
    text_hidden_size: int,
    num_text_heads: int,
    num_text_layers: int,
    use_gradient_checkpointing: bool = False,
    attention_fn: Callable[..., Any] | None = None,
    rngs: rnglib.Rngs | None = None,
    dtype: DTypeLike = jnp.float32,
    param_dtype: DTypeLike = jnp.float32,
    sharding: ShardingSpec = CLIPSharding,
):
    """Initialize CLIP text encoder.

    Args:
        context_length (int): Maximum sequence length.
        vocab_size (int): Size of vocabulary.
        text_hidden_size (int): Hidden dimension size of the text transformer.
        num_text_heads (int): Number of attention heads in the text transformer.
        num_text_layers (int): Number of transformer layers in the text transformer.
        use_gradient_checkpointing (bool): Enable gradient checkpointing.
        attention_fn (Callable[..., Any] | None, optional): Custom attention function (e.g. jimm.tokamax_attention or jimm.make_tokamax_attention("mosaic_tpu")). Defaults to None.
        rngs (rnglib.Rngs | None): RNG state. If None, initializes to nnx.Rngs(0).
        dtype (DTypeLike): Computation dtype.
        param_dtype (DTypeLike): Parameter dtype.
        sharding (ShardingSpec): Sharding specification for parameters.
    """
    if rngs is None:
        rngs = nnx.Rngs(0)
    self.context_length = context_length
    self.vocab_size = vocab_size
    self.text_hidden_size = text_hidden_size
    self.num_text_heads = num_text_heads
    self.num_text_layers = num_text_layers
    self.dtype = dtype

    self.token_embedding = nnx.Embed(
        num_embeddings=vocab_size,
        features=text_hidden_size,
        dtype=dtype,
        param_dtype=param_dtype,
        rngs=rngs,
        embedding_init=nnx.with_partitioning(
            nnx.initializers.xavier_uniform(),
            sharding.embed,
        ),
    )
    self.positional_embedding = nnx.Param(
        nnx.with_partitioning(
            nnx.initializers.truncated_normal(stddev=0.02),
            sharding.text_pos_embed,
        )(rngs.params(), (context_length, text_hidden_size))
    )

    attn_mask = jnp.tril(jnp.ones((context_length, context_length), dtype=dtype))
    self.transformer = Transformer(
        hidden_size=text_hidden_size,
        mlp_dim=text_hidden_size * 4,
        num_layers=num_text_layers,
        num_heads=num_text_heads,
        dropout_rate=0.0,
        attn_mask=attn_mask,
        use_quick_gelu=True,
        use_gradient_checkpointing=use_gradient_checkpointing,
        attention_fn=attention_fn,
        rngs=rngs,
        dtype=dtype,
        param_dtype=param_dtype,
        sharding=sharding,
    )

    self.ln_final = nnx.LayerNorm(
        text_hidden_size,
        epsilon=1e-5,
        dtype=dtype,
        param_dtype=param_dtype,
        rngs=rngs,
        scale_init=nnx.with_partitioning(
            nnx.initializers.ones_init(),
            sharding.layernorm,
        ),
        bias_init=nnx.with_partitioning(
            nnx.initializers.zeros_init(),
            sharding.layernorm,
        ),
    )

    self.text_projection = nnx.Linear(
        text_hidden_size,
        text_hidden_size,
        use_bias=False,
        dtype=dtype,
        param_dtype=param_dtype,
        rngs=rngs,
        kernel_init=nnx.with_partitioning(
            nnx.initializers.xavier_uniform(),
            sharding.proj_kernel,
        ),
    )

from_config(config, *, rngs=None, dtype=jnp.float32, param_dtype=jnp.float32, sharding=CLIPSharding, use_gradient_checkpointing=False, attention_fn=None) classmethod

Create model from HuggingFace-compatible config dict.

Parameters:

Name Type Description Default
config dict[str, Any]

Configuration with "text_config" key.

 required
rngs Rngs | None

Random number generator state.

 None
dtype DTypeLike

Data type for computations.

 float32
param_dtype DTypeLike

Data type for parameters.

 float32
sharding ShardingSpec

Sharding specification for parameters.

 CLIPSharding
use_gradient_checkpointing bool

Enable gradient checkpointing.

 False
attention_fn Callable[..., Any] | None

Custom attention function (e.g. jimm.tokamax_attention or jimm.make_tokamax_attention("mosaic_tpu")). Defaults to None.

 None

Returns:

Type Description
CLIPTextModel

CLIPTextModel with random weights.
Source code in src/jimm/models/clip/clip_model.py 
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
@classmethod
def from_config(
    cls,
    config: dict[str, Any],
    *,
    rngs: rnglib.Rngs | None = None,
    dtype: DTypeLike = jnp.float32,
    param_dtype: DTypeLike = jnp.float32,
    sharding: ShardingSpec = CLIPSharding,
    use_gradient_checkpointing: bool = False,
    attention_fn: Callable[..., Any] | None = None,
) -> "CLIPTextModel":
    """Create model from HuggingFace-compatible config dict.

    Args:
        config: Configuration with "text_config" key.
        rngs: Random number generator state.
        dtype: Data type for computations.
        param_dtype: Data type for parameters.
        sharding: Sharding specification for parameters.
        use_gradient_checkpointing: Enable gradient checkpointing.
        attention_fn: Custom attention function (e.g. jimm.tokamax_attention or jimm.make_tokamax_attention("mosaic_tpu")). Defaults to None.

    Returns:
        CLIPTextModel with random weights.
    """
    if rngs is None:
        rngs = nnx.Rngs(0)
    text_config = config["text_config"]

    return cls(
        context_length=text_config["max_position_embeddings"],
        vocab_size=text_config["vocab_size"],
        text_hidden_size=text_config["hidden_size"],
        num_text_heads=text_config["num_attention_heads"],
        num_text_layers=text_config["num_hidden_layers"],
        use_gradient_checkpointing=use_gradient_checkpointing,
        attention_fn=attention_fn,
        rngs=rngs,
        dtype=dtype,
        param_dtype=param_dtype,
        sharding=sharding,
    )

from_pretrained(model_name_or_path, use_pytorch=False, rngs=None, dtype=jnp.float32, param_dtype=jnp.float32, sharding=CLIPSharding, use_gradient_checkpointing=False, attention_fn=None) classmethod

Load pretrained text encoder from CLIP checkpoint.

Parameters:

Name Type Description Default
model_name_or_path str

Local path or HuggingFace model ID.

 required
use_pytorch bool

Load from PyTorch weights.

 False
rngs Rngs | None

RNG state. If None, initializes to nnx.Rngs(0).

 None
dtype DTypeLike

Computation dtype.

 float32
param_dtype DTypeLike

Parameter dtype.

 float32
sharding ShardingSpec

Sharding specification for parameters.

 CLIPSharding
use_gradient_checkpointing bool

Enable gradient checkpointing.

 False
attention_fn Callable[..., Any] | None

Custom attention function (e.g. jimm.tokamax_attention or jimm.make_tokamax_attention("mosaic_tpu")). Defaults to None.

 None

Returns:

Name Type Description
CLIPTextModel CLIPTextModel

Pretrained text model.
Source code in src/jimm/models/clip/clip_model.py 
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
@classmethod
def from_pretrained(
    cls,
    model_name_or_path: str,
    use_pytorch: bool = False,
    rngs: rnglib.Rngs | None = None,
    dtype: DTypeLike = jnp.float32,
    param_dtype: DTypeLike = jnp.float32,
    sharding: ShardingSpec = CLIPSharding,
    use_gradient_checkpointing: bool = False,
    attention_fn: Callable[..., Any] | None = None,
) -> "CLIPTextModel":
    """Load pretrained text encoder from CLIP checkpoint.

    Args:
        model_name_or_path (str): Local path or HuggingFace model ID.
        use_pytorch (bool): Load from PyTorch weights.
        rngs (rnglib.Rngs | None): RNG state. If None, initializes to nnx.Rngs(0).
        dtype (DTypeLike): Computation dtype.
        param_dtype (DTypeLike): Parameter dtype.
        sharding (ShardingSpec): Sharding specification for parameters.
        use_gradient_checkpointing (bool): Enable gradient checkpointing.
        attention_fn (Callable[..., Any] | None, optional): Custom attention function (e.g. jimm.tokamax_attention or jimm.make_tokamax_attention("mosaic_tpu")). Defaults to None.

    Returns:
        CLIPTextModel: Pretrained text model.
    """
    from .params import load_text_from_pretrained

    return load_text_from_pretrained(cls, model_name_or_path, use_pytorch, rngs, dtype, param_dtype, sharding, use_gradient_checkpointing, attention_fn)

save_pretrained(save_directory)

Save model weights and config in HuggingFace format.

Parameters:

Name Type Description Default
save_directory str

Directory path where the model will be saved.

 required
Source code in src/jimm/models/clip/clip_model.py 
401
402
403
404
405
406
407
408
409
def save_pretrained(self, save_directory: str) -> None:
    """Save model weights and config in HuggingFace format.

    Args:
        save_directory (str): Directory path where the model will be saved.
    """
    from .params import save_text_pretrained

    save_text_pretrained(self, save_directory)

jimm.models.clip.CLIP

Bases: Module

Source code in src/jimm/models/clip/clip_model.py 
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
class CLIP(nnx.Module):
    def __init__(
        self,
        image_resolution: int,
        vision_layers: int,
        vision_hidden_size: int,
        vision_patch_size: int,
        context_length: int,
        vocab_size: int,
        text_hidden_size: int,
        num_text_heads: int,
        num_text_layers: int,
        use_gradient_checkpointing: bool = False,
        attention_fn: Callable[..., Any] | None = None,
        vision_attention_fn: Callable[..., Any] | None = None,
        text_attention_fn: Callable[..., Any] | None = None,
        rngs: rnglib.Rngs | None = None,
        dtype: DTypeLike = jnp.float32,
        param_dtype: DTypeLike = jnp.float32,
        sharding: ShardingSpec = CLIPSharding,
    ):
        """Initialize the CLIP model.

        Args:
            image_resolution (int): The resolution of the input images.
            vision_layers (int): The number of layers in the vision transformer.
            vision_hidden_size (int): The hidden dimension size of the vision transformer.
            vision_patch_size (int): The patch size of the vision transformer.
            context_length (int): The maximum sequence length for text.
            vocab_size (int): The size of the vocabulary.
            text_hidden_size (int): The hidden dimension size of the text transformer.
            num_text_heads (int): The number of attention heads in the text transformer.
            num_text_layers (int): The number of layers in the text transformer.
            use_gradient_checkpointing (bool, optional): Whether to use gradient checkpointing. Defaults to False.
            attention_fn (Callable[..., Any] | None, optional): Custom attention function applied to both encoders. Defaults to None.
            vision_attention_fn (Callable[..., Any] | None, optional): Override attention_fn for the vision encoder only. Defaults to None.
            text_attention_fn (Callable[..., Any] | None, optional): Override attention_fn for the text encoder only. Defaults to None.
            rngs (rnglib.Rngs | None, optional): The random number generator state. If None, initializes to nnx.Rngs(0).
            dtype (DTypeLike, optional): The data type for computations. Defaults to jnp.float32.
            param_dtype (DTypeLike, optional): The data type for parameters. Defaults to jnp.float32.
            sharding (ShardingSpec): Sharding specification for parameters. Defaults to CLIPSharding.
        """
        if rngs is None:
            rngs = nnx.Rngs(0)
        self.vision_layers = vision_layers
        self.vision_hidden_size = vision_hidden_size
        self.vision_patch_size = vision_patch_size
        self.context_length = context_length
        self.vocab_size = vocab_size
        self.text_hidden_size = text_hidden_size
        self.num_text_heads = num_text_heads
        self.num_text_layers = num_text_layers
        self.dtype = dtype
        self._original_config = None

        self.vision_model = CLIPVisionModel(
            image_resolution=image_resolution,
            vision_layers=vision_layers,
            vision_hidden_size=vision_hidden_size,
            vision_patch_size=vision_patch_size,
            projection_dim=text_hidden_size,
            use_gradient_checkpointing=use_gradient_checkpointing,
            attention_fn=vision_attention_fn or attention_fn,
            rngs=rngs,
            dtype=dtype,
            param_dtype=param_dtype,
            sharding=sharding,
        )

        self.text_model = CLIPTextModel(
            context_length=context_length,
            vocab_size=vocab_size,
            text_hidden_size=text_hidden_size,
            num_text_heads=num_text_heads,
            num_text_layers=num_text_layers,
            use_gradient_checkpointing=use_gradient_checkpointing,
            attention_fn=text_attention_fn or attention_fn,
            rngs=rngs,
            dtype=dtype,
            param_dtype=param_dtype,
            sharding=sharding,
        )
        self.logit_scale = nnx.Param(nnx.with_partitioning(nnx.initializers.ones_init(), ())(rngs.params(), ()))

    def encode_image(self, image: Float[Array, "batch height width channels"], do_projection: bool = True) -> Float[Array, "batch text_hidden_size"]:
        """Encode images into embeddings.

        Args:
            image (Float[Array, "batch height width channels"]): Batch of input images.
            do_projection (bool): Whether the image encoder should do the visual projection layer. Defaults to true.

        Returns:
            Float[Array, "batch text_hidden_size"]: Image embeddings.
        """
        return self.vision_model(image, do_projection)

    def encode_text(self, text: Int[Array, "batch context_length"]) -> Float[Array, "batch text_hidden_size"]:
        """Encode text tokens into embeddings.

        Args:
            text (Int[Array, "batch context_length"]): Batch of token sequences.

        Returns:
            Float[Array, "batch text_hidden_size"]: Text embeddings.
        """
        return self.text_model(text, do_projection=True)

    def __call__(self, image: Float[Array, "batch height width channels"], text: Int[Array, "batch context_length"]) -> Float[Array, "batch batch"]:
        """Calculate similarity between image and text embeddings.

        Args:
            image (Float[Array, "batch height width channels"]): Batch of input images.
            text (Int[Array, "batch context_length"]): Batch of token sequences.

        Returns:
            Float[Array, "batch batch"]: Similarity scores between all pairs of images and texts.
        """
        image_features: Float[Array, "batch text_hidden_size"] = self.encode_image(image, do_projection=True)
        text_features: Float[Array, "batch text_hidden_size"] = self.encode_text(text)

        image_features: Float[Array, "batch text_hidden_size"] = image_features / jnp.linalg.norm(image_features, axis=-1, keepdims=True)
        text_features: Float[Array, "batch text_hidden_size"] = text_features / jnp.linalg.norm(text_features, axis=-1, keepdims=True)

        logit_scale: Float[Array, ""] = jnp.exp(self.logit_scale[...])
        image_spec = sharding_of(image_features).spec
        logits_sharding = named_sharding_like(image_features, P(image_spec[0], None))
        logits: Float[Array, "batch batch"] = logit_scale * jnp.matmul(image_features, text_features.T, out_sharding=logits_sharding)
        return logits

    @classmethod
    def from_pretrained(
        cls,
        model_name_or_path: str,
        use_pytorch: bool = False,
        rngs: rnglib.Rngs | None = None,
        dtype: DTypeLike = jnp.float32,
        param_dtype: DTypeLike = jnp.float32,
        sharding: ShardingSpec = CLIPSharding,
        use_gradient_checkpointing: bool = False,
        attention_fn: Callable[..., Any] | None = None,
    ) -> "CLIP":
        """Load a pretrained CLIP model from a local path or HuggingFace Hub.

        Args:
            model_name_or_path (str): Path to local weights or HuggingFace model ID.
            use_pytorch (bool): Whether to load from PyTorch weights. Defaults to False.
            rngs (rnglib.Rngs | None): Random number generator keys. If None, initializes to nnx.Rngs(0).
            dtype (DTypeLike): Data type for computations. Defaults to jnp.float32.
            param_dtype (DTypeLike): Data type for parameters. Defaults to jnp.float32.
            sharding (ShardingSpec): Sharding specification for parameters. Defaults to CLIPSharding.
            use_gradient_checkpointing (bool): Whether to use gradient checkpointing. Defaults to False.
            attention_fn (Callable[..., Any] | None, optional): Custom attention function (e.g. jimm.tokamax_attention or jimm.make_tokamax_attention("mosaic_tpu")). Defaults to None.

        Returns:
            CLIP: Pretrained CLIP model
        """
        from .params import load_from_pretrained

        return load_from_pretrained(cls, model_name_or_path, use_pytorch, rngs, dtype, param_dtype, sharding, use_gradient_checkpointing, attention_fn)

    @classmethod
    def from_config(
        cls,
        config: dict[str, Any],
        *,
        rngs: rnglib.Rngs | None = None,
        dtype: DTypeLike = jnp.float32,
        param_dtype: DTypeLike = jnp.float32,
        sharding: ShardingSpec = CLIPSharding,
        use_gradient_checkpointing: bool = False,
        attention_fn: Callable[..., Any] | None = None,
        vision_attention_fn: Callable[..., Any] | None = None,
        text_attention_fn: Callable[..., Any] | None = None,
    ) -> "CLIP":
        """Create model from HuggingFace-compatible config dict.

        Args:
            config: Configuration with "text_config" and "vision_config" keys.
            rngs: Random number generator state.
            dtype: Data type for computations.
            param_dtype: Data type for parameters.
            sharding: Sharding specification for parameters.
            use_gradient_checkpointing: Enable gradient checkpointing.
            attention_fn: Custom attention function applied to both encoders. Defaults to None.
            vision_attention_fn: Override attention_fn for the vision encoder only. Defaults to None.
            text_attention_fn: Override attention_fn for the text encoder only. Defaults to None.

        Returns:
            CLIP model with random weights.
        """
        if rngs is None:
            rngs = nnx.Rngs(0)
        text_config = config["text_config"]
        vision_config = config["vision_config"]

        return cls(
            image_resolution=vision_config["image_size"],
            vision_layers=vision_config["num_hidden_layers"],
            vision_hidden_size=vision_config["hidden_size"],
            vision_patch_size=vision_config["patch_size"],
            context_length=text_config["max_position_embeddings"],
            vocab_size=text_config["vocab_size"],
            text_hidden_size=text_config["hidden_size"],
            num_text_heads=text_config["num_attention_heads"],
            num_text_layers=text_config["num_hidden_layers"],
            use_gradient_checkpointing=use_gradient_checkpointing,
            attention_fn=attention_fn,
            vision_attention_fn=vision_attention_fn,
            text_attention_fn=text_attention_fn,
            rngs=rngs,
            dtype=dtype,
            param_dtype=param_dtype,
            sharding=sharding,
        )

    def save_pretrained(self, save_directory: str) -> None:
        """Save the model weights and config in HuggingFace format.

        Args:
            save_directory (str): Directory path where the model will be saved.
        """
        from .params import save_pretrained

        save_pretrained(self, save_directory)

__call__(image, text)

Calculate similarity between image and text embeddings.

Parameters:

Name Type Description Default
image Float[Array, 'batch height width channels']

Batch of input images.

 required
text Int[Array, 'batch context_length']

Batch of token sequences.

 required

Returns:

Type Description
Float[Array, 'batch batch']

Float[Array, "batch batch"]: Similarity scores between all pairs of images and texts.
Source code in src/jimm/models/clip/clip_model.py 
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
def __call__(self, image: Float[Array, "batch height width channels"], text: Int[Array, "batch context_length"]) -> Float[Array, "batch batch"]:
    """Calculate similarity between image and text embeddings.

    Args:
        image (Float[Array, "batch height width channels"]): Batch of input images.
        text (Int[Array, "batch context_length"]): Batch of token sequences.

    Returns:
        Float[Array, "batch batch"]: Similarity scores between all pairs of images and texts.
    """
    image_features: Float[Array, "batch text_hidden_size"] = self.encode_image(image, do_projection=True)
    text_features: Float[Array, "batch text_hidden_size"] = self.encode_text(text)

    image_features: Float[Array, "batch text_hidden_size"] = image_features / jnp.linalg.norm(image_features, axis=-1, keepdims=True)
    text_features: Float[Array, "batch text_hidden_size"] = text_features / jnp.linalg.norm(text_features, axis=-1, keepdims=True)

    logit_scale: Float[Array, ""] = jnp.exp(self.logit_scale[...])
    image_spec = sharding_of(image_features).spec
    logits_sharding = named_sharding_like(image_features, P(image_spec[0], None))
    logits: Float[Array, "batch batch"] = logit_scale * jnp.matmul(image_features, text_features.T, out_sharding=logits_sharding)
    return logits

__init__(image_resolution, vision_layers, vision_hidden_size, vision_patch_size, context_length, vocab_size, text_hidden_size, num_text_heads, num_text_layers, use_gradient_checkpointing=False, attention_fn=None, vision_attention_fn=None, text_attention_fn=None, rngs=None, dtype=jnp.float32, param_dtype=jnp.float32, sharding=CLIPSharding)

Initialize the CLIP model.

Parameters:

Name Type Description Default
image_resolution int

The resolution of the input images.

 required
vision_layers int

The number of layers in the vision transformer.

 required
vision_hidden_size int

The hidden dimension size of the vision transformer.

 required
vision_patch_size int

The patch size of the vision transformer.

 required
context_length int

The maximum sequence length for text.

 required
vocab_size int

The size of the vocabulary.

 required
text_hidden_size int

The hidden dimension size of the text transformer.

 required
num_text_heads int

The number of attention heads in the text transformer.

 required
num_text_layers int

The number of layers in the text transformer.

 required
use_gradient_checkpointing bool

Whether to use gradient checkpointing. Defaults to False.

 False
attention_fn Callable[..., Any] | None

Custom attention function applied to both encoders. Defaults to None.

 None
vision_attention_fn Callable[..., Any] | None

Override attention_fn for the vision encoder only. Defaults to None.

 None
text_attention_fn Callable[..., Any] | None

Override attention_fn for the text encoder only. Defaults to None.

 None
rngs Rngs | None

The random number generator state. If None, initializes to nnx.Rngs(0).

 None
dtype DTypeLike

The data type for computations. Defaults to jnp.float32.

 float32
param_dtype DTypeLike

The data type for parameters. Defaults to jnp.float32.

 float32
sharding ShardingSpec

Sharding specification for parameters. Defaults to CLIPSharding.

 CLIPSharding
Source code in src/jimm/models/clip/clip_model.py 
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
def __init__(
    self,
    image_resolution: int,
    vision_layers: int,
    vision_hidden_size: int,
    vision_patch_size: int,
    context_length: int,
    vocab_size: int,
    text_hidden_size: int,
    num_text_heads: int,
    num_text_layers: int,
    use_gradient_checkpointing: bool = False,
    attention_fn: Callable[..., Any] | None = None,
    vision_attention_fn: Callable[..., Any] | None = None,
    text_attention_fn: Callable[..., Any] | None = None,
    rngs: rnglib.Rngs | None = None,
    dtype: DTypeLike = jnp.float32,
    param_dtype: DTypeLike = jnp.float32,
    sharding: ShardingSpec = CLIPSharding,
):
    """Initialize the CLIP model.

    Args:
        image_resolution (int): The resolution of the input images.
        vision_layers (int): The number of layers in the vision transformer.
        vision_hidden_size (int): The hidden dimension size of the vision transformer.
        vision_patch_size (int): The patch size of the vision transformer.
        context_length (int): The maximum sequence length for text.
        vocab_size (int): The size of the vocabulary.
        text_hidden_size (int): The hidden dimension size of the text transformer.
        num_text_heads (int): The number of attention heads in the text transformer.
        num_text_layers (int): The number of layers in the text transformer.
        use_gradient_checkpointing (bool, optional): Whether to use gradient checkpointing. Defaults to False.
        attention_fn (Callable[..., Any] | None, optional): Custom attention function applied to both encoders. Defaults to None.
        vision_attention_fn (Callable[..., Any] | None, optional): Override attention_fn for the vision encoder only. Defaults to None.
        text_attention_fn (Callable[..., Any] | None, optional): Override attention_fn for the text encoder only. Defaults to None.
        rngs (rnglib.Rngs | None, optional): The random number generator state. If None, initializes to nnx.Rngs(0).
        dtype (DTypeLike, optional): The data type for computations. Defaults to jnp.float32.
        param_dtype (DTypeLike, optional): The data type for parameters. Defaults to jnp.float32.
        sharding (ShardingSpec): Sharding specification for parameters. Defaults to CLIPSharding.
    """
    if rngs is None:
        rngs = nnx.Rngs(0)
    self.vision_layers = vision_layers
    self.vision_hidden_size = vision_hidden_size
    self.vision_patch_size = vision_patch_size
    self.context_length = context_length
    self.vocab_size = vocab_size
    self.text_hidden_size = text_hidden_size
    self.num_text_heads = num_text_heads
    self.num_text_layers = num_text_layers
    self.dtype = dtype
    self._original_config = None

    self.vision_model = CLIPVisionModel(
        image_resolution=image_resolution,
        vision_layers=vision_layers,
        vision_hidden_size=vision_hidden_size,
        vision_patch_size=vision_patch_size,
        projection_dim=text_hidden_size,
        use_gradient_checkpointing=use_gradient_checkpointing,
        attention_fn=vision_attention_fn or attention_fn,
        rngs=rngs,
        dtype=dtype,
        param_dtype=param_dtype,
        sharding=sharding,
    )

    self.text_model = CLIPTextModel(
        context_length=context_length,
        vocab_size=vocab_size,
        text_hidden_size=text_hidden_size,
        num_text_heads=num_text_heads,
        num_text_layers=num_text_layers,
        use_gradient_checkpointing=use_gradient_checkpointing,
        attention_fn=text_attention_fn or attention_fn,
        rngs=rngs,
        dtype=dtype,
        param_dtype=param_dtype,
        sharding=sharding,
    )
    self.logit_scale = nnx.Param(nnx.with_partitioning(nnx.initializers.ones_init(), ())(rngs.params(), ()))

encode_image(image, do_projection=True)

Encode images into embeddings.

Parameters:

Name Type Description Default
image Float[Array, 'batch height width channels']

Batch of input images.

 required
do_projection bool

Whether the image encoder should do the visual projection layer. Defaults to true.

 True

Returns:

Type Description
Float[Array, 'batch text_hidden_size']

Float[Array, "batch text_hidden_size"]: Image embeddings.
Source code in src/jimm/models/clip/clip_model.py 
496
497
498
499
500
501
502
503
504
505
506
def encode_image(self, image: Float[Array, "batch height width channels"], do_projection: bool = True) -> Float[Array, "batch text_hidden_size"]:
    """Encode images into embeddings.

    Args:
        image (Float[Array, "batch height width channels"]): Batch of input images.
        do_projection (bool): Whether the image encoder should do the visual projection layer. Defaults to true.

    Returns:
        Float[Array, "batch text_hidden_size"]: Image embeddings.
    """
    return self.vision_model(image, do_projection)

encode_text(text)

Encode text tokens into embeddings.

Parameters:

Name Type Description Default
text Int[Array, 'batch context_length']

Batch of token sequences.

 required

Returns:

Type Description
Float[Array, 'batch text_hidden_size']

Float[Array, "batch text_hidden_size"]: Text embeddings.
Source code in src/jimm/models/clip/clip_model.py 
508
509
510
511
512
513
514
515
516
517
def encode_text(self, text: Int[Array, "batch context_length"]) -> Float[Array, "batch text_hidden_size"]:
    """Encode text tokens into embeddings.

    Args:
        text (Int[Array, "batch context_length"]): Batch of token sequences.

    Returns:
        Float[Array, "batch text_hidden_size"]: Text embeddings.
    """
    return self.text_model(text, do_projection=True)

from_config(config, *, rngs=None, dtype=jnp.float32, param_dtype=jnp.float32, sharding=CLIPSharding, use_gradient_checkpointing=False, attention_fn=None, vision_attention_fn=None, text_attention_fn=None) classmethod

Create model from HuggingFace-compatible config dict.

Parameters:

Name Type Description Default
config dict[str, Any]

Configuration with "text_config" and "vision_config" keys.

 required
rngs Rngs | None

Random number generator state.

 None
dtype DTypeLike

Data type for computations.

 float32
param_dtype DTypeLike

Data type for parameters.

 float32
sharding ShardingSpec

Sharding specification for parameters.

 CLIPSharding
use_gradient_checkpointing bool

Enable gradient checkpointing.

 False
attention_fn Callable[..., Any] | None

Custom attention function applied to both encoders. Defaults to None.

 None
vision_attention_fn Callable[..., Any] | None

Override attention_fn for the vision encoder only. Defaults to None.

 None
text_attention_fn Callable[..., Any] | None

Override attention_fn for the text encoder only. Defaults to None.

 None

Returns:

Type Description
CLIP

CLIP model with random weights.
Source code in src/jimm/models/clip/clip_model.py 
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
@classmethod
def from_config(
    cls,
    config: dict[str, Any],
    *,
    rngs: rnglib.Rngs | None = None,
    dtype: DTypeLike = jnp.float32,
    param_dtype: DTypeLike = jnp.float32,
    sharding: ShardingSpec = CLIPSharding,
    use_gradient_checkpointing: bool = False,
    attention_fn: Callable[..., Any] | None = None,
    vision_attention_fn: Callable[..., Any] | None = None,
    text_attention_fn: Callable[..., Any] | None = None,
) -> "CLIP":
    """Create model from HuggingFace-compatible config dict.

    Args:
        config: Configuration with "text_config" and "vision_config" keys.
        rngs: Random number generator state.
        dtype: Data type for computations.
        param_dtype: Data type for parameters.
        sharding: Sharding specification for parameters.
        use_gradient_checkpointing: Enable gradient checkpointing.
        attention_fn: Custom attention function applied to both encoders. Defaults to None.
        vision_attention_fn: Override attention_fn for the vision encoder only. Defaults to None.
        text_attention_fn: Override attention_fn for the text encoder only. Defaults to None.

    Returns:
        CLIP model with random weights.
    """
    if rngs is None:
        rngs = nnx.Rngs(0)
    text_config = config["text_config"]
    vision_config = config["vision_config"]

    return cls(
        image_resolution=vision_config["image_size"],
        vision_layers=vision_config["num_hidden_layers"],
        vision_hidden_size=vision_config["hidden_size"],
        vision_patch_size=vision_config["patch_size"],
        context_length=text_config["max_position_embeddings"],
        vocab_size=text_config["vocab_size"],
        text_hidden_size=text_config["hidden_size"],
        num_text_heads=text_config["num_attention_heads"],
        num_text_layers=text_config["num_hidden_layers"],
        use_gradient_checkpointing=use_gradient_checkpointing,
        attention_fn=attention_fn,
        vision_attention_fn=vision_attention_fn,
        text_attention_fn=text_attention_fn,
        rngs=rngs,
        dtype=dtype,
        param_dtype=param_dtype,
        sharding=sharding,
    )

from_pretrained(model_name_or_path, use_pytorch=False, rngs=None, dtype=jnp.float32, param_dtype=jnp.float32, sharding=CLIPSharding, use_gradient_checkpointing=False, attention_fn=None) classmethod

Load a pretrained CLIP model from a local path or HuggingFace Hub.

Parameters:

Name Type Description Default
model_name_or_path str

Path to local weights or HuggingFace model ID.

 required
use_pytorch bool

Whether to load from PyTorch weights. Defaults to False.

 False
rngs Rngs | None

Random number generator keys. If None, initializes to nnx.Rngs(0).

 None
dtype DTypeLike

Data type for computations. Defaults to jnp.float32.

 float32
param_dtype DTypeLike

Data type for parameters. Defaults to jnp.float32.

 float32
sharding ShardingSpec

Sharding specification for parameters. Defaults to CLIPSharding.

 CLIPSharding
use_gradient_checkpointing bool

Whether to use gradient checkpointing. Defaults to False.

 False
attention_fn Callable[..., Any] | None

Custom attention function (e.g. jimm.tokamax_attention or jimm.make_tokamax_attention("mosaic_tpu")). Defaults to None.

 None

Returns:

Name Type Description
CLIP CLIP

Pretrained CLIP model
Source code in src/jimm/models/clip/clip_model.py 
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
@classmethod
def from_pretrained(
    cls,
    model_name_or_path: str,
    use_pytorch: bool = False,
    rngs: rnglib.Rngs | None = None,
    dtype: DTypeLike = jnp.float32,
    param_dtype: DTypeLike = jnp.float32,
    sharding: ShardingSpec = CLIPSharding,
    use_gradient_checkpointing: bool = False,
    attention_fn: Callable[..., Any] | None = None,
) -> "CLIP":
    """Load a pretrained CLIP model from a local path or HuggingFace Hub.

    Args:
        model_name_or_path (str): Path to local weights or HuggingFace model ID.
        use_pytorch (bool): Whether to load from PyTorch weights. Defaults to False.
        rngs (rnglib.Rngs | None): Random number generator keys. If None, initializes to nnx.Rngs(0).
        dtype (DTypeLike): Data type for computations. Defaults to jnp.float32.
        param_dtype (DTypeLike): Data type for parameters. Defaults to jnp.float32.
        sharding (ShardingSpec): Sharding specification for parameters. Defaults to CLIPSharding.
        use_gradient_checkpointing (bool): Whether to use gradient checkpointing. Defaults to False.
        attention_fn (Callable[..., Any] | None, optional): Custom attention function (e.g. jimm.tokamax_attention or jimm.make_tokamax_attention("mosaic_tpu")). Defaults to None.

    Returns:
        CLIP: Pretrained CLIP model
    """
    from .params import load_from_pretrained

    return load_from_pretrained(cls, model_name_or_path, use_pytorch, rngs, dtype, param_dtype, sharding, use_gradient_checkpointing, attention_fn)

save_pretrained(save_directory)

Save the model weights and config in HuggingFace format.

Parameters:

Name Type Description Default
save_directory str

Directory path where the model will be saved.

 required
Source code in src/jimm/models/clip/clip_model.py 
627
628
629
630
631
632
633
634
635
def save_pretrained(self, save_directory: str) -> None:
    """Save the model weights and config in HuggingFace format.

    Args:
        save_directory (str): Directory path where the model will be saved.
    """
    from .params import save_pretrained

    save_pretrained(self, save_directory)