__init__.py 49.8 KB
Newer Older
1
2
3
# ruff: noqa: F821
# the above line disables the `undefined-name` rule for the model type variables

4
import torch
5
import enum
Nicolas Patry's avatar
Nicolas Patry committed
6
import os
7

8
from loguru import logger
9
from transformers.configuration_utils import PretrainedConfig
10
from transformers.models.auto import modeling_auto
Nicolas Patry's avatar
Nicolas Patry committed
11
from huggingface_hub import hf_hub_download, HfApi
12
from typing import Optional, List, Dict
13
from pathlib import Path
14

Nicolas Patry's avatar
Nicolas Patry committed
15
from text_generation_server.utils.speculate import get_speculate, set_speculate
16
from text_generation_server.models.model import Model
17
18
19
20
21
from text_generation_server.models.causal_lm import CausalLM, CausalLMBatchKeysLast
from text_generation_server.models.custom_modeling.opt_modeling import OPTForCausalLM
from text_generation_server.models.custom_modeling.mpt_modeling import (
    MPTForCausalLM,
)
22
from text_generation_server.models.bloom import BloomCausalLMBatch
23
24
25
from text_generation_server.models.custom_modeling.bloom_modeling import (
    BloomForCausalLM,
)
26
from text_generation_server.models.seq2seq_lm import Seq2SeqLM
27
28
29
30
31
32
33
34
from text_generation_server.models.galactica import GalacticaCausalLMBatch
from text_generation_server.models.custom_modeling.neox_modeling import (
    GPTNeoxForCausalLM,
)
from text_generation_server.models.custom_modeling.phi_modeling import (
    PhiConfig,
    PhiForCausalLM,
)
drbh's avatar
drbh committed
35
36
37
from text_generation_server.models.custom_modeling.flash_phi_moe_modeling import (
    PhiMoEConfig,
)
38
39
40
from text_generation_server.models.custom_modeling.t5_modeling import (
    T5ForConditionalGeneration,
)
41

42
43
44
45
46
47
48
49
50
51

from text_generation_server.utils.adapter import (
    AdapterParameters,
    build_layer_weight_lookup,
    load_and_merge_adapters,
    AdapterInfo,
)
from text_generation_server.adapters.lora import LoraWeights


52
from text_generation_server.utils.import_utils import SYSTEM
53
from text_generation_server.utils.log import log_master
54

55
56
57
58
59
60
61
62
63
64
65
66
67
68
# The flag below controls whether to allow TF32 on matmul. This flag defaults to False
# in PyTorch 1.12 and later.
torch.backends.cuda.matmul.allow_tf32 = True

# The flag below controls whether to allow TF32 on cuDNN. This flag defaults to True.
torch.backends.cudnn.allow_tf32 = True

# Disable gradients
torch.set_grad_enabled(False)

__all__ = [
    "Model",
    "CausalLM",
    "Seq2SeqLM",
69
    "get_model_with_lora_adapters",
70
71
]

72
FLASH_ATT_ERROR_MESSAGE = "{} requires Flash Attention enabled models."
73

74
FLASH_ATTENTION = True
75

76
try:
77
    from text_generation_server.models.flash_causal_lm import FlashCausalLM
78
    from text_generation_server.models.vlm_causal_lm import VlmCausalLM
Nicolas Patry's avatar
Nicolas Patry committed
79
    from text_generation_server.models.mllama_causal_lm import MllamaCausalLM
80
81
82
83
    from text_generation_server.models.custom_modeling.flash_deepseek_v2_modeling import (
        FlashDeepseekV2ForCausalLM,
        DeepseekV2Config,
    )
84
85
    from text_generation_server.models.custom_modeling.flash_llama_modeling import (
        FlashLlamaForCausalLM,
86
    )
87
88
    from text_generation_server.models.custom_modeling.flash_cohere_modeling import (
        FlashCohereForCausalLM,
OlivierDehaene's avatar
OlivierDehaene committed
89
    )
90
91
    from text_generation_server.models.custom_modeling.flash_gemma_modeling import (
        FlashGemmaForCausalLM,
OlivierDehaene's avatar
OlivierDehaene committed
92
    )
93
94
    from text_generation_server.models.custom_modeling.flash_gemma2_modeling import (
        FlashGemma2ForCausalLM,
95
    )
96
97
98
99
100
101
102
103
104
105
    from text_generation_server.models.custom_modeling.flash_dbrx_modeling import (
        FlashDbrxForCausalLM,
        DbrxConfig,
    )
    from text_generation_server.models.custom_modeling.flash_rw_modeling import (
        RWConfig,
        FlashRWForCausalLM,
    )
    from text_generation_server.models.custom_modeling.flash_neox_modeling import (
        FlashGPTNeoXForCausalLM,
Nicolas Patry's avatar
Nicolas Patry committed
106
    )
drbh's avatar
drbh committed
107
    from text_generation_server.models.pali_gemma import (
108
        PaliGemmaBatch,
drbh's avatar
drbh committed
109
    )
110
111
112
113
114
    from text_generation_server.models.custom_modeling.flash_pali_gemma_modeling import (
        PaliGemmaForConditionalGeneration,
    )
    from text_generation_server.models.custom_modeling.flash_phi_modeling import (
        FlashPhiForCausalLM,
115
    )
Nicolas Patry's avatar
Nicolas Patry committed
116
117
118
119
120
    from text_generation_server.models.idefics_causal_lm import IdeficsCausalLM
    from text_generation_server.models.mllama_causal_lm import MllamaCausalLMBatch
    from text_generation_server.models.custom_modeling.mllama import (
        MllamaForConditionalGeneration,
    )
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
    from text_generation_server.models.custom_modeling.llava_next import (
        LlavaNextForConditionalGeneration,
    )

    from text_generation_server.models.custom_modeling.flash_santacoder_modeling import (
        FlashSantacoderForCausalLM,
    )
    from text_generation_server.models.custom_modeling.flash_starcoder2_modeling import (
        FlashStarcoder2ForCausalLM,
    )
    from text_generation_server.models.custom_modeling.flash_qwen2_modeling import (
        Qwen2ForCausalLM,
    )
    from text_generation_server.models.custom_modeling.flash_mistral_modeling import (
        FlashMistralForCausalLM,
    )
    from text_generation_server.models.custom_modeling.flash_mixtral_modeling import (
        FlashMixtralForCausalLM,
    )
    from text_generation_server.models.custom_modeling.flash_gpt2_modeling import (
        FlashGPT2ForCausalLM,
    )
143
144
145
    from text_generation_server.models.custom_modeling.flash_gptj_modeling import (
        FlashGPTJForCausalLM,
    )
146
147
148
    from text_generation_server.models.custom_modeling.idefics2 import (
        Idefics2ForConditionalGeneration,
    )
149
    from text_generation_server.layers.attention import SUPPORTS_WINDOWING
150
except ImportError as e:
151
    log_master(logger.warning, f"Could not import Flash Attention enabled models: {e}")
152
    SUPPORTS_WINDOWING = False
153
    FLASH_ATTENTION = False
154

155
if FLASH_ATTENTION:
156
    __all__.append(FlashCausalLM)
Nicolas Patry's avatar
Nicolas Patry committed
157
    __all__.append(IdeficsCausalLM)
OlivierDehaene's avatar
OlivierDehaene committed
158

drbh's avatar
drbh committed
159
160
161
162
MAMBA_AVAILABLE = True
try:
    from text_generation_server.models.mamba import Mamba
except ImportError as e:
163
    log_master(logger.warning, f"Could not import Mamba: {e}")
drbh's avatar
drbh committed
164
165
166
167
    MAMBA_AVAILABLE = False

if MAMBA_AVAILABLE:
    __all__.append(Mamba)
OlivierDehaene's avatar
OlivierDehaene committed
168

169

170
class ModelType(enum.Enum):
171
172
173
174
175
    DEEPSEEK_V2 = {
        "type": "deepseek_v2",
        "name": "Deepseek V2",
        "url": "https://huggingface.co/deepseek-ai/DeepSeek-V2",
    }
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
    IDEFICS2 = {
        "type": "idefics2",
        "name": "Idefics 2",
        "url": "https://huggingface.co/HuggingFaceM4/idefics2-8b",
        "multimodal": True,
    }
    LLAVA_NEXT = {
        "type": "llava_next",
        "name": "Llava Next (1.6)",
        "url": "https://huggingface.co/llava-hf/llava-v1.6-vicuna-13b-hf",
        "multimodal": True,
    }
    LLAMA = {
        "type": "llama",
        "name": "Llama",
191
        "url": "https://huggingface.co/collections/meta-llama/llama-31-669fc079a0c406a149a5738f",
192
193
194
195
196
197
    }
    PHI3 = {
        "type": "phi3",
        "name": "Phi 3",
        "url": "https://huggingface.co/microsoft/Phi-3-mini-4k-instruct",
    }
198
199
200
201
202
    GRANITE = {
        "type": "granite",
        "name": "Granite",
        "url": "https://huggingface.co/ibm-granite/granite-3.0-8b-instruct",
    }
203
204
205
206
207
    GEMMA = {
        "type": "gemma",
        "name": "Gemma",
        "url": "https://huggingface.co/google/gemma-7b",
    }
208
209
210
211
212
    PALIGEMMA = {
        "type": "paligemma",
        "name": "PaliGemma",
        "url": "https://huggingface.co/google/paligemma-3b-pt-224",
    }
Nicolas Patry's avatar
Nicolas Patry committed
213
214
215
    GEMMA2 = {
        "type": "gemma2",
        "name": "Gemma2",
216
        "url": "https://huggingface.co/collections/google/gemma-2-release-667d6600fd5220e7b967f315",
Nicolas Patry's avatar
Nicolas Patry committed
217
    }
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
    COHERE = {
        "type": "cohere",
        "name": "Cohere",
        "url": "https://huggingface.co/CohereForAI/c4ai-command-r-plus",
    }
    DBRX = {
        "type": "dbrx",
        "name": "Dbrx",
        "url": "https://huggingface.co/databricks/dbrx-instruct",
    }
    MAMBA = {
        "type": "ssm",
        "name": "Mamba",
        "url": "https://huggingface.co/state-spaces/mamba-2.8b-slimpj",
    }
    MISTRAL = {
        "type": "mistral",
        "name": "Mistral",
236
        "url": "https://huggingface.co/mistralai/Mistral-Nemo-Instruct-2407",
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
    }
    MIXTRAL = {
        "type": "mixtral",
        "name": "Mixtral",
        "url": "https://huggingface.co/mistralai/Mixtral-8x22B-Instruct-v0.1",
    }
    GPT_BIGCODE = {
        "type": "gpt_bigcode",
        "name": "Gpt Bigcode",
        "url": "https://huggingface.co/bigcode/gpt_bigcode-santacoder",
    }
    PHI = {
        "type": "phi",
        "name": "Phi",
        "url": "https://huggingface.co/microsoft/phi-1_5",
    }
drbh's avatar
drbh committed
253
254
255
256
257
    PHI_MOE = {
        "type": "phimoe",
        "name": "PhiMoe",
        "url": "https://huggingface.co/microsoft/Phi-3.5-MoE-instruct",
    }
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
    BAICHUAN = {
        "type": "baichuan",
        "name": "Baichuan",
        "url": "https://huggingface.co/baichuan-inc/Baichuan2-7B-Chat",
    }
    FALCON = {
        "type": "falcon",
        "name": "Falcon",
        "url": "https://huggingface.co/tiiuae/falcon-7b-instruct",
    }
    STARCODER2 = {
        "type": "starcoder2",
        "name": "StarCoder 2",
        "url": "https://huggingface.co/bigcode/starcoder2-15b-instruct-v0.1",
    }
    QWEN2 = {
        "type": "qwen2",
        "name": "Qwen 2",
276
        "url": "https://huggingface.co/collections/Qwen/qwen2-6659360b33528ced941e557f",
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
    }
    OPT = {
        "type": "opt",
        "name": "Opt",
        "url": "https://huggingface.co/facebook/opt-6.7b",
    }
    T5 = {
        "type": "t5",
        "name": "T5",
        "url": "https://huggingface.co/google/flan-t5-xxl",
    }
    GALACTICA = {
        "type": "galactica",
        "name": "Galactica",
        "url": "https://huggingface.co/facebook/galactica-120b",
    }
    SANTACODER = {
        "type": "santacoder",
        "name": "SantaCoder",
        "url": "https://huggingface.co/bigcode/santacoder",
    }
    BLOOM = {
        "type": "bloom",
        "name": "Bloom",
        "url": "https://huggingface.co/bigscience/bloom-560m",
    }
    MPT = {
        "type": "mpt",
        "name": "Mpt",
        "url": "https://huggingface.co/mosaicml/mpt-7b-instruct",
    }
    GPT2 = {
        "type": "gpt2",
        "name": "Gpt2",
        "url": "https://huggingface.co/openai-community/gpt2",
    }
    GPT_NEOX = {
        "type": "gpt_neox",
        "name": "Gpt Neox",
        "url": "https://huggingface.co/EleutherAI/gpt-neox-20b",
    }
318
319
320
321
322
    GPTJ = {
        "type": "gptj",
        "name": "Gptj",
        "url": "https://huggingface.co/EleutherAI/gpt-j-6b",
    }
323
324
325
326
327
328
    IDEFICS = {
        "type": "idefics",
        "name": "Idefics",
        "url": "https://huggingface.co/HuggingFaceM4/idefics-9b",
        "multimodal": True,
    }
Nicolas Patry's avatar
Nicolas Patry committed
329
330
331
332
333
334
    MLLAMA = {
        "type": "mllama",
        "name": "Mllama",
        "url": "https://huggingface.co/meta-llama/Llama-3.2-11B-Vision-Instruct",
        "multimodal": True,
    }
335
336
337
338
339
340
341


__GLOBALS = locals()
for data in ModelType:
    __GLOBALS[data.name] = data.value["type"]


342
def get_model(
343
    model_id: str,
drbh's avatar
drbh committed
344
    lora_adapter_ids: Optional[List[str]],
345
346
347
    revision: Optional[str],
    sharded: bool,
    quantize: Optional[str],
Nicolas Patry's avatar
Nicolas Patry committed
348
    speculate: Optional[int],
349
    dtype: Optional[str],
350
    kv_cache_dtype: Optional[str],
351
    trust_remote_code: bool,
352
    max_input_tokens: int,
353
) -> Model:
354
    global FLASH_ATTENTION
355
356
357
358
359
360
361

    config_dict, _ = PretrainedConfig.get_config_dict(
        model_id, revision=revision, trust_remote_code=trust_remote_code
    )
    model_type = config_dict.get("model_type", None)

    quantization_config = config_dict.get("quantization_config", None)
362
    compression_config = config_dict.get("compression_config", None)
363
364
    if quantization_config is not None and quantize is None:
        method = quantization_config.get("quant_method", None)
365
        config_groups = quantization_config.get("config_groups", None)
366
367
368
        if method in {"gptq", "awq", "exl2"}:
            log_master(logger.info, f"Auto selecting quantization method {method}")
            quantize = method
369
        elif method == "fbgemm_fp8" or method == "fp8":
370
371
            log_master(logger.info, "Auto selecting quantization method fp8")
            quantize = "fp8"
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
        elif config_groups is not None:
            # TODO: at some point we should probably fully parse the compression
            # configuration to know which parameters are compressed.
            for _, group in config_groups.items():
                weights_config = group.get("weights")
                if weights_config is not None:
                    if (
                        weights_config["type"] == "float"
                        and weights_config["num_bits"] == 8
                    ):
                        log_master(
                            logger.info, "Auto selecting quantization method fp8"
                        )
                        quantize = "fp8"
                        break
387
388
        else:
            log_master(logger.warning, f"Unknown quantization method {method}")
389
    elif compression_config is not None:
390
        # `compression_config` renamed to `quantization_config`; support retained for backward compatibility.
391
392
393
394
395
396
397
398
399
400
401
402
403
404
        config_groups = compression_config.get("config_groups")
        if config_groups is not None:
            for _, group in config_groups.items():
                weights_config = group.get("weights")
                if weights_config is not None:
                    if (
                        weights_config["type"] == "float"
                        and weights_config["num_bits"] == 8
                    ):
                        log_master(
                            logger.info, "Auto selecting quantization method fp8"
                        )
                        quantize = "fp8"
                        break
405

406
    if dtype is None:
407
        if quantize in ["awq", "exl2", "gptq", "marlin"]:
Nicolas Patry's avatar
Nicolas Patry committed
408
409
410
411
412
            if SYSTEM == "ipex" and not hasattr(torch, "xpu"):
                dtype = torch.bfloat16
            else:
                # These quantizers only work with float16 params.
                dtype = torch.float16
413
414
415
416
        else:
            # Keep it as default for now and let
            # every model resolve their own default dtype.
            dtype = None
417
418
419
420
421
422
423
    elif dtype == "float16":
        dtype = torch.float16
    elif dtype == "bfloat16":
        dtype = torch.bfloat16
    else:
        raise RuntimeError(f"Unknown dtype {dtype}")

424
425
    if kv_cache_dtype is None:
        kv_cache_dtype = dtype
426
427
    elif kv_cache_dtype == "fp8_e4m3fn":
        kv_cache_dtype = torch.float8_e4m3fn
428
429
430
431
432
    elif kv_cache_dtype == "fp8_e5m2":
        kv_cache_dtype = torch.float8_e5m2
    else:
        raise RuntimeError(f"Unknown kv_cache_dtype: {kv_cache_dtype}")

Nicolas Patry's avatar
Nicolas Patry committed
433
434
435
436
437
    if speculate is not None:
        set_speculate(speculate)
    else:
        set_speculate(0)

Nicolas Patry's avatar
Nicolas Patry committed
438
    speculator = None
Nicolas Patry's avatar
Nicolas Patry committed
439
    if "medusa_num_heads" in config_dict:
440
441
        medusa_model_id = model_id
        medusa_revision = revision
Nicolas Patry's avatar
Nicolas Patry committed
442
443
444
445
446
        model_id = config_dict["base_model_name_or_path"]
        revision = "main"
        speculate_medusa = config_dict["medusa_num_heads"]
        if speculate is not None:
            if speculate > speculate_medusa:
OlivierDehaene's avatar
OlivierDehaene committed
447
                raise RuntimeError(
OlivierDehaene's avatar
OlivierDehaene committed
448
                    f"Speculate is set to `{speculate}` but this medusa models only has `{speculate_medusa}` heads, please make them match"
OlivierDehaene's avatar
OlivierDehaene committed
449
                )
Nicolas Patry's avatar
Nicolas Patry committed
450
451
452
453
454
455
456
457
            else:
                set_speculate(speculate)
        else:
            set_speculate(speculate_medusa)

        config_dict, _ = PretrainedConfig.get_config_dict(
            model_id, revision=revision, trust_remote_code=trust_remote_code
        )
Nicolas Patry's avatar
Nicolas Patry committed
458
459
        # Reload model type from parent.
        model_type = config_dict.get("model_type", None)
460
461
462
463
464
465
466
467
468
469
        is_local = Path(medusa_model_id).exists()
        if not is_local:
            medusa_config = hf_hub_download(
                medusa_model_id, revision=medusa_revision, filename="config.json"
            )
            hf_hub_download(
                medusa_model_id,
                revision=medusa_revision,
                filename="medusa_lm_head.safetensors",
            )
Nicolas Patry's avatar
Nicolas Patry committed
470
471
472
473
            speculator = {
                "path": Path(medusa_config).parent,
                "model_paths": ["medusa_lm_head.safetensors"],
            }
474
        else:
Nicolas Patry's avatar
Nicolas Patry committed
475
476
477
478
            speculator = {
                "path": Path(medusa_model_id),
                "model_paths": ["medusa_lm_head.safetensors"],
            }
479

Nicolas Patry's avatar
Nicolas Patry committed
480
        method = "medusa"
Nicolas Patry's avatar
Nicolas Patry committed
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
    elif model_type == "mlp_speculator":
        mlp_model_id = model_id
        mlp_revision = revision
        model_id = config_dict["base_model_name_or_path"]
        revision = "main"
        speculate_mlp = config_dict["n_predict"]
        if speculate is not None:
            if speculate > speculate_mlp:
                raise RuntimeError(
                    f"Speculate is set to `{speculate}` but this mlp_speculator models only has `{speculate_mlp}` heads, please make them match"
                )
            else:
                set_speculate(speculate)
        else:
            set_speculate(speculate_mlp)

        config_dict, _ = PretrainedConfig.get_config_dict(
            model_id, revision=revision, trust_remote_code=trust_remote_code
        )
        # Reload model type from parent.
        model_type = config_dict.get("model_type", None)
        is_local = Path(mlp_model_id).exists()
        extension = ".safetensors"
        if not is_local:
            mlp_speculator_config = hf_hub_download(
                mlp_model_id, revision=mlp_revision, filename="config.json"
            )
            api = HfApi()
            info = api.model_info(mlp_model_id, revision=mlp_revision)
            filenames = [
                s.rfilename
                for s in info.siblings
                if s.rfilename.endswith(extension)
                and len(s.rfilename.split("/")) == 1
                and "arguments" not in s.rfilename
                and "args" not in s.rfilename
                and "training" not in s.rfilename
            ]
            for filename in filenames:
                hf_hub_download(
                    mlp_model_id,
                    revision=mlp_revision,
                    filename=filename,
                )
525
526
527
528
529
            speculator_dir_path = Path(mlp_speculator_config).parent
            # if these are downloaded, they get converted to safetensors
            filenames.extend(
                [p for p in os.listdir(speculator_dir_path) if p.endswith(extension)]
            )
Nicolas Patry's avatar
Nicolas Patry committed
530
531
532
533
534
535
536
537
538
            speculator = {
                "path": Path(mlp_speculator_config).parent,
                "model_paths": filenames,
            }
        else:
            speculator = Path(mlp_model_id)
            filenames = [p for p in os.listdir(speculator) if p.endswith(extension)]
            speculator = {"path": speculator, "model_paths": filenames}
        method = "mlp_speculator"
Nicolas Patry's avatar
Nicolas Patry committed
539
540
541
542
543
    else:
        method = "n-gram"

    speculate = get_speculate()
    if speculate > 0:
544
545
546
        log_master(
            logger.info, f"Using speculation {method} with {speculate} input ids."
        )
Nicolas Patry's avatar
Nicolas Patry committed
547

drbh's avatar
drbh committed
548
549
550
551
552
553
554
555
556
557
    if model_type is None:
        # TODO: fix how we determine model type for Mamba
        if "ssm_cfg" in config_dict:
            # *only happens in Mamba case
            model_type = "ssm"
        else:
            raise RuntimeError(
                f"Could not determine model type for {model_id} revision {revision}"
            )

558
559
560
561
    if quantize == "exl2" and sharded:
        raise RuntimeError(
            "Sharding is currently not supported with `exl2` quantization"
        )
drbh's avatar
drbh committed
562
563
564
565
566
567

    sliding_window = (
        config_dict.get("sliding_window")
        if config_dict.get("sliding_window") is not None
        else -1
    )
568

569
570
571
    use_sliding_window = sliding_window is not None and sliding_window != -1
    needs_sliding_window = (
        max_input_tokens is not None and max_input_tokens > sliding_window
572
    )
573
574
575
576
    if use_sliding_window and needs_sliding_window and not SUPPORTS_WINDOWING:
        raise ValueError(
            f"The backend {SYSTEM} does not support sliding window attention that is used by the model type {model_type}. To use this model nonetheless with the {SYSTEM} backend, please launch TGI with the argument `--max-input-tokens` smaller than sliding_window={sliding_window} (got here max_input_tokens={max_input_tokens})."
        )
577

578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
    if model_type == DEEPSEEK_V2:
        if FLASH_ATTENTION:
            head_size = max(
                config_dict.get("qk_nope_dim", 128)
                + config_dict.get("qk_rope_dim", 64),
                config_dict.get("v_head_dim", 128),
            )
            return FlashCausalLM(
                model_id=model_id,
                model_class=FlashDeepseekV2ForCausalLM,
                revision=revision,
                quantize=quantize,
                speculator=speculator,
                default_dtype=torch.bfloat16,
                dtype=dtype,
593
                kv_cache_dtype=kv_cache_dtype,
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
                trust_remote_code=trust_remote_code,
                lora_adapter_ids=lora_adapter_ids,
                config_class=DeepseekV2Config,
                head_size=head_size,
            )
        elif sharded:
            raise NotImplementedError(
                FLASH_ATT_ERROR_MESSAGE.format("Sharded Deepseek V2")
            )
        else:
            return CausalLM.fallback(
                model_id,
                revision,
                quantize=quantize,
                speculator=speculator,
                dtype=dtype,
                trust_remote_code=trust_remote_code,
            )
    elif model_type == MAMBA:
drbh's avatar
drbh committed
613
614
615
616
        return Mamba(
            model_id,
            revision,
            quantize=quantize,
Nicolas Patry's avatar
Nicolas Patry committed
617
            speculator=speculator,
drbh's avatar
drbh committed
618
619
620
            dtype=dtype,
            trust_remote_code=trust_remote_code,
        )
621

OlivierDehaene's avatar
OlivierDehaene committed
622
    if model_id.startswith("facebook/galactica"):
623
624
625
626
627
        return CausalLM(
            model_id=model_id,
            # Yes galactica is just an OPT model.
            model_class=OPTForCausalLM,
            revision=revision,
OlivierDehaene's avatar
OlivierDehaene committed
628
            quantize=quantize,
Nicolas Patry's avatar
Nicolas Patry committed
629
            speculator=speculator,
OlivierDehaene's avatar
OlivierDehaene committed
630
631
            dtype=dtype,
            trust_remote_code=trust_remote_code,
632
            batch_class=GalacticaCausalLMBatch,
OlivierDehaene's avatar
OlivierDehaene committed
633
634
        )

635
    if (
636
637
        model_type == GPT_BIGCODE
        or model_type == GPT2
638
639
        and model_id.startswith("bigcode/")
    ):
640
        if FLASH_ATTENTION:
641
642
643
644
            return FlashCausalLM(
                model_id=model_id,
                model_class=FlashSantacoderForCausalLM,
                revision=revision,
645
                quantize=quantize,
Nicolas Patry's avatar
Nicolas Patry committed
646
                speculator=speculator,
647
                dtype=dtype,
648
                kv_cache_dtype=kv_cache_dtype,
649
                trust_remote_code=trust_remote_code,
650
651
652
                lora_adapter_ids=lora_adapter_ids,
                aliases={"transformer.wte.weight": ["lm_head.weight"]},
                num_kv_heads=1,
653
            )
654
655
656
657
        elif sharded:
            raise NotImplementedError(
                FLASH_ATT_ERROR_MESSAGE.format("Sharded Santacoder")
            )
658
        else:
659
660
661
            return CausalLM.fallback(
                model_id=model_id,
                revision=revision,
662
                quantize=quantize,
Nicolas Patry's avatar
Nicolas Patry committed
663
                speculator=speculator,
664
                dtype=dtype,
665
666
                trust_remote_code=trust_remote_code,
            )
667

668
    if model_type == BLOOM:
669
670
671
672
        return CausalLM(
            model_id=model_id,
            model_class=BloomForCausalLM,
            revision=revision,
673
            quantize=quantize,
Nicolas Patry's avatar
Nicolas Patry committed
674
            speculator=speculator,
675
676
            dtype=dtype,
            trust_remote_code=trust_remote_code,
677
            batch_class=BloomCausalLMBatch,
678
        )
679
    elif model_type == MPT:
680
681
682
683
        return CausalLM(
            model_id=model_id,
            model_class=MPTForCausalLM,
            revision=revision,
OlivierDehaene's avatar
OlivierDehaene committed
684
            quantize=quantize,
Nicolas Patry's avatar
Nicolas Patry committed
685
            speculator=speculator,
OlivierDehaene's avatar
OlivierDehaene committed
686
687
            dtype=dtype,
            trust_remote_code=trust_remote_code,
688
            batch_class=CausalLMBatchKeysLast,
689
        )
690
    elif model_type == GPT2:
691
        if FLASH_ATTENTION:
692
            try:
693
694
695
696
                return FlashCausalLM(
                    model_id=model_id,
                    model_class=FlashGPT2ForCausalLM,
                    revision=revision,
697
698
699
                    quantize=quantize,
                    speculator=speculator,
                    dtype=dtype,
700
                    kv_cache_dtype=kv_cache_dtype,
701
                    trust_remote_code=trust_remote_code,
702
                    lora_adapter_ids=lora_adapter_ids,
703
704
705
                )
            except RuntimeError as e:
                # Lots of legacy models with various weight names.
706
                log_master(logger.warning, f"Couldn't load flash gpt2 variant: {e}")
707
                return CausalLM.fallback(
708
709
710
711
712
713
714
                    model_id,
                    revision,
                    quantize=quantize,
                    speculator=speculator,
                    dtype=dtype,
                    trust_remote_code=trust_remote_code,
                )
715
716
717
        elif sharded:
            raise NotImplementedError(FLASH_ATT_ERROR_MESSAGE.format("Sharded GPT-2"))
        else:
718
            return CausalLM.fallback(
719
720
721
722
723
724
725
                model_id,
                revision,
                quantize=quantize,
                speculator=speculator,
                dtype=dtype,
                trust_remote_code=trust_remote_code,
            )
726
727
728
729
730
731
732
733
734
735
    elif model_type == GPTJ:
        if FLASH_ATTENTION:
            try:
                return FlashCausalLM(
                    model_id=model_id,
                    model_class=FlashGPTJForCausalLM,
                    revision=revision,
                    quantize=quantize,
                    speculator=speculator,
                    dtype=dtype,
736
                    kv_cache_dtype=kv_cache_dtype,
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
                    trust_remote_code=trust_remote_code,
                    lora_adapter_ids=lora_adapter_ids,
                )
            except RuntimeError as e:
                # Lots of legacy models with various weight names.
                log_master(logger.warning, f"Couldn't load flash gptj variant: {e}")
                return CausalLM.fallback(
                    model_id,
                    revision,
                    quantize=quantize,
                    speculator=speculator,
                    dtype=dtype,
                    trust_remote_code=trust_remote_code,
                )
        elif sharded:
            raise NotImplementedError(FLASH_ATT_ERROR_MESSAGE.format("Sharded GPT-J"))
        else:
            return CausalLM.fallback(
                model_id,
                revision,
                quantize=quantize,
                speculator=speculator,
                dtype=dtype,
                trust_remote_code=trust_remote_code,
            )
762
    elif model_type == GPT_NEOX:
763
        if FLASH_ATTENTION:
764
765
766
767
            from text_generation_server.models.custom_modeling.flash_neox_modeling import (
                GPTNeoXConfig,
            )

768
769
770
771
            return FlashCausalLM(
                model_id=model_id,
                model_class=FlashGPTNeoXForCausalLM,
                revision=revision,
772
                quantize=quantize,
Nicolas Patry's avatar
Nicolas Patry committed
773
                speculator=speculator,
774
                dtype=dtype,
775
                kv_cache_dtype=kv_cache_dtype,
776
                trust_remote_code=trust_remote_code,
777
                lora_adapter_ids=lora_adapter_ids,
778
                config_class=GPTNeoXConfig,
779
780
            )
        elif sharded:
781
782
783
784
            return CausalLM(
                model_id=model_id,
                model_class=GPTNeoxForCausalLM,
                revision=revision,
785
                quantize=quantize,
Nicolas Patry's avatar
Nicolas Patry committed
786
                speculator=speculator,
787
                dtype=dtype,
788
789
                trust_remote_code=trust_remote_code,
            )
790
        else:
791
            return CausalLM.fallback(
792
793
794
                model_id,
                revision,
                quantize=quantize,
Nicolas Patry's avatar
Nicolas Patry committed
795
                speculator=speculator,
796
                dtype=dtype,
797
798
                trust_remote_code=trust_remote_code,
            )
OlivierDehaene's avatar
OlivierDehaene committed
799

800
    elif model_type == PHI:
drbh's avatar
drbh committed
801
        if FLASH_ATTENTION:
802
803
804
805
            return FlashCausalLM(
                model_id=model_id,
                model_class=FlashPhiForCausalLM,
                revision=revision,
drbh's avatar
drbh committed
806
                quantize=quantize,
Nicolas Patry's avatar
Nicolas Patry committed
807
                speculator=speculator,
drbh's avatar
drbh committed
808
                dtype=dtype,
809
                kv_cache_dtype=kv_cache_dtype,
drbh's avatar
drbh committed
810
                trust_remote_code=trust_remote_code,
811
                lora_adapter_ids=lora_adapter_ids,
drbh's avatar
drbh committed
812
813
            )
        else:
814
            return CausalLM.fallback(
drbh's avatar
drbh committed
815
816
817
                model_id,
                revision,
                quantize=quantize,
Nicolas Patry's avatar
Nicolas Patry committed
818
                speculator=speculator,
drbh's avatar
drbh committed
819
820
821
822
                dtype=dtype,
                trust_remote_code=trust_remote_code,
            )

drbh's avatar
drbh committed
823
824
825
826
827
828
829
830
831
832
    elif model_type == PHI_MOE:
        if FLASH_ATTENTION:
            return FlashCausalLM(
                model_id=model_id,
                model_class=FlashLlamaForCausalLM,
                config_class=PhiMoEConfig,
                revision=revision,
                quantize=quantize,
                speculator=speculator,
                dtype=dtype,
833
                kv_cache_dtype=kv_cache_dtype,
drbh's avatar
drbh committed
834
835
836
837
838
839
840
841
842
843
844
845
846
                trust_remote_code=trust_remote_code,
                lora_adapter_ids=lora_adapter_ids,
            )
        else:
            return CausalLM.fallback(
                model_id,
                revision,
                quantize=quantize,
                speculator=speculator,
                dtype=dtype,
                trust_remote_code=trust_remote_code,
            )

drbh's avatar
drbh committed
847
848
    elif model_type == "phi-msft":
        if FLASH_ATTENTION:
OlivierDehaene's avatar
OlivierDehaene committed
849
850
851
            raise NotImplementedError(
                "Legacy phi-msft is not supported with Flash Attention"
            )
drbh's avatar
drbh committed
852
        else:
853
854
855
856
857
            return CausalLM(
                model_id=model_id,
                model_class=PhiForCausalLM,
                config_class=PhiConfig,
                revision=revision,
drbh's avatar
drbh committed
858
                quantize=quantize,
Nicolas Patry's avatar
Nicolas Patry committed
859
                speculator=speculator,
drbh's avatar
drbh committed
860
861
862
                dtype=dtype,
                trust_remote_code=trust_remote_code,
            )
863

864
865
866
867
868
869
    elif (
        model_type == LLAMA
        or model_type == BAICHUAN
        or model_type == PHI3
        or model_type == GRANITE
    ):
870
        if FLASH_ATTENTION:
871
872
873
874
            return FlashCausalLM(
                model_id=model_id,
                model_class=FlashLlamaForCausalLM,
                revision=revision,
875
                quantize=quantize,
Nicolas Patry's avatar
Nicolas Patry committed
876
                speculator=speculator,
877
                dtype=dtype,
878
                kv_cache_dtype=kv_cache_dtype,
879
                trust_remote_code=trust_remote_code,
drbh's avatar
drbh committed
880
                lora_adapter_ids=lora_adapter_ids,
881
            )
882
        elif sharded:
883
884
885
            raise NotImplementedError(
                FLASH_ATT_ERROR_MESSAGE.format(f"Sharded {model_type}")
            )
886
        else:
887
            return CausalLM.fallback(
888
889
890
                model_id,
                revision,
                quantize=quantize,
Nicolas Patry's avatar
Nicolas Patry committed
891
                speculator=speculator,
892
                dtype=dtype,
893
894
                trust_remote_code=trust_remote_code,
            )
895
    if model_type == GEMMA:
896
        if FLASH_ATTENTION:
897
898
899
900
            return FlashCausalLM(
                model_id=model_id,
                model_class=FlashGemmaForCausalLM,
                revision=revision,
901
                quantize=quantize,
Nicolas Patry's avatar
Nicolas Patry committed
902
                speculator=speculator,
903
                dtype=dtype,
904
                kv_cache_dtype=kv_cache_dtype,
905
906
                # Works better for these models
                default_dtype=torch.bfloat16,
907
                trust_remote_code=trust_remote_code,
908
                lora_adapter_ids=lora_adapter_ids,
909
910
            )
        elif sharded:
OlivierDehaene's avatar
OlivierDehaene committed
911
            raise NotImplementedError(FLASH_ATT_ERROR_MESSAGE.format("Sharded Gemma"))
912
        else:
913
            return CausalLM.fallback(
914
915
916
                model_id,
                revision,
                quantize=quantize,
Nicolas Patry's avatar
Nicolas Patry committed
917
                speculator=speculator,
918
919
920
                dtype=dtype,
                trust_remote_code=trust_remote_code,
            )
Nicolas Patry's avatar
Nicolas Patry committed
921
922
    elif model_type == GEMMA2:
        if FLASH_ATTENTION:
923
924
925
926
            return FlashCausalLM(
                model_id=model_id,
                model_class=FlashGemma2ForCausalLM,
                revision=revision,
Nicolas Patry's avatar
Nicolas Patry committed
927
928
929
                quantize=quantize,
                speculator=speculator,
                dtype=dtype,
930
                kv_cache_dtype=kv_cache_dtype,
931
932
                # Works better for these models
                default_dtype=torch.bfloat16,
Nicolas Patry's avatar
Nicolas Patry committed
933
                trust_remote_code=trust_remote_code,
934
                lora_adapter_ids=lora_adapter_ids,
Nicolas Patry's avatar
Nicolas Patry committed
935
936
937
938
            )
        elif sharded:
            raise NotImplementedError(FLASH_ATT_ERROR_MESSAGE.format("Sharded Gemma2"))
        else:
939
            return CausalLM.fallback(
Nicolas Patry's avatar
Nicolas Patry committed
940
941
942
943
944
945
946
                model_id,
                revision,
                quantize=quantize,
                speculator=speculator,
                dtype=dtype,
                trust_remote_code=trust_remote_code,
            )
947

948
    if model_type == COHERE:
OlivierDehaene's avatar
OlivierDehaene committed
949
        if FLASH_ATTENTION:
950
951
952
953
            return FlashCausalLM(
                model_id=model_id,
                model_class=FlashCohereForCausalLM,
                revision=revision,
OlivierDehaene's avatar
OlivierDehaene committed
954
                quantize=quantize,
Nicolas Patry's avatar
Nicolas Patry committed
955
                speculator=speculator,
OlivierDehaene's avatar
OlivierDehaene committed
956
                dtype=dtype,
957
                kv_cache_dtype=kv_cache_dtype,
OlivierDehaene's avatar
OlivierDehaene committed
958
                trust_remote_code=trust_remote_code,
959
                lora_adapter_ids=lora_adapter_ids,
OlivierDehaene's avatar
OlivierDehaene committed
960
961
962
963
            )
        elif sharded:
            raise NotImplementedError(FLASH_ATT_ERROR_MESSAGE.format("Sharded Cohere"))
        else:
964
            return CausalLM.fallback(
OlivierDehaene's avatar
OlivierDehaene committed
965
966
967
                model_id,
                revision,
                quantize=quantize,
Nicolas Patry's avatar
Nicolas Patry committed
968
                speculator=speculator,
OlivierDehaene's avatar
OlivierDehaene committed
969
970
971
972
                dtype=dtype,
                trust_remote_code=trust_remote_code,
            )

973
    if model_type == DBRX:
974
        if FLASH_ATTENTION:
975
976
977
978
            return FlashCausalLM(
                model_id=model_id,
                model_class=FlashDbrxForCausalLM,
                revision=revision,
979
                quantize=quantize,
Nicolas Patry's avatar
Nicolas Patry committed
980
                speculator=speculator,
981
                dtype=dtype,
982
                kv_cache_dtype=kv_cache_dtype,
983
984
                # Dbrx works better in bfloat16.
                default_dtype=torch.bfloat16,
985
                trust_remote_code=trust_remote_code,
986
987
                lora_adapter_ids=lora_adapter_ids,
                config_class=DbrxConfig,
988
989
990
991
            )
        elif sharded:
            raise NotImplementedError(FLASH_ATT_ERROR_MESSAGE.format("Sharded DBRX"))
        else:
992
            return CausalLM.fallback(
993
994
995
                model_id,
                revision,
                quantize=quantize,
Nicolas Patry's avatar
Nicolas Patry committed
996
                speculator=speculator,
997
998
999
1000
                dtype=dtype,
                trust_remote_code=trust_remote_code,
            )

1001
    if model_type in ["RefinedWeb", "RefinedWebModel", FALCON]:
1002
1003
        if sharded:
            if FLASH_ATTENTION:
1004
                if config_dict.get("alibi", False):
1005
                    raise NotImplementedError("sharded is not supported for this model")
1006
1007
1008
1009
                return FlashCausalLM(
                    model_id=model_id,
                    model_class=FlashRWForCausalLM,
                    revision=revision,
1010
                    quantize=quantize,
Nicolas Patry's avatar
Nicolas Patry committed
1011
                    speculator=speculator,
1012
                    dtype=dtype,
1013
                    kv_cache_dtype=kv_cache_dtype,
1014
1015
1016
1017
                    aliases={
                        "lm_head.weight": ["transformer.word_embeddings.weight"],
                        "transformer.word_embeddings.weight": ["lm_head.weight"],
                    },
1018
                    trust_remote_code=trust_remote_code,
1019
1020
                    lora_adapter_ids=lora_adapter_ids,
                    config_class=RWConfig,
1021
                )
1022
            raise NotImplementedError(FLASH_ATT_ERROR_MESSAGE.format("Sharded Falcon"))
1023
        else:
1024
            if FLASH_ATTENTION and not config_dict.get("alibi", False):
1025
1026
1027
1028
                return FlashCausalLM(
                    model_id=model_id,
                    model_class=FlashRWForCausalLM,
                    revision=revision,
1029
                    quantize=quantize,
Nicolas Patry's avatar
Nicolas Patry committed
1030
                    speculator=speculator,
1031
                    dtype=dtype,
1032
                    kv_cache_dtype=kv_cache_dtype,
1033
1034
1035
1036
                    aliases={
                        "lm_head.weight": ["transformer.word_embeddings.weight"],
                        "transformer.word_embeddings.weight": ["lm_head.weight"],
                    },
1037
                    trust_remote_code=trust_remote_code,
1038
1039
                    lora_adapter_ids=lora_adapter_ids,
                    config_class=RWConfig,
1040
1041
                )
            else:
1042
                return CausalLM.fallback(
1043
1044
1045
                    model_id,
                    revision,
                    quantize=quantize,
Nicolas Patry's avatar
Nicolas Patry committed
1046
                    speculator=speculator,
1047
                    dtype=dtype,
1048
1049
1050
                    trust_remote_code=trust_remote_code,
                )

1051
    if model_type == MISTRAL:
1052
        if FLASH_ATTENTION:
1053
1054
1055
1056
            return FlashCausalLM(
                model_id=model_id,
                model_class=FlashMistralForCausalLM,
                revision=revision,
1057
                quantize=quantize,
Nicolas Patry's avatar
Nicolas Patry committed
1058
                speculator=speculator,
1059
                dtype=dtype,
1060
                kv_cache_dtype=kv_cache_dtype,
1061
                trust_remote_code=trust_remote_code,
1062
                lora_adapter_ids=lora_adapter_ids,
1063
            )
OlivierDehaene's avatar
OlivierDehaene committed
1064
1065
1066
        elif sharded:
            raise NotImplementedError(FLASH_ATT_ERROR_MESSAGE.format("Sharded Mistral"))
        else:
1067
            return CausalLM.fallback(
OlivierDehaene's avatar
OlivierDehaene committed
1068
1069
1070
                model_id,
                revision,
                quantize=quantize,
Nicolas Patry's avatar
Nicolas Patry committed
1071
                speculator=speculator,
OlivierDehaene's avatar
OlivierDehaene committed
1072
1073
1074
                dtype=dtype,
                trust_remote_code=trust_remote_code,
            )
OlivierDehaene's avatar
OlivierDehaene committed
1075

1076
    if model_type == MIXTRAL:
1077
        if FLASH_ATTENTION:
1078
1079
1080
1081
            return FlashCausalLM(
                model_id=model_id,
                model_class=FlashMixtralForCausalLM,
                revision=revision,
OlivierDehaene's avatar
OlivierDehaene committed
1082
                quantize=quantize,
Nicolas Patry's avatar
Nicolas Patry committed
1083
                speculator=speculator,
OlivierDehaene's avatar
OlivierDehaene committed
1084
                dtype=dtype,
1085
                kv_cache_dtype=kv_cache_dtype,
OlivierDehaene's avatar
OlivierDehaene committed
1086
                trust_remote_code=trust_remote_code,
1087
                lora_adapter_ids=lora_adapter_ids,
OlivierDehaene's avatar
OlivierDehaene committed
1088
            )
OlivierDehaene's avatar
OlivierDehaene committed
1089
1090
1091
        elif sharded:
            raise NotImplementedError(FLASH_ATT_ERROR_MESSAGE.format("Sharded Mixtral"))
        else:
1092
            return CausalLM.fallback(
OlivierDehaene's avatar
OlivierDehaene committed
1093
1094
1095
                model_id,
                revision,
                quantize=quantize,
Nicolas Patry's avatar
Nicolas Patry committed
1096
                speculator=speculator,
OlivierDehaene's avatar
OlivierDehaene committed
1097
1098
1099
1100
                dtype=dtype,
                trust_remote_code=trust_remote_code,
            )

1101
    if model_type == STARCODER2:
1102
        if FLASH_ATTENTION:
1103
1104
1105
1106
            return FlashCausalLM(
                model_id=model_id,
                model_class=FlashStarcoder2ForCausalLM,
                revision=revision,
OlivierDehaene's avatar
OlivierDehaene committed
1107
                quantize=quantize,
1108
                speculator=speculator,
OlivierDehaene's avatar
OlivierDehaene committed
1109
                dtype=dtype,
1110
                kv_cache_dtype=kv_cache_dtype,
OlivierDehaene's avatar
OlivierDehaene committed
1111
                trust_remote_code=trust_remote_code,
1112
                lora_adapter_ids=lora_adapter_ids,
OlivierDehaene's avatar
OlivierDehaene committed
1113
1114
1115
1116
1117
1118
            )
        elif sharded:
            raise NotImplementedError(
                FLASH_ATT_ERROR_MESSAGE.format("Sharded Starcoder2")
            )
        else:
1119
            return CausalLM.fallback(
OlivierDehaene's avatar
OlivierDehaene committed
1120
1121
1122
                model_id,
                revision,
                quantize=quantize,
Nicolas Patry's avatar
Nicolas Patry committed
1123
                speculator=speculator,
OlivierDehaene's avatar
OlivierDehaene committed
1124
1125
1126
1127
                dtype=dtype,
                trust_remote_code=trust_remote_code,
            )

1128
    if model_type == QWEN2:
1129
        if FLASH_ATTENTION:
1130
1131
1132
1133
            return FlashCausalLM(
                model_id=model_id,
                model_class=Qwen2ForCausalLM,
                revision=revision,
OlivierDehaene's avatar
OlivierDehaene committed
1134
                quantize=quantize,
1135
                speculator=speculator,
OlivierDehaene's avatar
OlivierDehaene committed
1136
                dtype=dtype,
1137
                kv_cache_dtype=kv_cache_dtype,
OlivierDehaene's avatar
OlivierDehaene committed
1138
                trust_remote_code=trust_remote_code,
1139
                lora_adapter_ids=lora_adapter_ids,
OlivierDehaene's avatar
OlivierDehaene committed
1140
1141
1142
1143
            )
        elif sharded:
            raise NotImplementedError(FLASH_ATT_ERROR_MESSAGE.format("Sharded Qwen2"))
        else:
1144
            return CausalLM.fallback(
OlivierDehaene's avatar
OlivierDehaene committed
1145
1146
1147
                model_id,
                revision,
                quantize=quantize,
Nicolas Patry's avatar
Nicolas Patry committed
1148
                speculator=speculator,
OlivierDehaene's avatar
OlivierDehaene committed
1149
                dtype=dtype,
OlivierDehaene's avatar
OlivierDehaene committed
1150
1151
                trust_remote_code=trust_remote_code,
            )
1152

1153
    if model_type == OPT:
1154
1155
1156
1157
        return CausalLM(
            model_id=model_id,
            model_class=OPTForCausalLM,
            revision=revision,
1158
            quantize=quantize,
Nicolas Patry's avatar
Nicolas Patry committed
1159
            speculator=speculator,
1160
1161
            dtype=dtype,
            trust_remote_code=trust_remote_code,
1162
        )
1163

1164
    if model_type == T5:
1165
1166
1167
1168
        return Seq2SeqLM(
            model_id=model_id,
            model_class=T5ForConditionalGeneration,
            revision=revision,
1169
            quantize=quantize,
Nicolas Patry's avatar
Nicolas Patry committed
1170
            speculator=speculator,
1171
            dtype=dtype,
1172
            trust_remote_code=trust_remote_code,
1173
1174
1175
1176
1177
1178
            aliases={
                "shared.weight": [
                    "encoder.embed_tokens.weight",
                    "decoder.embed_tokens.weight",
                ]
            },
1179
        )
1180
    if model_type == IDEFICS:
1181
        if FLASH_ATTENTION:
Nicolas Patry's avatar
Nicolas Patry committed
1182
            return IdeficsCausalLM(
OlivierDehaene's avatar
OlivierDehaene committed
1183
1184
1185
                model_id,
                revision,
                quantize=quantize,
Nicolas Patry's avatar
Nicolas Patry committed
1186
                speculator=speculator,
OlivierDehaene's avatar
OlivierDehaene committed
1187
1188
1189
                dtype=dtype,
                trust_remote_code=trust_remote_code,
            )
1190
1191
        else:
            raise NotImplementedError(FLASH_ATT_ERROR_MESSAGE.format("Idefics"))
Nicolas Patry's avatar
Nicolas Patry committed
1192
1193
1194
1195
1196
1197
1198
1199
1200
1201
1202
1203
1204
1205
1206
1207
    if model_type == MLLAMA:
        if FLASH_ATTENTION:
            return MllamaCausalLM(
                model_id=model_id,
                model_class=MllamaForConditionalGeneration,
                batch_class=MllamaCausalLMBatch,
                revision=revision,
                quantize=quantize,
                speculator=speculator,
                dtype=dtype,
                default_dtype=torch.bfloat16,
                trust_remote_code=trust_remote_code,
                lora_adapter_ids=lora_adapter_ids,
            )
        else:
            raise NotImplementedError(FLASH_ATT_ERROR_MESSAGE.format("Mllama"))
1208
    if model_type == IDEFICS2:
Nicolas Patry's avatar
Nicolas Patry committed
1209
        if FLASH_ATTENTION:
1210
1211
1212
1213
            return VlmCausalLM(
                model_id=model_id,
                model_class=Idefics2ForConditionalGeneration,
                revision=revision,
Nicolas Patry's avatar
Nicolas Patry committed
1214
                quantize=quantize,
Nicolas Patry's avatar
Nicolas Patry committed
1215
                speculator=speculator,
Nicolas Patry's avatar
Nicolas Patry committed
1216
                dtype=dtype,
1217
                kv_cache_dtype=kv_cache_dtype,
Nicolas Patry's avatar
Nicolas Patry committed
1218
                trust_remote_code=trust_remote_code,
1219
1220
1221
1222
                lora_adapter_ids=lora_adapter_ids,
                # XXX: Extremely important to cap resolution in order to limit
                # VRAM usage.
                processor_kwargs={"size": {"longest_edge": 448, "shortest_edge": 378}},
Nicolas Patry's avatar
Nicolas Patry committed
1223
1224
1225
            )
        else:
            raise NotImplementedError(FLASH_ATT_ERROR_MESSAGE.format("Idefics"))
1226
    if model_type == PALIGEMMA:
drbh's avatar
drbh committed
1227
        if FLASH_ATTENTION:
1228
1229
1230
1231
            return VlmCausalLM(
                model_id=model_id,
                model_class=PaliGemmaForConditionalGeneration,
                revision=revision,
drbh's avatar
drbh committed
1232
1233
1234
                quantize=quantize,
                speculator=speculator,
                dtype=dtype,
1235
                kv_cache_dtype=kv_cache_dtype,
1236
1237
                # Works better for these models
                default_dtype=torch.bfloat16,
drbh's avatar
drbh committed
1238
                trust_remote_code=trust_remote_code,
1239
1240
                lora_adapter_ids=lora_adapter_ids,
                batch_class=PaliGemmaBatch,
drbh's avatar
drbh committed
1241
1242
1243
            )
        else:
            raise NotImplementedError(FLASH_ATT_ERROR_MESSAGE.format("Idefics"))
1244

1245
    if model_type == LLAVA_NEXT:
1246
        if FLASH_ATTENTION:
1247
1248
1249
1250
            return VlmCausalLM(
                model_class=LlavaNextForConditionalGeneration,
                model_id=model_id,
                revision=revision,
1251
                quantize=quantize,
Nicolas Patry's avatar
Nicolas Patry committed
1252
                speculator=speculator,
1253
                dtype=dtype,
1254
                kv_cache_dtype=kv_cache_dtype,
1255
1256
1257
1258
1259
                trust_remote_code=trust_remote_code,
            )
        else:
            raise NotImplementedError(FLASH_ATT_ERROR_MESSAGE.format("LlavaNext"))

1260
    if sharded:
1261
        raise NotImplementedError("sharded is not supported for AutoModel")
1262
    if quantize == "gptq":
1263
        raise NotImplementedError(
1264
1265
            "gptq quantization is not supported for AutoModel, you can try to quantize it with `text-generation-server quantize ORIGINAL_MODEL_ID NEW_MODEL_ID`"
        )
1266
    if quantize == "awq":
1267
        raise NotImplementedError("awq quantization is not supported for AutoModel")
Nicolas Patry's avatar
Nicolas Patry committed
1268
    elif (quantize == "bitsandbytes-fp4") or (quantize == "bitsandbytes-nf4"):
1269
        raise NotImplementedError("4bit quantization is not supported for AutoModel")
OlivierDehaene's avatar
OlivierDehaene committed
1270
    elif quantize == "eetq":
1271
        raise NotImplementedError("Eetq quantization is not supported for AutoModel")
1272
1273
    elif quantize == "exl2":
        raise NotImplementedError("exl2 quantization is not supported for AutoModel")
1274
    if model_type in modeling_auto.MODEL_FOR_CAUSAL_LM_MAPPING_NAMES:
1275
        return CausalLM.fallback(
1276
1277
1278
            model_id,
            revision,
            quantize=quantize,
Nicolas Patry's avatar
Nicolas Patry committed
1279
            speculator=speculator,
1280
1281
            dtype=dtype,
            trust_remote_code=trust_remote_code,
1282
        )
1283
    if model_type in modeling_auto.MODEL_FOR_SEQ_TO_SEQ_CAUSAL_LM_MAPPING_NAMES:
1284
        return Seq2SeqLM.fallback(
1285
1286
1287
            model_id,
            revision,
            quantize=quantize,
Nicolas Patry's avatar
Nicolas Patry committed
1288
            speculator=speculator,
1289
1290
            dtype=dtype,
            trust_remote_code=trust_remote_code,
1291
1292
        )

1293
    auto_map = config_dict.get("auto_map", None)
1294
1295
    if trust_remote_code and auto_map is not None:
        if "AutoModelForCausalLM" in auto_map.keys():
1296
            return CausalLM.fallback(
1297
1298
1299
                model_id,
                revision,
                quantize=quantize,
Nicolas Patry's avatar
Nicolas Patry committed
1300
                speculator=speculator,
1301
                dtype=dtype,
1302
1303
                trust_remote_code=trust_remote_code,
            )
1304
        if "AutoModelForSeq2SeqLM" in auto_map.keys():
1305
            return Seq2SeqLM.fallback(
1306
1307
1308
                model_id,
                revision,
                quantize=quantize,
Nicolas Patry's avatar
Nicolas Patry committed
1309
                speculator=speculator,
1310
                dtype=dtype,
1311
1312
                trust_remote_code=trust_remote_code,
            )
1313
1314

    raise ValueError(f"Unsupported model type {model_type}")
1315
1316
1317
1318
1319
1320
1321
1322
1323
1324
1325
1326


# get_model_with_lora_adapters wraps the internal get_model function and adds support for loading adapters
# this provides a post model loading hook to load adapters into the model after the model has been loaded
def get_model_with_lora_adapters(
    model_id: str,
    lora_adapters: Optional[List[AdapterInfo]],
    revision: Optional[str],
    sharded: bool,
    quantize: Optional[str],
    speculate: Optional[int],
    dtype: Optional[str],
1327
    kv_cache_dtype: Optional[str],
1328
1329
1330
1331
1332
1333
1334
1335
1336
1337
1338
1339
1340
    trust_remote_code: bool,
    max_input_tokens: int,
    adapter_to_index: Dict[str, int],
):
    lora_adapter_ids = [adapter.id for adapter in lora_adapters]
    model = get_model(
        model_id,
        lora_adapter_ids,
        revision,
        sharded,
        quantize,
        speculate,
        dtype,
1341
        kv_cache_dtype,
1342
1343
1344
1345
1346
1347
1348
1349
1350
1351
1352
1353
1354
1355
1356
1357
1358
1359
1360
1361
1362
1363
1364
1365
1366
1367
1368
1369
1370
1371
1372
1373
1374
1375
1376
1377
1378
1379
1380
1381
1382
1383
1384
1385
1386
1387
1388
1389
1390
1391
1392
1393
        trust_remote_code,
        max_input_tokens,
    )

    if len(lora_adapters) > 0:
        target_to_layer = build_layer_weight_lookup(model.model)

        for index, adapter in enumerate(lora_adapters):
            # The AdapterParameters object allows for merging multiple adapters into a single adapter.
            # At the moment, we only support loading a single adapter into the model, but we keep the
            # AdapterParameters object for easier extension in the future.
            adapter_parameters = AdapterParameters(
                adapter_info=[adapter],
                # when merging multiple adapters we can weight them differently
                # if this is not set, all adapters will be weighted equally
                # see: text_generation_server.utils.merges.strategies for impl
                weights=None,
                merge_strategy=0,
                density=1.0,
                majority_sign_method=0,
            )

            adapter_index = index + 1
            adapter_to_index[adapter.id] = adapter_index

            logger.info(
                f"Loading adapter weights into model: {','.join([adapter.id for adapter in adapter_parameters.adapter_info])}"
            )
            weight_names = tuple([v[0] for v in target_to_layer.values()])
            (
                module_map,
                adapter_config,
                adapter_weight_names,
                adapter_tokenizer,
            ) = load_and_merge_adapters(
                model.model_id,
                adapter_parameters,
                adapter_index,
                weight_names,
                False,
            )

            unused_weight_names = adapter_weight_names.copy()

            adapter_layers = [
                "q_proj",
                "k_proj",
                "v_proj",
                "o_proj",
                "gate_proj",
                "up_proj",
                "down_proj",
1394
                "qkv_proj",
1395
1396
1397
1398
1399
1400
1401
1402
1403
1404
1405
1406
1407
1408
1409
1410
1411
1412
1413
1414
1415
1416
1417
1418
1419
1420
1421
            ]

            for layer_name in adapter_layers:
                nlayers = (
                    1 if layer_name == "lm_head" else len(model.model.model.layers)
                )
                adapter_weights = LoraWeights.prepare_weights(
                    config=adapter_config,
                    module_map=module_map,
                    layer_type=layer_name,
                    unused_weight_names=unused_weight_names,
                    nlayers=nlayers,
                    dtype=model.dtype,
                    world_size=model.world_size,
                    process_group=model.process_group,
                    target_to_layer=target_to_layer,
                )

                if adapter_weights is None:
                    continue

                model.layer_to_adapter_weights[layer_name].add_adapter(
                    adapter_index, adapter_weights
                )

            if len(unused_weight_names) > 0:
                logger.warning(
1422
                    f"{','.join([a.id for a in lora_adapters])} unused adapter weights: {unused_weight_names}"
1423
1424
1425
1426
1427
1428
1429
1430
                )

            if adapter_tokenizer is not None:
                model.tokenizers.add_tokenizer(adapter_index, adapter_tokenizer)

            model.loaded_adapters.add(adapter_index)

    return model