linear.py 65 KB
Newer Older
1
# SPDX-License-Identifier: Apache-2.0
2
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
3

4
import itertools
5
from abc import abstractmethod
6
from typing import Any, Literal, Optional, Union
7
8

import torch
9
import torch.nn as nn
10
from torch.nn.parameter import Parameter, UninitializedParameter
11

12
from vllm import envs
13
14
15
16
17
from vllm.distributed import (divide, get_tensor_model_parallel_rank,
                              get_tensor_model_parallel_world_size,
                              split_tensor_along_last_dim,
                              tensor_model_parallel_all_gather,
                              tensor_model_parallel_all_reduce)
18
from vllm.logger import init_logger
19
20
from vllm.model_executor.layers.quantization.base_config import (
    QuantizationConfig, QuantizeMethodBase)
21
from vllm.model_executor.layers.utils import dispatch_unquantized_gemm
22
# yapf: disable
23
from vllm.model_executor.parameter import (BasevLLMParameter,
24
                                           BlockQuantScaleParameter,
25
                                           PackedColumnParameter,
26
                                           PackedvLLMParameter,
27
28
                                           PerTensorScaleParameter,
                                           RowvLLMParameter)
29
# yapf: enable
30
from vllm.model_executor.utils import set_weight_attrs
31
from vllm.platforms import current_platform
32
33
34

logger = init_logger(__name__)

35
WEIGHT_LOADER_V2_SUPPORTED = [
36
    "CompressedTensorsLinearMethod",
37
38
    "BitBLASLinearMethod",
    "GPTQBitBLASLinearMethod",
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
    "AWQMarlinLinearMethod",
    "AWQLinearMethod",
    "GPTQMarlinLinearMethod",
    "Fp8LinearMethod",
    "MarlinLinearMethod",
    "QQQLinearMethod",
    "GPTQMarlin24LinearMethod",
    "TPUInt8LinearMethod",
    "GPTQLinearMethod",
    "FBGEMMFp8LinearMethod",
    "ModelOptFp8LinearMethod",
    "IPEXAWQLinearMethod",
    "IPEXGPTQLinearMethod",
    "HQQMarlinMethod",
    "QuarkLinearMethod",
    "ModelOptNvFp4LinearMethod",
55
]
56

57

58
59
60
61
62
63
64
65
66
def adjust_bitblas_shard(param, shard_size, shard_offset):
    bitblas_tile_size = getattr(param, "bitblas_tile_size", None)
    if bitblas_tile_size is not None:
        return (shard_size // bitblas_tile_size,
                shard_offset // bitblas_tile_size)

    return shard_size, shard_offset


67
68
69
70
71
72
73
74
def adjust_marlin_shard(param, shard_size, shard_offset):
    marlin_tile_size = getattr(param, "marlin_tile_size", None)
    if marlin_tile_size is None:
        return shard_size, shard_offset

    return shard_size * marlin_tile_size, shard_offset * marlin_tile_size


75
def adjust_bitsandbytes_4bit_shard(param: Parameter,
76
77
                                   shard_offsets: dict[str, tuple[int, int]],
                                   loaded_shard_id: str) -> tuple[int, int]:
78
79
    """Adjust the quantization offsets and sizes for BitsAndBytes sharding."""

80
81
    total, _ = shard_offsets["total"]
    orig_offset, orig_size = shard_offsets[loaded_shard_id]
82
83
84
85
86
87
88
89

    quantized_total = param.data.shape[0]
    quantized_offset = orig_offset * quantized_total // total
    quantized_size = orig_size * quantized_total // total

    return quantized_size, quantized_offset


90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
def adjust_scalar_to_fused_array(param, loaded_weight, shard_id):
    """For fused modules (QKV and MLP) we have an array of length
    N that holds 1 scale for each "logical" matrix. So the param
    is an array of length N. The loaded_weight corresponds to 
    one of the shards on disk. Here, we slice the param based on 
    the shard_id for loading.
    """
    qkv_idxs = {"q": 0, "k": 1, "v": 2}

    if isinstance(shard_id, str):
        shard_id = qkv_idxs[shard_id]
    elif not isinstance(shard_id, int):
        raise ValueError(f"Unknown Shard Id {shard_id}")

    # AutoFP8 scales do not have a shape
    # compressed-tensors scales do have a shape
    if len(loaded_weight.shape) != 0:
        assert loaded_weight.shape[0] == 1
        loaded_weight = loaded_weight[0]

    return param[shard_id], loaded_weight


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
# TODO(Isotr0py): We might need a more flexible structure to handle
# bitsandbytes shard offsets.
def left_shift_bitsandbytes_4bit_shard(bnb_weight_attrs: dict[str, Any]):
    """
    Separate the BitsAndBytes 4-bit shard.

    For example, given bnb weight attributes as below:
    {
        'bnb_shard_offsets': array([0, 4, 8, 16]), 
        'bnb_quant_state': {0: ..., 1: ..., 2: ...},
    }

    The function will return:
    {
        'bnb_shard_offsets': array([0, 4]), 
        'bnb_quant_state': {0: ...},
    }
    and
    {
        'bnb_shard_offsets': array([0, 4, 12]),
        'bnb_quant_state': {0: ..., 1: ...},
    }
    """
    shard_offsets = bnb_weight_attrs["bnb_shard_offsets"]
    offset_l = shard_offsets[:2]
    offset_r = shard_offsets[1:] - shard_offsets[1]
    quant_state_l = {0: bnb_weight_attrs["bnb_quant_state"][0]}
    quant_state_r = {
        i - 1: bnb_weight_attrs["bnb_quant_state"][i]
        for i in range(1,
                       len(shard_offsets) - 1)
    }
    left = dict(bnb_shard_offsets=offset_l, bnb_quant_state=quant_state_l)
    right = dict(bnb_shard_offsets=offset_r, bnb_quant_state=quant_state_r)
    return left, right


150
class LinearMethodBase(QuantizeMethodBase):
151
152
153
    """Base class for different (maybe quantized) linear methods."""

    @abstractmethod
154
155
    def create_weights(self, layer: torch.nn.Module,
                       input_size_per_partition: int,
156
                       output_partition_sizes: list[int], input_size: int,
157
158
                       output_size: int, params_dtype: torch.dtype,
                       **extra_weight_attrs):
159
160
        """Create weights for a linear layer. 
           The weights will be set as attributes of the layer.
161

162
163
164
165
166
167
168
169
170
171
        Args:
            layer: The layer that is using the LinearMethodBase factory.
            input_size_per_partition: Size of the weight input dim on rank X.
            output_partition_sizes: Sizes of the output dim of each logical 
                weight on rank X. E.g., output_partition_sizes for QKVLinear
                is a list contains the width of Wq, Wk, Wv on rank X.
            input_size: Size of the input dim of the weight across all ranks.
            output_size: Size of the output dim of the weight across all ranks.
            params_dtype: Datatype of the parameters.
        """
172
173
174
        raise NotImplementedError

    @abstractmethod
175
176
177
178
    def apply(self,
              layer: torch.nn.Module,
              x: torch.Tensor,
              bias: Optional[torch.Tensor] = None) -> torch.Tensor:
179
180
        """Apply the weights in layer to the input tensor.
        Expects create_weights to have been called before on the layer."""
181
182
183
184
        raise NotImplementedError


class UnquantizedLinearMethod(LinearMethodBase):
185
    """Linear method without quantization."""
186

187
188
    def create_weights(self, layer: torch.nn.Module,
                       input_size_per_partition: int,
189
                       output_partition_sizes: list[int], input_size: int,
190
191
                       output_size: int, params_dtype: torch.dtype,
                       **extra_weight_attrs):
192
        weight = Parameter(torch.empty(sum(output_partition_sizes),
CHU Tianxiang's avatar
CHU Tianxiang committed
193
                                       input_size_per_partition,
194
195
196
                                       dtype=params_dtype),
                           requires_grad=False)
        set_weight_attrs(weight, {"input_dim": 1, "output_dim": 0})
197
198
        layer.register_parameter("weight", weight)
        set_weight_attrs(weight, extra_weight_attrs)
199

200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
    def process_weights_after_loading(self, layer: torch.nn.Module) -> None:
        if current_platform.is_cpu() and envs.VLLM_CPU_SGL_KERNEL:
            N, K = layer.weight.size()
            dtype = layer.weight.dtype
            if (torch._C._cpu._is_amx_tile_supported()
                    and dtype == torch.bfloat16 and N % 32 == 0
                    and K % 32 == 0):
                packed_weight = torch.ops._C.convert_weight_packed(
                    layer.weight)
                assert packed_weight.size() == layer.weight.size()
                layer.weight.copy_(packed_weight)
                if layer.bias is not None:
                    layer.bias = Parameter(layer.bias.to(torch.float32),
                                           requires_grad=False)
                layer.use_cpu_sgl = True
            else:
                logger.warning(
                    "CPU SGL kernels require Intel AMX support,"
                    " bfloat16 weight, IC and OC are divisible by 32.")
                layer.use_cpu_sgl = False

221
222
223
224
    def apply(self,
              layer: torch.nn.Module,
              x: torch.Tensor,
              bias: Optional[torch.Tensor] = None) -> torch.Tensor:
225

226
        return dispatch_unquantized_gemm()(layer, x, layer.weight, bias)
227
228


229
230
class LinearBase(torch.nn.Module):
    """Base linear layer.
231
232
233
234
235
236
237

    Args:
        input_size: input dimension of the linear layer.
        output_size: output dimension of the linear layer.
        bias: If true, add bias.
        skip_bias_add: If true, skip adding bias but instead return it.
        params_dtype: Data type for the parameters.
238
        quant_config: Quantization configure.
239
        return_bias: If true, return bias together with outputs in forward pass.
240
241
242
243
244
245
246
247
    """

    def __init__(
        self,
        input_size: int,
        output_size: int,
        skip_bias_add: bool = False,
        params_dtype: Optional[torch.dtype] = None,
248
        quant_config: Optional[QuantizationConfig] = None,
249
        prefix: str = "",
250
251
        *,
        return_bias: bool = True,
252
253
254
255
256
257
258
259
260
261
    ):
        super().__init__()

        # Keep input parameters
        self.input_size = input_size
        self.output_size = output_size
        self.skip_bias_add = skip_bias_add
        if params_dtype is None:
            params_dtype = torch.get_default_dtype()
        self.params_dtype = params_dtype
262
        if quant_config is None:
263
264
            self.quant_method: Optional[
                QuantizeMethodBase] = UnquantizedLinearMethod()
265
        else:
266
267
            self.quant_method = quant_config.get_quant_method(self,
                                                              prefix=prefix)
268
        self.return_bias = return_bias
269

270
271
272
    def forward(
        self, x: torch.Tensor
    ) -> Union[torch.Tensor, tuple[torch.Tensor, Optional[Parameter]]]:
273
274
275
276
277
278
279
280
281
282
283
284
285
        raise NotImplementedError


class ReplicatedLinear(LinearBase):
    """Replicated linear layer.

    Args:
        input_size: input dimension of the linear layer.
        output_size: output dimension of the linear layer.
        bias: If true, add bias.
        skip_bias_add: If true, skip adding bias but instead return it.
        params_dtype: Data type for the parameters.
        quant_config: Quantization configure.
286
287
        prefix: The name of the layer in the state dict, including all parents
                        (e.g. model.layers.0.qkv_proj)
288
        return_bias: If true, return bias together with outputs in forward pass.
289
290
    """

291
292
293
294
295
296
297
298
299
300
301
302
    def __init__(
        self,
        input_size: int,
        output_size: int,
        bias: bool = True,
        skip_bias_add: bool = False,
        params_dtype: Optional[torch.dtype] = None,
        quant_config: Optional[QuantizationConfig] = None,
        prefix: str = "",
        *,
        return_bias: bool = True,
    ):
303
304
305
306
307
        super().__init__(input_size,
                         output_size,
                         skip_bias_add,
                         params_dtype,
                         quant_config,
308
309
                         prefix=prefix,
                         return_bias=return_bias)
310

311
312
        # All the linear layer supports quant method.
        assert self.quant_method is not None
313
314
315
316
317
        self.quant_method.create_weights(self,
                                         self.input_size, [self.output_size],
                                         self.input_size,
                                         self.output_size,
                                         self.params_dtype,
318
                                         weight_loader=self.weight_loader)
319

320
321
        if bias:
            self.bias = Parameter(
322
                torch.empty(self.output_size, dtype=self.params_dtype))
323
324
325
326
            set_weight_attrs(self.bias, {
                "output_dim": 0,
                "weight_loader": self.weight_loader,
            })
327
328
329
        else:
            self.register_parameter("bias", None)

330
331
332
    def weight_loader(self, param: Parameter, loaded_weight: torch.Tensor):
        # If the weight on disk does not have a shape, give it one
        # (such scales for AutoFp8).
333
334
335
336
337
338
339
340
341
342
343
        # Special case for GGUF

        is_gguf_weight = getattr(param, "is_gguf_weight", False)
        is_gguf_weight_type = getattr(param, "is_gguf_weight_type", False)
        if is_gguf_weight_type:
            param.weight_type = loaded_weight.item()

        # Materialize GGUF UninitializedParameter
        if is_gguf_weight and isinstance(param, UninitializedParameter):
            param.materialize(loaded_weight.shape, dtype=loaded_weight.dtype)

344
345
346
        if len(loaded_weight.shape) == 0:
            loaded_weight = loaded_weight.reshape(1)

347
348
349
        assert param.size() == loaded_weight.size(), (
            f"Tried to load weights of size {loaded_weight.size()}"
            f"to a parameter of size {param.size()}")
350
351
        param.data.copy_(loaded_weight)

352
353
354
    def forward(
        self, x: torch.Tensor
    ) -> Union[torch.Tensor, tuple[torch.Tensor, Optional[Parameter]]]:
355
        bias = self.bias if not self.skip_bias_add else None
356
        assert self.quant_method is not None
357
        output = self.quant_method.apply(self, x, bias)
358
        output_bias = self.bias if self.skip_bias_add else None
359
360
        if not self.return_bias:
            return output
361
362
        return output, output_bias

363
364
365
366
367
368
    def extra_repr(self) -> str:
        s = f"in_features={self.input_size}"
        s += f", output_features={self.output_size}"
        s += f", bias={self.bias is not None}"
        return s

369

370
class ColumnParallelLinear(LinearBase):
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
    """Linear layer with column parallelism.

    The linear layer is defined as Y = XA + b. A is parallelized along
    its second dimension as A = [A_1, ..., A_p].

    Args:
        input_size: first dimension of matrix A.
        output_size: second dimension of matrix A.
        bias: If true, add bias.
        gather_output: If true, call all-gather on output and make Y available
                       to all GPUs, otherwise, every GPU will have its output
                       which is Y_i = XA_i
        skip_bias_add: This was added to enable performance optimizations where
                       bias can be fused with other element-wise operations. we
                       skip adding bias but instead return it.
        params_dtype: Data type for the parameters.
387
        quant_config: Quantization configure.
James Fleming's avatar
James Fleming committed
388
389
        output_sizes: list of output sizes packed into one output, like for QKV
                       the list would be size 3.
390
391
        prefix: The name of the layer in the state dict, including all parents
                        (e.g. model.layers.0.qkv_proj) 
392
393
    """

394
395
396
397
398
399
400
401
402
403
404
405
406
407
    def __init__(
        self,
        input_size: int,
        output_size: int,
        bias: bool = True,
        gather_output: bool = False,
        skip_bias_add: bool = False,
        params_dtype: Optional[torch.dtype] = None,
        quant_config: Optional[QuantizationConfig] = None,
        output_sizes: Optional[list[int]] = None,
        prefix: str = "",
        *,
        return_bias: bool = True,
    ):
408
        # Divide the weight matrix along the last dimension.
409
410
411
        self.tp_size = get_tensor_model_parallel_world_size()
        self.input_size_per_partition = input_size
        self.output_size_per_partition = divide(output_size, self.tp_size)
412
413
414
415
        self.output_partition_sizes = [self.output_size_per_partition]
        # If QKV or MergedColumn, use output size of each partition.
        if hasattr(self, "output_sizes"):
            self.output_partition_sizes = [
416
                divide(output_size, self.tp_size)
417
418
419
                for output_size in self.output_sizes
            ]

420
421
422
423
424
425
426
        super().__init__(input_size,
                         output_size,
                         skip_bias_add,
                         params_dtype,
                         quant_config,
                         prefix,
                         return_bias=return_bias)
427
428
429

        self.gather_output = gather_output

James Fleming's avatar
James Fleming committed
430
431
        if output_sizes is None:
            output_sizes = [output_size]
432

433
        assert self.quant_method is not None
434
435
        self.quant_method.create_weights(
            layer=self,
436
            input_size_per_partition=self.input_size_per_partition,
437
438
439
440
            output_partition_sizes=self.output_partition_sizes,
            input_size=self.input_size,
            output_size=self.output_size,
            params_dtype=self.params_dtype,
441
442
            weight_loader=(
                self.weight_loader_v2 if self.quant_method.__class__.__name__
443
                in WEIGHT_LOADER_V2_SUPPORTED else self.weight_loader))
444
445
446
447
448
449
450
451
452
453
454
455
456
457
        if bias:
            self.bias = Parameter(
                torch.empty(self.output_size_per_partition,
                            dtype=params_dtype))
            set_weight_attrs(self.bias, {
                "output_dim": 0,
                "weight_loader": self.weight_loader,
            })
        else:
            self.register_parameter("bias", None)

    def weight_loader(self, param: Parameter, loaded_weight: torch.Tensor):
        tp_rank = get_tensor_model_parallel_rank()
        output_dim = getattr(param, "output_dim", None)
458

459
460
461
462
463
464
        is_sharded_weight = getattr(param, "is_sharded_weight", False)
        use_bitsandbytes_4bit = getattr(param, "use_bitsandbytes_4bit", False)
        # bitsandbytes loads the weights of the specific portion
        # no need to narrow
        is_sharded_weight = is_sharded_weight or use_bitsandbytes_4bit

465
466
467
468
469
470
471
472
        # Special case for GGUF
        is_gguf_weight = getattr(param, "is_gguf_weight", False)
        is_gguf_weight_type = getattr(param, "is_gguf_weight_type", False)
        if is_gguf_weight_type:
            param.weight_type = loaded_weight.item()

        # Materialize GGUF UninitializedParameter
        if is_gguf_weight and isinstance(param, UninitializedParameter):
473
474
475
476
477
478
            final_shape = list(loaded_weight.shape)
            if output_dim is not None:
                tp_size = get_tensor_model_parallel_world_size()
                assert final_shape[output_dim] % tp_size == 0
                final_shape[output_dim] = final_shape[output_dim] // tp_size
            param.materialize(final_shape, dtype=loaded_weight.dtype)
479

480
        param_data = param.data
481
        if output_dim is not None and not is_sharded_weight:
482
483
484
485
            shard_size = param_data.shape[output_dim]
            start_idx = tp_rank * shard_size
            loaded_weight = loaded_weight.narrow(output_dim, start_idx,
                                                 shard_size)
486
487
488
489
490

        # Special case for loading scales off disk, which often do not
        # have a shape (such as in the case of AutoFP8).
        if len(loaded_weight.shape) == 0:
            loaded_weight = loaded_weight.reshape(1)
491

492
493
494
        assert param_data.shape == loaded_weight.shape
        param_data.copy_(loaded_weight)

495
    def weight_loader_v2(self, param: Parameter, loaded_weight: torch.Tensor):
496
497
498
499
500
        # Special case for loading scales off disk, which often do not
        # have a shape (such as in the case of AutoFP8).
        if len(loaded_weight.shape) == 0:
            assert loaded_weight.numel() == 1
            loaded_weight = loaded_weight.reshape(1)
501
502
        param.load_column_parallel_weight(loaded_weight=loaded_weight)

503
504
505
    def forward(
        self, input_
    ) -> Union[torch.Tensor, tuple[torch.Tensor, Optional[Parameter]]]:
506
507
508
        bias = self.bias if not self.skip_bias_add else None

        # Matrix multiply.
509
        assert self.quant_method is not None
510
        output_parallel = self.quant_method.apply(self, input_, bias)
511
512
513
514
515
516
        if self.gather_output:
            # All-gather across the partitions.
            output = tensor_model_parallel_all_gather(output_parallel)
        else:
            output = output_parallel
        output_bias = self.bias if self.skip_bias_add else None
517
518
        if not self.return_bias:
            return output
519
520
        return output, output_bias

521
522
523
524
525
526
527
528
    def extra_repr(self) -> str:
        s = f"in_features={self.input_size}"
        s += f", output_features={self.output_size_per_partition}"
        s += f", bias={self.bias is not None}"
        s += f", tp_size={get_tensor_model_parallel_world_size()}"
        s += f", gather_output={self.gather_output}"
        return s

529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547

class MergedColumnParallelLinear(ColumnParallelLinear):
    """Packed linear layers with column parallelism.

    Similar to ColumnParallelLinear, but the weight matrix is concatenated
    along the output dimension. When the weight matrix is loaded, the
    different partitions are sharded separately.

    Args:
        input_size: input dimension of the linear layer.
        output_sizes: list of output dimensions of the linear layer.
        bias: If true, add bias.
        gather_output: If true, call all-gather on output and make the output
                       available to all GPUs, otherwise, every GPU will have
                       its own output.
        skip_bias_add: This was added to enable performance optimizations where
                       bias can be fused with other element-wise operations. we
                       skip adding bias but instead return it.
        params_dtype: Data type for the parameters.
548
        quant_config: Quantization configure.
549
550
        prefix: The name of the layer in the state dict, including all parents
                        (e.g. model.layers.0.qkv_proj)
551
        return_bias: If true, return bias together with outputs in forward pass.
552
553
    """

554
555
556
557
558
559
560
561
562
563
564
565
566
    def __init__(
        self,
        input_size: int,
        output_sizes: list[int],
        bias: bool = True,
        gather_output: bool = False,
        skip_bias_add: bool = False,
        params_dtype: Optional[torch.dtype] = None,
        quant_config: Optional[QuantizationConfig] = None,
        prefix: str = "",
        *,
        return_bias: bool = True,
    ):
567
568
569
        self.output_sizes = output_sizes
        tp_size = get_tensor_model_parallel_world_size()
        assert all(output_size % tp_size == 0 for output_size in output_sizes)
570
571
572
573
574
575
        super().__init__(input_size=input_size,
                         output_size=sum(output_sizes),
                         bias=bias,
                         gather_output=gather_output,
                         skip_bias_add=skip_bias_add,
                         params_dtype=params_dtype,
576
                         quant_config=quant_config,
577
578
                         prefix=prefix,
                         return_bias=return_bias)
579
580
581
582
583

    def weight_loader(self,
                      param: Parameter,
                      loaded_weight: torch.Tensor,
                      loaded_shard_id: Optional[int] = None):
James Fleming's avatar
James Fleming committed
584

585
586
587
588
589
        # Special case for GGUF
        # initialize GGUF param after we know the quantize type
        is_gguf_weight = getattr(param, "is_gguf_weight", False)
        is_gguf_weight_type = getattr(param, "is_gguf_weight_type", False)
        if is_gguf_weight_type:
590
591
592
593
594
595
596
597
            if loaded_shard_id is not None:
                param.data[loaded_shard_id].copy_(loaded_weight)
                param.shard_weight_type[loaded_shard_id] = loaded_weight.item()
            else:
                param.shard_weight_type = {
                    i: loaded_weight.item()
                    for i, _ in enumerate(self.output_sizes)
                }
598
599
            return

600
601
602
603
604
605
606
        if is_gguf_weight:
            tp_size = get_tensor_model_parallel_world_size()
            tp_rank = get_tensor_model_parallel_rank()

            output_dim = getattr(param, "output_dim", None)
            shard_size = loaded_weight.size(output_dim) // tp_size
            start_idx = tp_rank * shard_size
607

608
609
610
611
612
613
614
            if loaded_shard_id is not None:
                loaded_weight = loaded_weight.narrow(output_dim, start_idx,
                                                     shard_size)
                param.shard_id.append(loaded_shard_id)
                param.shard_id_map[loaded_shard_id] = len(param.data_container)
                param.data_container.append(loaded_weight)
                return
615

616
617
        param_data = param.data
        output_dim = getattr(param, "output_dim", None)
618
        # Special case for AQLM codebooks.
James Fleming's avatar
James Fleming committed
619
        is_metadata = getattr(param, "is_metadata", False)
620
621
        # Special case for per-tensor scale to load scalar into fused array.
        needs_scalar_to_array = getattr(param, "needs_scalar_to_array", False)
622

623
        if loaded_shard_id is None:
624
625
            # Loaded weight is already fused on disk (mlp).
            # (e.g., Phi-3's gate_up_proj).
626
            if output_dim is None:
627
                if needs_scalar_to_array:
628
629
                    param_data, loaded_weight = adjust_scalar_to_fused_array(
                        param_data, loaded_weight, 0)
630

631
632
633
634
                assert param_data.shape == loaded_weight.shape
                param_data.copy_(loaded_weight)
                return
            current_shard_offset = 0
635
636
            use_bitsandbytes_4bit = getattr(param, "use_bitsandbytes_4bit",
                                            False)
637
            shard_offsets: list[tuple[int, int, int]] = []
638
639
640
641
642
            for i, output_size in enumerate(self.output_sizes):
                shard_offsets.append((i, current_shard_offset, output_size))
                current_shard_offset += output_size
            packed_dim = getattr(param, "packed_dim", None)
            for shard_id, shard_offset, shard_size in shard_offsets:
643
                # Special case for Quantization.
644
645
646
647
648
                # If quantized, we need to adjust the offset and size to account
                # for the packing.
                if packed_dim == output_dim:
                    shard_size = shard_size // param.pack_factor
                    shard_offset = shard_offset // param.pack_factor
649
                    # Special case for Marlin.
650
651
                    shard_size, shard_offset = adjust_marlin_shard(
                        param, shard_size, shard_offset)
652

653
654
655
                shard_size, shard_offset = adjust_bitblas_shard(
                    param, shard_size, shard_offset)

656
                if use_bitsandbytes_4bit:
657
658
659
660
661
662
663
664
665
                    index = list(itertools.accumulate([0] + self.output_sizes))
                    orig_offsets = {
                        str(i): (index[i], size)
                        for i, size in enumerate(self.output_sizes)
                    }
                    orig_offsets["total"] = (self.output_size, 0)
                    shard_size, shard_offset = adjust_bitsandbytes_4bit_shard(
                        param, orig_offsets, str(shard_id))

666
667
668
669
670
671
672
673
674
675
676
                loaded_weight_shard = loaded_weight.narrow(
                    output_dim, shard_offset, shard_size)
                self.weight_loader(param, loaded_weight_shard, shard_id)
            return

        assert loaded_shard_id < len(self.output_sizes)
        tp_rank = get_tensor_model_parallel_rank()
        tp_size = get_tensor_model_parallel_world_size()
        if output_dim is not None:
            shard_offset = sum(self.output_sizes[:loaded_shard_id]) // tp_size
            shard_size = self.output_sizes[loaded_shard_id] // tp_size
677
            # Special case for quantization.
678
679
680
681
682
683
            # If quantized, we need to adjust the offset and size to account
            # for the packing.
            packed_dim = getattr(param, "packed_dim", None)
            if packed_dim == output_dim:
                shard_size = shard_size // param.pack_factor
                shard_offset = shard_offset // param.pack_factor
684
                # Special case for Marlin.
685
686
                shard_size, shard_offset = adjust_marlin_shard(
                    param, shard_size, shard_offset)
687
688
            shard_size, shard_offset = adjust_bitblas_shard(
                param, shard_size, shard_offset)
689

690
691
            use_bitsandbytes_4bit = getattr(param, "use_bitsandbytes_4bit",
                                            False)
692
693
694
695
696
            is_sharded_weight = getattr(param, "is_sharded_weight", False)
            # bitsandbytes loads the weights of the specific portion
            # no need to narrow
            is_sharded_weight = is_sharded_weight or use_bitsandbytes_4bit

697
            if use_bitsandbytes_4bit:
698
699
700
701
                shard_size = loaded_weight.shape[output_dim]
                shard_offset = loaded_weight.shape[output_dim] * \
                    loaded_shard_id

702
703
704
            param_data = param_data.narrow(output_dim, shard_offset,
                                           shard_size)
            start_idx = tp_rank * shard_size
705
            if not is_sharded_weight:
706
707
                loaded_weight = loaded_weight.narrow(output_dim, start_idx,
                                                     shard_size)
708
        # Special case for AQLM codebooks.
James Fleming's avatar
James Fleming committed
709
710
711
712
713
        elif is_metadata:
            # metadata indicates fixed size concatenated along dim 0
            shard_size = loaded_weight.shape[0]
            shard_offset = loaded_shard_id * shard_size
            param_data = param_data.narrow(0, shard_offset, shard_size)
714

715
716
717
        # Special case for per-tensor scales in fused case.
        elif needs_scalar_to_array:
            param_data, loaded_weight = adjust_scalar_to_fused_array(
718
719
                param_data, loaded_weight, loaded_shard_id)

720
        else:
CHU Tianxiang's avatar
CHU Tianxiang committed
721
722
723
724
725
726
            ignore_warning = getattr(param, "ignore_warning", False)
            if not ignore_warning:
                logger.warning(
                    "Loading a weight without `output_dim` attribute in "
                    "MergedColumnParallelLinear, assume the weight is "
                    "the same for all partitions.")
727

728
729
730
        assert param_data.shape == loaded_weight.shape
        param_data.copy_(loaded_weight)

731
732
733
734
735
736
737
738
739
740
741
742
743
    def _load_fused_module_from_checkpoint(self, param: BasevLLMParameter,
                                           loaded_weight: torch.Tensor):
        """
        Handle special case for models where MLP layers are already
        fused on disk. In this case, we have no shard id. This function
        determmines the shard id by splitting these layers and then calls
        the weight loader using the shard id.

        An example of a model with these fused layers:
        https://huggingface.co/microsoft/Phi-3-mini-4k-instruct
        """

        current_shard_offset = 0
744
        shard_offsets: list[tuple[int, int, int]] = []
745
746
747
748
749
750
751
752
        for i, output_size in enumerate(self.output_sizes):
            shard_offsets.append((i, current_shard_offset, output_size))
            current_shard_offset += output_size

        for shard_id, shard_offset, shard_size in shard_offsets:
            # Special case for Quantization.
            # If quantized, we need to adjust the offset and size to account
            # for the packing.
753
754
            if isinstance(param, (PackedColumnParameter, PackedvLLMParameter
                                  )) and param.packed_dim == param.output_dim:
755
756
                shard_size, shard_offset = \
                    param.adjust_shard_indexes_for_packing(
757
758
759
760
761
762
763
764
765
766
767
768
                    shard_size=shard_size, shard_offset=shard_offset)

            loaded_weight_shard = loaded_weight.narrow(param.output_dim,
                                                       shard_offset,
                                                       shard_size)
            self.weight_loader_v2(param, loaded_weight_shard, shard_id)

    def weight_loader_v2(self,
                         param: BasevLLMParameter,
                         loaded_weight: torch.Tensor,
                         loaded_shard_id: Optional[int] = None):
        if loaded_shard_id is None:
769
770
771
772
            if isinstance(param, PerTensorScaleParameter):
                param.load_merged_column_weight(loaded_weight=loaded_weight,
                                                shard_id=0)
                return
773
            elif type(param) in (RowvLLMParameter, BasevLLMParameter):
774
                param.load_merged_column_weight(loaded_weight=loaded_weight)
775
                return
776
            # TODO: @dsikka - move to parameter.py
777
778
779
780
781
782
            self._load_fused_module_from_checkpoint(param, loaded_weight)
            return

        assert loaded_shard_id < len(self.output_sizes)

        tp_size = get_tensor_model_parallel_world_size()
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800

        if isinstance(param, BlockQuantScaleParameter):
            from vllm.model_executor.layers.quantization.fp8 import (
                Fp8LinearMethod, Fp8MoEMethod)
            assert self.quant_method is not None
            assert isinstance(self.quant_method,
                              (Fp8LinearMethod, Fp8MoEMethod))
            weight_block_size = self.quant_method.quant_config.weight_block_size
            assert weight_block_size is not None
            block_n, _ = weight_block_size[0], weight_block_size[1]
            shard_offset = (
                (sum(self.output_sizes[:loaded_shard_id]) + block_n - 1) //
                block_n) // tp_size
            shard_size = ((self.output_sizes[loaded_shard_id] + block_n - 1) //
                          block_n // tp_size)
        else:
            shard_offset = sum(self.output_sizes[:loaded_shard_id]) // tp_size
            shard_size = self.output_sizes[loaded_shard_id] // tp_size
801
802
803
804
805
806

        param.load_merged_column_weight(loaded_weight=loaded_weight,
                                        shard_id=loaded_shard_id,
                                        shard_offset=shard_offset,
                                        shard_size=shard_size)

807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828

class QKVParallelLinear(ColumnParallelLinear):
    """Linear layers for the attention's QKV transformation.

    Linear layers for the linear transformation of the query, key, and value
    vectors in the attention layer. The weight matrix is concatenated along
    the output dimension. The layer is parallelized along the head dimension.
    When the number of key/value heads is smaller than the number of query
    heads (e.g., multi-query/grouped-query attention), the key/value head may
    be replicated while the query heads are partitioned.

    Args:
        hidden_size: input hidden state size of the transformer.
        head_size: size of each attention head.
        total_num_heads: total number of attention query heads.
        total_num_kv_heads: total number of attention key/value heads. If
                            None, assume total_num_kv_heads = total_num_heads.
        bias: If true, add bias.
        skip_bias_add: This was added to enable performance optimizations where
                       bias can be fused with other element-wise operations. we
                       skip adding bias but instead return it.
        params_dtype: Data type for the parameters.
829
        quant_config: Quantization configure.
830
831
        prefix: The name of the layer in the state dict, including all parents
                        (e.g. model.layers.0.qkv_proj)
832
        return_bias: If true, return bias together with outputs in forward pass.
833
834
    """

835
836
837
838
839
840
841
842
843
844
845
846
847
848
    def __init__(
        self,
        hidden_size: int,
        head_size: int,
        total_num_heads: int,
        total_num_kv_heads: Optional[int] = None,
        bias: bool = True,
        skip_bias_add: bool = False,
        params_dtype: Optional[torch.dtype] = None,
        quant_config: Optional[QuantizationConfig] = None,
        prefix: str = "",
        *,
        return_bias: bool = True,
    ):
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
        self.hidden_size = hidden_size
        self.head_size = head_size
        self.total_num_heads = total_num_heads
        if total_num_kv_heads is None:
            total_num_kv_heads = total_num_heads
        self.total_num_kv_heads = total_num_kv_heads
        # Divide the weight matrix along the last dimension.
        tp_size = get_tensor_model_parallel_world_size()
        self.num_heads = divide(self.total_num_heads, tp_size)
        if tp_size >= self.total_num_kv_heads:
            self.num_kv_heads = 1
            self.num_kv_head_replicas = divide(tp_size,
                                               self.total_num_kv_heads)
        else:
            self.num_kv_heads = divide(self.total_num_kv_heads, tp_size)
            self.num_kv_head_replicas = 1
        input_size = self.hidden_size
        output_size = (self.num_heads +
                       2 * self.num_kv_heads) * tp_size * self.head_size
868
869
870
871
        self.output_sizes = [
            self.num_heads * self.head_size * tp_size,  # q_proj
            self.num_kv_heads * self.head_size * tp_size,  # k_proj
            self.num_kv_heads * self.head_size * tp_size,  # v_proj 
James Fleming's avatar
James Fleming committed
872
873
        ]

874
875
876
877
878
879
        super().__init__(input_size=input_size,
                         output_size=output_size,
                         bias=bias,
                         gather_output=False,
                         skip_bias_add=skip_bias_add,
                         params_dtype=params_dtype,
880
                         quant_config=quant_config,
881
882
                         prefix=prefix,
                         return_bias=return_bias)
883

884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
    def _get_shard_offset_mapping(self, loaded_shard_id: str):
        shard_offset_mapping = {
            "q": 0,
            "k": self.num_heads * self.head_size,
            "v": (self.num_heads + self.num_kv_heads) * self.head_size,
            "total": (self.num_heads + 2 * self.num_kv_heads) * self.head_size
        }
        return shard_offset_mapping.get(loaded_shard_id)

    def _get_shard_size_mapping(self, loaded_shard_id: str):
        shard_size_mapping = {
            "q": self.num_heads * self.head_size,
            "k": self.num_kv_heads * self.head_size,
            "v": self.num_kv_heads * self.head_size,
        }
        return shard_size_mapping.get(loaded_shard_id)

    def _load_fused_module_from_checkpoint(self, param: BasevLLMParameter,
                                           loaded_weight: torch.Tensor):
        """
        Handle special case for models where QKV layers are already 
        fused on disk. In this case, we have no shard id. This function
        determmines the shard id by splitting these layers and then calls
        the weight loader using the shard id.

        An example of a model with these fused layers:
        https://huggingface.co/microsoft/Phi-3-mini-4k-instruct
        """
        shard_offsets = [
            # (shard_id, shard_offset, shard_size)
            ("q", 0, self.total_num_heads * self.head_size),
            ("k", self.total_num_heads * self.head_size,
             self.total_num_kv_heads * self.head_size),
            ("v",
             (self.total_num_heads + self.total_num_kv_heads) * self.head_size,
             self.total_num_kv_heads * self.head_size),
        ]

        for shard_id, shard_offset, shard_size in shard_offsets:
            # Special case for Quantization.
            # If quantized, we need to adjust the offset and size to account
            # for the packing.
926
927
            if isinstance(param, (PackedColumnParameter, PackedvLLMParameter
                                  )) and param.packed_dim == param.output_dim:
928
929
                shard_size, shard_offset = \
                    param.adjust_shard_indexes_for_packing(
930
931
932
933
934
935
936
937
938
939
940
941
                    shard_size=shard_size, shard_offset=shard_offset)

            loaded_weight_shard = loaded_weight.narrow(param.output_dim,
                                                       shard_offset,
                                                       shard_size)
            self.weight_loader_v2(param, loaded_weight_shard, shard_id)

    def weight_loader_v2(self,
                         param: BasevLLMParameter,
                         loaded_weight: torch.Tensor,
                         loaded_shard_id: Optional[str] = None):
        if loaded_shard_id is None:  # special case for certain models
942
            if isinstance(param, PerTensorScaleParameter):
943
                param.load_qkv_weight(loaded_weight=loaded_weight, shard_id=0)
944
                return
945
946
            elif type(param) in (RowvLLMParameter, BasevLLMParameter):
                param.load_qkv_weight(loaded_weight=loaded_weight)
947
                return
948
            # TODO: @dsikka - move to parameter.py
949
950
951
952
953
954
955
956
            self._load_fused_module_from_checkpoint(param, loaded_weight)
            return

        assert loaded_shard_id in ["q", "k", "v"]

        shard_offset = self._get_shard_offset_mapping(loaded_shard_id)
        shard_size = self._get_shard_size_mapping(loaded_shard_id)

957
958
959
960
961
962
963
964
965
        # Note(simon): This is needed for Qwen3's fp8 quantization.
        if isinstance(param, BlockQuantScaleParameter):
            assert self.quant_method is not None
            assert hasattr(self.quant_method, "quant_config")
            weight_block_size = self.quant_method.quant_config.weight_block_size
            block_n, _ = weight_block_size[0], weight_block_size[1]
            shard_offset = (shard_offset + block_n - 1) // block_n
            shard_size = (shard_size + block_n - 1) // block_n

966
967
968
969
970
971
        param.load_qkv_weight(loaded_weight=loaded_weight,
                              num_heads=self.num_kv_head_replicas,
                              shard_id=loaded_shard_id,
                              shard_offset=shard_offset,
                              shard_size=shard_size)

972
973
974
975
    def weight_loader(self,
                      param: Parameter,
                      loaded_weight: torch.Tensor,
                      loaded_shard_id: Optional[str] = None):
976
977
978
979
980

        # Special case for GGUF
        # initialize GGUF param after we know the quantize type
        is_gguf_weight = getattr(param, "is_gguf_weight", False)
        is_gguf_weight_type = getattr(param, "is_gguf_weight_type", False)
981
        if is_gguf_weight_type:
982
            idx_map = {"q": 0, "k": 1, "v": 2}
983
984
985
986
987
988
989
990
            if loaded_shard_id is not None:
                param.data[idx_map[loaded_shard_id]].copy_(loaded_weight)
                param.shard_weight_type[loaded_shard_id] = loaded_weight.item()
            else:
                param.shard_weight_type = {
                    k: loaded_weight.item()
                    for k in idx_map
                }
991
992
            return

993
994
995
996
997
998
999
1000
        if is_gguf_weight:
            tp_size = get_tensor_model_parallel_world_size()
            tp_rank = get_tensor_model_parallel_rank()

            output_dim = getattr(param, "output_dim", None)
            shard_size = loaded_weight.size(output_dim) // tp_size
            start_idx = tp_rank * shard_size

1001
1002
1003
1004
1005
1006
1007
            if loaded_shard_id is not None:
                loaded_weight = loaded_weight.narrow(output_dim, start_idx,
                                                     shard_size)
                param.shard_id.append(loaded_shard_id)
                param.shard_id_map[loaded_shard_id] = len(param.data_container)
                param.data_container.append(loaded_weight)
                return
1008

1009
1010
        param_data = param.data
        output_dim = getattr(param, "output_dim", None)
1011
        # Special case for AQLM codebooks.
James Fleming's avatar
James Fleming committed
1012
        is_metadata = getattr(param, "is_metadata", False)
1013

1014
1015
        # Special case for per-tensor scales in fused case.
        needs_scalar_to_array = getattr(param, "needs_scalar_to_array", False)
1016

1017
        if loaded_shard_id is None:
1018
1019
            # Loaded weight is already fused on disk (qkv).
            # (e.g., Phi-3's qkv_proj).
1020
            if output_dim is None:
1021
                if needs_scalar_to_array:
1022
1023
                    param_data, loaded_weight = adjust_scalar_to_fused_array(
                        param_data, loaded_weight, 0)
1024

1025
1026
1027
1028
1029
1030
1031
1032
1033
1034
1035
                assert param_data.shape == loaded_weight.shape
                param_data.copy_(loaded_weight)
                return
            shard_offsets = [
                # (shard_id, shard_offset, shard_size)
                ("q", 0, self.total_num_heads * self.head_size),
                ("k", self.total_num_heads * self.head_size,
                 self.total_num_kv_heads * self.head_size),
                ("v", (self.total_num_heads + self.total_num_kv_heads) *
                 self.head_size, self.total_num_kv_heads * self.head_size),
            ]
1036
1037
1038
            use_bitsandbytes_4bit = getattr(param, "use_bitsandbytes_4bit",
                                            False)

1039
1040
            packed_dim = getattr(param, "packed_dim", None)
            for shard_id, shard_offset, shard_size in shard_offsets:
1041
                # Special case for Quantized Weights.
1042
1043
1044
1045
1046
                # If quantized, we need to adjust the offset and size to account
                # for the packing.
                if packed_dim == output_dim:
                    shard_size = shard_size // param.pack_factor
                    shard_offset = shard_offset // param.pack_factor
1047

1048
                    # Special case for Marlin.
1049
1050
1051
                    shard_size, shard_offset = adjust_marlin_shard(
                        param, shard_size, shard_offset)

1052
1053
1054
1055
1056
1057
1058
1059
1060
1061
1062
1063
1064
1065
1066
1067
1068
                if use_bitsandbytes_4bit:
                    orig_qkv_offsets = {
                        "q": (0, self.total_num_heads * self.head_size),
                        "k": (self.total_num_heads * self.head_size,
                              self.total_num_kv_heads * self.head_size),
                        "v":
                        ((self.total_num_heads + self.total_num_kv_heads) *
                         self.head_size,
                         self.total_num_kv_heads * self.head_size),
                        "total":
                        ((self.total_num_heads + 2 * self.total_num_kv_heads) *
                         self.head_size, 0)
                    }

                    shard_size, shard_offset = adjust_bitsandbytes_4bit_shard(
                        param, orig_qkv_offsets, shard_id)

1069
1070
1071
1072
1073
1074
1075
                loaded_weight_shard = loaded_weight.narrow(
                    output_dim, shard_offset, shard_size)
                self.weight_loader(param, loaded_weight_shard, shard_id)
            return

        tp_rank = get_tensor_model_parallel_rank()
        assert loaded_shard_id in ["q", "k", "v"]
1076
1077

        # If output dim is defined, use the default loading process.
1078
1079
1080
1081
1082
1083
1084
1085
1086
1087
1088
        if output_dim is not None:
            if loaded_shard_id == "q":
                shard_offset = 0
                shard_size = self.num_heads * self.head_size
            elif loaded_shard_id == "k":
                shard_offset = self.num_heads * self.head_size
                shard_size = self.num_kv_heads * self.head_size
            elif loaded_shard_id == "v":
                shard_offset = (self.num_heads +
                                self.num_kv_heads) * self.head_size
                shard_size = self.num_kv_heads * self.head_size
1089
            # Special case for Quantized Weights.
1090
1091
1092
1093
1094
1095
            # If quantized, we need to adjust the offset and size to account
            # for the packing.
            packed_dim = getattr(param, "packed_dim", None)
            if packed_dim == output_dim:
                shard_size = shard_size // param.pack_factor
                shard_offset = shard_offset // param.pack_factor
1096

1097
                # Special case for Marlin.
1098
1099
1100
                shard_size, shard_offset = adjust_marlin_shard(
                    param, shard_size, shard_offset)

1101
1102
            use_bitsandbytes_4bit = getattr(param, "use_bitsandbytes_4bit",
                                            False)
1103
1104
1105
1106
1107
            is_sharded_weight = getattr(param, "is_sharded_weight", False)
            # bitsandbytes loads the weights of the specific portion
            # no need to narrow
            is_sharded_weight = is_sharded_weight or use_bitsandbytes_4bit

1108
            if use_bitsandbytes_4bit:
1109
1110
1111
1112
1113
1114
1115
1116
1117
1118
1119
                orig_qkv_offsets = {
                    "q": (0, self.num_heads * self.head_size),
                    "k": (self.num_heads * self.head_size,
                          self.num_kv_heads * self.head_size),
                    "v":
                    ((self.num_heads + self.num_kv_heads) * self.head_size,
                     self.num_kv_heads * self.head_size),
                    "total":
                    ((self.num_heads + 2 * self.num_kv_heads) * self.head_size,
                     0)
                }
1120
                shard_size, shard_offset = adjust_bitsandbytes_4bit_shard(
1121
1122
                    param, orig_qkv_offsets, loaded_shard_id)

1123
1124
            param_data = param_data.narrow(output_dim, shard_offset,
                                           shard_size)
1125
1126
1127
1128
            if loaded_shard_id == "q":
                shard_id = tp_rank
            else:
                shard_id = tp_rank // self.num_kv_head_replicas
1129
            start_idx = shard_id * shard_size
1130

1131
            if not is_sharded_weight:
1132
1133
1134
                loaded_weight = loaded_weight.narrow(output_dim, start_idx,
                                                     shard_size)

1135
        # Special case for for AQLM codebooks.
James Fleming's avatar
James Fleming committed
1136
1137
1138
1139
1140
1141
        elif is_metadata:
            # metadata indicates fixed size concatenated along dim 0
            shard_size = loaded_weight.shape[0]
            shard_index = ["q", "k", "v"].index(loaded_shard_id)
            param_data = param_data.narrow(0, shard_index * shard_size,
                                           shard_size)
1142
1143
1144
        # Special case for per-tensor scales in fused case.
        elif needs_scalar_to_array:
            param_data, loaded_weight = adjust_scalar_to_fused_array(
1145
                param_data, loaded_weight, loaded_shard_id)
1146
        else:
CHU Tianxiang's avatar
CHU Tianxiang committed
1147
1148
1149
1150
1151
1152
            ignore_warning = getattr(param, "ignore_warning", False)
            if not ignore_warning:
                logger.warning(
                    "Loading a weight without `output_dim` attribute in "
                    "QKVParallelLinear, assume the weight is the same "
                    "for all partitions.")
1153

1154
1155
1156
1157
        assert param_data.shape == loaded_weight.shape
        param_data.copy_(loaded_weight)


1158
class RowParallelLinear(LinearBase):
1159
1160
1161
1162
1163
1164
1165
1166
1167
1168
1169
1170
1171
1172
1173
1174
1175
1176
1177
1178
1179
1180
    """Linear layer with row parallelism.

    The linear layer is defined as Y = XA + b. A is parallelized along
    its first dimension and X along its second dimension as:
               -   -
              | A_1 |
              | .   |
          A = | .   |        X = [X_1, ..., X_p]
              | .   |
              | A_p |
               -   -
    Arguments:
        input_size: first dimension of matrix A.
        output_size: second dimension of matrix A.
        bias: If true, add bias. Note that bias is not parallelized.
        input_is_parallel: If true, we assume that the input is already
                           split across the GPUs and we do not split
                           again.
        skip_bias_add: This was added to enable performance optimization where
                       bias can be fused with other element-wise operations.
                       We skip adding bias but instead return it.
        params_dtype: Data type for the parameters.
1181
1182
1183
        reduce_results: If true, call all-reduce on output and make Y available
                       to all GPUs, otherwise, every GPU will have its output
                       which is Y = X_iA_i
1184
        quant_config: Quantization configure.
1185
1186
1187
        prefix: The name of the layer in the state dict, including all parents
                        (e.g. model.layers.0.down_proj)
        return_bias: If true, return bias together with outputs in forward pass.
1188
1189
    """

1190
1191
1192
1193
1194
1195
1196
1197
1198
1199
1200
1201
1202
1203
    def __init__(
        self,
        input_size: int,
        output_size: int,
        bias: bool = True,
        input_is_parallel: bool = True,
        skip_bias_add: bool = False,
        params_dtype: Optional[torch.dtype] = None,
        reduce_results: bool = True,
        quant_config: Optional[QuantizationConfig] = None,
        prefix: str = "",
        *,
        return_bias: bool = True,
    ):
1204
1205
1206
1207
1208
1209
1210
        # Divide the weight matrix along the first dimension.
        self.tp_rank = get_tensor_model_parallel_rank()
        self.tp_size = get_tensor_model_parallel_world_size()
        self.input_size_per_partition = divide(input_size, self.tp_size)
        self.output_size_per_partition = output_size
        self.output_partition_sizes = [output_size]

1211
1212
1213
1214
1215
1216
1217
        super().__init__(input_size,
                         output_size,
                         skip_bias_add,
                         params_dtype,
                         quant_config,
                         prefix,
                         return_bias=return_bias)
1218

1219
1220
1221
        self.input_is_parallel = input_is_parallel
        self.reduce_results = reduce_results

1222
        assert self.quant_method is not None
1223
1224
1225
        self.quant_method.create_weights(
            layer=self,
            input_size_per_partition=self.input_size_per_partition,
1226
            output_partition_sizes=self.output_partition_sizes,
1227
1228
1229
            input_size=self.input_size,
            output_size=self.output_size,
            params_dtype=self.params_dtype,
1230
1231
            weight_loader=(
                self.weight_loader_v2 if self.quant_method.__class__.__name__
1232
                in WEIGHT_LOADER_V2_SUPPORTED else self.weight_loader))
1233
1234
1235
1236
1237
1238
        if not reduce_results and (bias and not skip_bias_add):
            raise ValueError("When not reduce the results, adding bias to the "
                             "results can lead to incorrect results")

        if bias:
            self.bias = Parameter(
1239
                torch.empty(self.output_size, dtype=params_dtype))
1240
1241
1242
1243
1244
1245
1246
1247
1248
            set_weight_attrs(self.bias, {
                "output_dim": 0,
                "weight_loader": self.weight_loader,
            })
        else:
            self.register_parameter("bias", None)

    def weight_loader(self, param: Parameter, loaded_weight: torch.Tensor):
        tp_rank = get_tensor_model_parallel_rank()
1249
        tp_size = get_tensor_model_parallel_world_size()
1250
        input_dim = getattr(param, "input_dim", None)
1251
        use_bitsandbytes_4bit = getattr(param, "use_bitsandbytes_4bit", False)
1252
1253
1254
1255
        is_sharded_weight = getattr(param, "is_sharded_weight", False)
        # bitsandbytes loads the weights of the specific portion
        # no need to narrow
        is_sharded_weight = is_sharded_weight or use_bitsandbytes_4bit
1256
1257
1258
1259
1260
1261
1262
1263
1264

        # Special case for GGUF
        is_gguf_weight = getattr(param, "is_gguf_weight", False)
        is_gguf_weight_type = getattr(param, "is_gguf_weight_type", False)
        if is_gguf_weight_type:
            param.weight_type = loaded_weight.item()

        # Materialize GGUF UninitializedParameter
        if is_gguf_weight and isinstance(param, UninitializedParameter):
1265
1266
1267
1268
            weight_shape = list(loaded_weight.shape)
            if input_dim:
                weight_shape[input_dim] = weight_shape[input_dim] // tp_size
            param.materialize(tuple(weight_shape), dtype=loaded_weight.dtype)
1269

1270
        param_data = param.data
1271
        if input_dim is not None and not is_sharded_weight:
1272
1273
1274
1275
            shard_size = param_data.shape[input_dim]
            start_idx = tp_rank * shard_size
            loaded_weight = loaded_weight.narrow(input_dim, start_idx,
                                                 shard_size)
1276

1277
1278
1279
        # Special case for loading scales off disk, which often do not
        # have a shape (such as in the case of AutoFP8).
        if len(loaded_weight.shape) == 0:
1280
1281
            loaded_weight = loaded_weight.reshape(1)

1282
1283
1284
        assert param_data.shape == loaded_weight.shape
        param_data.copy_(loaded_weight)

1285
1286
    def weight_loader_v2(self, param: BasevLLMParameter,
                         loaded_weight: torch.Tensor):
1287
1288
1289
1290
1291
1292
1293

        # Special case for loading scales off disk, which often do not
        # have a shape (such as in the case of AutoFP8).
        if len(loaded_weight.shape) == 0:
            assert loaded_weight.numel() == 1
            loaded_weight = loaded_weight.reshape(1)

1294
1295
        param.load_row_parallel_weight(loaded_weight=loaded_weight)

1296
1297
1298
    def forward(
        self, input_
    ) -> Union[torch.Tensor, tuple[torch.Tensor, Optional[Parameter]]]:
1299
1300
1301
1302
1303
1304
1305
1306
1307
        if self.input_is_parallel:
            input_parallel = input_
        else:
            tp_rank = get_tensor_model_parallel_rank()
            splitted_input = split_tensor_along_last_dim(
                input_, num_partitions=self.tp_size)
            input_parallel = splitted_input[tp_rank].contiguous()

        # Matrix multiply.
1308
        assert self.quant_method is not None
1309
1310
1311
1312
1313
1314
        # Only fuse bias add into GEMM for rank 0 (this ensures that
        # bias will not get added more than once in TP>1 case)
        bias_ = None if (self.tp_rank > 0 or self.skip_bias_add) else self.bias
        output_parallel = self.quant_method.apply(self,
                                                  input_parallel,
                                                  bias=bias_)
1315
        if self.reduce_results and self.tp_size > 1:
1316
            output = tensor_model_parallel_all_reduce(output_parallel)
1317
        else:
1318
1319
1320
            output = output_parallel

        output_bias = self.bias if self.skip_bias_add else None
1321

1322
1323
        if not self.return_bias:
            return output
1324
        return output, output_bias
1325
1326
1327
1328
1329
1330
1331
1332

    def extra_repr(self) -> str:
        s = f"input_features={self.input_size_per_partition}"
        s += f", output_features={self.output_size}"
        s += f", bias={self.bias is not None}"
        s += f", tp_size={self.tp_size}"
        s += f", reduce_results={self.reduce_results}"
        return s
1333
1334


1335
1336
1337
1338
1339
1340
1341
1342
1343
1344
1345
1346
1347
1348
1349
1350
1351
1352
class QKVCrossParallelLinear(LinearBase):
    """Linear layers for efficient cross-attention's QKV transformation.

    Args:
        hidden_size: input hidden state size of the transformer.
        head_size: size of each attention head.
        total_num_heads: total number of attention query heads.
        total_num_kv_heads: total number of attention key/value heads. If
                            None, assume total_num_kv_heads = total_num_heads.
        bias: If true, add bias.
        skip_bias_add: This was added to enable performance optimizations where
                       bias can be fused with other element-wise operations. we
                       skip adding bias but instead return it.
        params_dtype: Data type for the parameters.
        quant_config: Quantization configure.
        prefix: The name of the layer in the state dict, including all parents
                        (e.g. model.layers.0.qkv_proj)
    """
1353
1354
1355
1356
1357
1358
1359
1360
1361
1362
1363

    def __init__(self,
                 hidden_size: int,
                 head_size: int,
                 total_num_heads: int,
                 total_num_kv_heads: Optional[int] = None,
                 bias: bool = True,
                 skip_bias_add: bool = False,
                 params_dtype: Optional[torch.dtype] = None,
                 quant_config: Optional[QuantizationConfig] = None,
                 prefix: str = ""):
1364
1365
1366
1367
1368
1369
1370
1371
1372
1373
1374
1375
        # input_size and output_size are not used, just for alignment
        input_size = hidden_size
        output_size = (total_num_heads + (total_num_kv_heads or 0)) * head_size
        super().__init__(input_size=input_size,
                         output_size=output_size,
                         skip_bias_add=skip_bias_add,
                         params_dtype=params_dtype,
                         quant_config=quant_config,
                         prefix=prefix)

        self.quant_config = quant_config

1376
        # Empty placeholders for loading as a single module.
1377
1378
1379
1380
1381
1382
1383
1384
1385
        placeholder_size = 0
        assert self.quant_method is not None
        self.quant_method.create_weights(self,
                                         placeholder_size, [placeholder_size],
                                         placeholder_size,
                                         placeholder_size,
                                         self.params_dtype,
                                         weight_loader=self.weight_loader)

1386
1387
1388
1389
1390
1391
1392
1393
1394
1395
1396
1397
1398
1399
1400
1401
1402
1403
1404
1405
1406
1407
1408
1409
        # Use a dictionary to avoid submodules parameters auto-registration:
        # drop-in replacement for a `QKVParallelLinear` module.
        self.proj = dict()
        self.proj["q_proj_decoder"] = ColumnParallelLinear(
            input_size=hidden_size,
            output_size=total_num_heads * head_size,
            bias=bias,
            quant_config=quant_config,
            skip_bias_add=skip_bias_add,
            params_dtype=params_dtype,
            prefix=f"{prefix}.q_proj_decoder")

        self.proj["kv_proj_encoder"] = QKVParallelLinear(
            hidden_size=hidden_size,
            head_size=head_size,
            total_num_heads=0,
            total_num_kv_heads=total_num_kv_heads,
            bias=bias,
            quant_config=quant_config,
            skip_bias_add=skip_bias_add,
            params_dtype=params_dtype,
            prefix=f"{prefix}.kv_proj_encoder")

        # `kv_proj_encoder.num_kv_heads` accounts for sharding with tp>1.
1410
        self.q_size = self.q_proj_decoder.output_size_per_partition
1411
1412
1413
1414
1415
        self.kv_size = self.kv_proj_encoder.num_kv_heads * head_size

        if bias:
            self.bias = torch.nn.Parameter()
            set_weight_attrs(self.bias, {
1416
1417
                "output_dim": 0,
                "weight_loader": self.weight_loader,
1418
            })
1419
1420
        else:
            self.bias = None
1421

1422
1423
1424
1425
1426
    def process_weights_after_loading(self):
        for layer in self.proj.values():
            if self.quant_method is not None:
                self.quant_method.process_weights_after_loading(layer)

1427
    @property
1428
1429
1430
    def q_proj_decoder(self) -> ColumnParallelLinear:
        layer = self.proj["q_proj_decoder"]
        for name, param in self.named_parameters():
1431
1432
1433
1434
1435
            target_param = getattr(layer, name, None)
            if target_param is not None:
                self.sync_weight_attrs(param,
                                       target_param,
                                       mode="q_proj_decoder")
1436
        return layer
1437
1438

    @property
1439
1440
1441
    def kv_proj_encoder(self) -> QKVParallelLinear:
        layer = self.proj["kv_proj_encoder"]
        for name, param in self.named_parameters():
1442
1443
1444
1445
1446
            target_param = getattr(layer, name, None)
            if target_param is not None:
                self.sync_weight_attrs(param,
                                       target_param,
                                       mode="kv_proj_encoder")
1447
1448
1449
1450
1451
1452
1453
1454
1455
1456
        return layer

    def sync_weight_attrs(
        self,
        src_param: nn.Parameter,
        tgt_param: nn.Parameter,
        mode: Literal["q_proj_decoder", "kv_proj_encoder"],
    ):
        missing_attrs_dict = {
            k: getattr(src_param, k)
1457
1458
            for k in (set(vars(src_param).keys()) -
                      set(vars(tgt_param).keys()))
1459
1460
1461
1462
1463
1464
1465
1466
1467
1468
1469
1470
1471
        }
        # TODO(Isotr0py): handle bitsandbytes 8bit
        use_bitsandbytes_4bit = getattr(src_param, "use_bitsandbytes_4bit",
                                        False)
        if (missing_attrs_dict and use_bitsandbytes_4bit):
            q_proj_attrs, kv_proj_attrs = left_shift_bitsandbytes_4bit_shard(
                missing_attrs_dict)
            if mode == "q_proj_decoder":
                set_weight_attrs(tgt_param, q_proj_attrs)
            elif mode == "kv_proj_encoder":
                set_weight_attrs(tgt_param, kv_proj_attrs)
        else:
            set_weight_attrs(tgt_param, missing_attrs_dict)
1472

1473
1474
1475
1476
1477
1478
1479
1480
1481
1482
1483
1484
1485
1486
1487
1488
1489
1490
1491
1492
1493
1494
1495
1496
1497
1498
1499
1500
1501
1502
1503
1504
1505
1506
1507
1508
1509
1510
1511
1512
1513
1514
    def _is_same_param(
        self,
        src_param: torch.nn.Parameter,
        map_param: torch.nn.Parameter,
    ) -> bool:
        """Check if two parameters are exactly pointing to same things."""
        # ignore weight_loader because it's always different
        key_to_ignore = ["weight_loader", "_weight_loader"]
        has_same_type_name = type(src_param) is type(map_param)
        src_param_attrs = {
            k: v
            for k, v in src_param.__dict__.items() if k not in key_to_ignore
        }
        map_param_attrs = {
            k: v
            for k, v in map_param.__dict__.items() if k not in key_to_ignore
        }
        has_same_attrs = src_param_attrs == map_param_attrs
        return has_same_type_name and has_same_attrs

    def select_proj_params(
        self,
        layer: nn.Module,
        param: nn.Parameter,
    ) -> nn.Parameter:
        """
        Given the placeholder param, 
        return the corresponding param in the proj layers.
        """
        target_param_list = [
            v for _, v in layer.named_parameters()
            if self._is_same_param(param, v)
        ]
        assert len(target_param_list) == 1
        target_param = target_param_list[0]
        return target_param

    def forward(  # type: ignore[override]
        self,
        decoder_hidden_states: torch.Tensor,
        encoder_hidden_states: torch.Tensor,
    ) -> tuple[torch.Tensor, ...]:
1515
1516
1517
1518
1519
1520
1521
1522
1523
1524
1525
1526
        q, _ = self.q_proj_decoder(decoder_hidden_states)
        if encoder_hidden_states is None:
            # Encoder KV already cached.
            k = None
            v = None
        else:
            # Prefill phase, encoder KV cached here.
            kv_enc, _ = self.kv_proj_encoder(encoder_hidden_states)
            # Split kv in half
            k, v = kv_enc.split(self.kv_size, dim=-1)
        return q, k, v

1527
1528
1529
1530
1531
1532
1533
1534
    def weight_loader(self,
                      param: torch.nn.Parameter,
                      loaded_weight: torch.Tensor,
                      loaded_shard_id: Optional[str] = None):
        layer = (self.q_proj_decoder
                 if loaded_shard_id == "q" else self.kv_proj_encoder)
        target_param = self.select_proj_params(layer, param)
        shard_id_args = (loaded_shard_id, ) if loaded_shard_id != "q" else ()
1535
1536
1537
1538
        if self.quant_method.__class__.__name__ in WEIGHT_LOADER_V2_SUPPORTED:
            layer.weight_loader_v2(target_param, loaded_weight, *shard_id_args)
        else:
            layer.weight_loader(target_param, loaded_weight, *shard_id_args)
1539
1540
1541

    def extra_repr(self) -> str:
        s = f"in_features={self.input_size}"
1542
        s += f", q_size={self.q_size}"
1543
1544
1545
1546
1547
        s += f", kv_size={self.kv_size}"
        s += f", bias={self.bias is not None}"
        s += f", tp_size={get_tensor_model_parallel_world_size()}"
        s += ", gather_output=False"
        return s