linear.py 62.1 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

7
from typing import Any
8
from vllm import envs
9
10

import torch
11
from torch.nn.parameter import Parameter, UninitializedParameter
12

13
14
15
16
17
18
19
20
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,
)
21
from vllm.logger import init_logger
22
from vllm.model_executor.custom_op import CustomOp
23
from vllm.model_executor.layers.quantization.base_config import (
24
25
26
    QuantizationConfig,
    QuantizeMethodBase,
)
27
from vllm.model_executor.layers.utils import dispatch_unquantized_gemm
28
29
30
31
32
33
34
35
36
from vllm.model_executor.parameter import (
    BasevLLMParameter,
    BlockQuantScaleParameter,
    ModelWeightParameter,
    PackedColumnParameter,
    PackedvLLMParameter,
    PerTensorScaleParameter,
    RowvLLMParameter,
)
37
from vllm.model_executor.utils import set_weight_attrs
38
from vllm.platforms import current_platform
gaoqiong's avatar
gaoqiong committed
39

zhuwenwen's avatar
zhuwenwen committed
40
import os
41
from vllm.model_executor.utils import gemm_bank_conf
42
43
44

logger = init_logger(__name__)

45
WEIGHT_LOADER_V2_SUPPORTED = [
46
    "UnquantizedLinearMethod",
47
    "CompressedTensorsLinearMethod",
48
    "CompressedTensorsLinearTransformMethod",
49
50
    "BitBLASLinearMethod",
    "GPTQBitBLASLinearMethod",
51
52
53
54
55
56
57
58
59
60
    "AWQMarlinLinearMethod",
    "AWQLinearMethod",
    "GPTQMarlinLinearMethod",
    "Fp8LinearMethod",
    "MarlinLinearMethod",
    "GPTQMarlin24LinearMethod",
    "TPUInt8LinearMethod",
    "GPTQLinearMethod",
    "FBGEMMFp8LinearMethod",
    "ModelOptFp8LinearMethod",
61
62
    "ModelOptFp8PcPtLinearMethod",
    "ModelOptFp8PbWoLinearMethod",
63
64
65
66
    "IPEXAWQLinearMethod",
    "IPEXGPTQLinearMethod",
    "QuarkLinearMethod",
    "ModelOptNvFp4LinearMethod",
67
    "PetitNvFp4LinearMethod",
zhuwenwen's avatar
zhuwenwen committed
68
    "BlockInt8LinearMethod",
69
]
70

71

72
73
74
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:
75
        return (shard_size // bitblas_tile_size, shard_offset // bitblas_tile_size)
76
77
78
79

    return shard_size, shard_offset


80
81
82
83
84
85
86
87
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


88
89
90
91
92
93
94
95
def adjust_block_scale_shard(weight_block_size, shard_size, shard_offset):
    assert weight_block_size is not None
    block_n = weight_block_size[0]
    shard_offset = (shard_offset + block_n - 1) // block_n
    shard_size = (shard_size + block_n - 1) // block_n
    return shard_size, shard_offset


96
97
98
def adjust_bitsandbytes_4bit_shard(
    param: Parameter, shard_offsets: dict[str, tuple[int, int]], loaded_shard_id: str
) -> tuple[int, int]:
99
100
    """Adjust the quantization offsets and sizes for BitsAndBytes sharding."""

101
102
    total, _ = shard_offsets["total"]
    orig_offset, orig_size = shard_offsets[loaded_shard_id]
103
104
105
106
107
108
109
110

    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


111
112
113
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
114
115
    is an array of length N. The loaded_weight corresponds to
    one of the shards on disk. Here, we slice the param based on
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
    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]

131
132
133
134
    if envs.VLLM_USE_NN:
        return param[shard_id], loaded_weight.t()
    else:
        return param[shard_id], loaded_weight
135
136


137
138
139
140
141
142
143
144
# 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:
    {
145
        'bnb_shard_offsets': array([0, 4, 8, 16]),
146
147
148
149
150
        'bnb_quant_state': {0: ..., 1: ..., 2: ...},
    }

    The function will return:
    {
151
        'bnb_shard_offsets': array([0, 4]),
152
153
154
155
156
157
158
159
160
161
162
163
164
165
        '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]
166
        for i in range(1, len(shard_offsets) - 1)
167
168
169
170
171
172
    }
    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


173
class LinearMethodBase(QuantizeMethodBase):
174
175
176
    """Base class for different (maybe quantized) linear methods."""

    @abstractmethod
177
178
179
180
181
182
183
184
185
186
187
    def create_weights(
        self,
        layer: torch.nn.Module,
        input_size_per_partition: int,
        output_partition_sizes: list[int],
        input_size: int,
        output_size: int,
        params_dtype: torch.dtype,
        **extra_weight_attrs,
    ):
        """Create weights for a linear layer.
188
           The weights will be set as attributes of the layer.
189

190
191
192
        Args:
            layer: The layer that is using the LinearMethodBase factory.
            input_size_per_partition: Size of the weight input dim on rank X.
193
            output_partition_sizes: Sizes of the output dim of each logical
194
195
196
197
198
199
                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.
        """
200
201
202
        raise NotImplementedError

    @abstractmethod
203
204
205
206
    def apply(
        self,
        layer: torch.nn.Module,
        x: torch.Tensor,
207
        bias: torch.Tensor | None = None,
208
    ) -> torch.Tensor:
209
210
        """Apply the weights in layer to the input tensor.
        Expects create_weights to have been called before on the layer."""
211
212
213
214
        raise NotImplementedError


class UnquantizedLinearMethod(LinearMethodBase):
215
    """Linear method without quantization."""
216
217
    
    def __init__(self):
zhuwenwen's avatar
zhuwenwen committed
218
        self.use_llama_nn = os.environ.get('LLAMA_NN') == '1'
219
        self.use_gemm_pad = os.environ.get('GEMM_PAD') == '1'
gaoqiong's avatar
gaoqiong committed
220
        
221

222

223
224
225
226
227
228
229
230
231
232
    def create_weights(
        self,
        layer: torch.nn.Module,
        input_size_per_partition: int,
        output_partition_sizes: list[int],
        input_size: int,
        output_size: int,
        params_dtype: torch.dtype,
        **extra_weight_attrs,
    ):
233
234
235
236
        # This method creates unquantized linear weights.
        # The weights are not quantized, and they are not sharded.
        # The amount of memory allocated for the weights is
        # sum(output_partition_sizes) * input_size_per_partition.
237
        weight_loader = extra_weight_attrs.pop("weight_loader")
238
239
240
        if envs.VLLM_USE_NN:
            weight = ModelWeightParameter(
                data=torch.empty(
241
242
                    input_size_per_partition,
                    sum(output_partition_sizes),
243
244
245
246
247
248
249
250
251
                    dtype=params_dtype,
                ),
                input_dim=1,
                output_dim=0,
                weight_loader=weight_loader,
            )
        else:
            weight = ModelWeightParameter(
                data=torch.empty(
252
253
                    sum(output_partition_sizes),
                    input_size_per_partition,
254
255
256
257
258
259
                    dtype=params_dtype,
                ),
                input_dim=1,
                output_dim=0,
                weight_loader=weight_loader,
            )
260

261
262
        layer.register_parameter("weight", weight)
        set_weight_attrs(weight, extra_weight_attrs)
263

264
    def process_weights_after_loading(self, layer: torch.nn.Module) -> None:
265
        if current_platform.is_cpu():
266
            from vllm.model_executor.layers.utils import dispatch_cpu_unquantized_gemm
267

268
            dispatch_cpu_unquantized_gemm(layer, remove_weight=True)
269

270
271
272
273
    def apply(
        self,
        layer: torch.nn.Module,
        x: torch.Tensor,
274
        bias: torch.Tensor | None = None,
275
    ) -> torch.Tensor:
zhuwenwen's avatar
zhuwenwen committed
276
        if self.use_llama_nn:
zhuwenwen's avatar
zhuwenwen committed
277
278
            # if os.environ['GEMM_PAD'] == '1' and gemm_bank_conf(layer.weight.shape[1] - 32):
            #     layer.weight = layer.weight[:,:-32]
zhuwenwen's avatar
zhuwenwen committed
279
            if bias is not None:
zhuwenwen's avatar
zhuwenwen committed
280
                if len(x.shape) == 2: 
281
                    return torch.addmm(bias, x, layer.weight)
zhuwenwen's avatar
zhuwenwen committed
282
                else:
283
                    return torch.matmul(x, layer.weight) + bias
zhuwenwen's avatar
zhuwenwen committed
284
            else:
285
                return torch.matmul(x, layer.weight)
zhuwenwen's avatar
zhuwenwen committed
286
        else:
zhuwenwen's avatar
zhuwenwen committed
287
288
289
290
291
292
293
294
295
296
            # if envs.VLLM_USE_NN and x.shape[-1] == layer.weight.shape[0]:
            #     return dispatch_unquantized_gemm()(layer, x, layer.weight.t(), bias)
            if envs.VLLM_USE_NN:
                if bias is not None:
                    if len(x.shape) == 2: 
                        return torch.addmm(bias, x, layer.weight)
                    else:
                        return torch.matmul(x, layer.weight) + bias
                else:
                    return torch.matmul(x, layer.weight)
297
            else:
zhuwenwen's avatar
zhuwenwen committed
298
                return dispatch_unquantized_gemm()(layer, x, layer.weight, bias)
299

300

301
class LinearBase(CustomOp):
302
    """Base linear layer.
303
304
305
306
307
308

    Args:
        input_size: input dimension of the linear layer.
        output_size: output dimension of the linear layer.
        skip_bias_add: If true, skip adding bias but instead return it.
        params_dtype: Data type for the parameters.
309
        quant_config: Quantization configure.
310
        prefix: Prefix for parameter names.
311
        return_bias: If true, return bias together with outputs in forward pass.
312
        disable_tp: If true, tensor parallelism will be disabled for this layer.
313
314
315
316
317
318
319
    """

    def __init__(
        self,
        input_size: int,
        output_size: int,
        skip_bias_add: bool = False,
320
321
        params_dtype: torch.dtype | None = None,
        quant_config: QuantizationConfig | None = None,
322
        prefix: str = "",
323
324
        *,
        return_bias: bool = True,
325
        disable_tp: bool = False,
326
327
328
329
330
331
332
333
334
335
    ):
        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
336
337
        self.quant_config = quant_config
        self.prefix = prefix
338
        self.allow_fp8_block_shape_mismatch = False
339
        if quant_config is None:
340
            self.quant_method: QuantizeMethodBase | None = UnquantizedLinearMethod()
341
        else:
342
            self.quant_method = quant_config.get_quant_method(self, prefix=prefix)
343
        self.return_bias = return_bias
344
        self.disable_tp = disable_tp
345
346
        self.tp_rank = get_tensor_model_parallel_rank() if not disable_tp else 0
        self.tp_size = get_tensor_model_parallel_world_size() if not disable_tp else 1
347

348
    def update_param_tp_status(self):
349
350
351
352
        for param in self.parameters():
            if isinstance(param, BasevLLMParameter):
                param.tp_rank = self.tp_rank
                param.tp_size = self.tp_size
353
354


355
# --8<-- [start:replicated_linear]
356
@CustomOp.register("replicated_linear")
357
358
359
360
361
362
363
364
365
366
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.
367
368
        prefix: The name of the layer in the state dict, including all parents
                        (e.g. model.layers.0.qkv_proj)
369
        return_bias: If true, return bias together with outputs in forward pass.
370
        disable_tp: Take no effect for replicated linear layers.
371
372
    """

373
374
    # --8<-- [end:replicated_linear]

375
376
377
378
379
380
    def __init__(
        self,
        input_size: int,
        output_size: int,
        bias: bool = True,
        skip_bias_add: bool = False,
381
382
        params_dtype: torch.dtype | None = None,
        quant_config: QuantizationConfig | None = None,
383
384
385
        prefix: str = "",
        *,
        return_bias: bool = True,
386
        disable_tp: bool = False,
387
    ):
388
389
390
391
392
        # If MergedReplicatedLinear, use output size of each partition.
        if hasattr(self, "output_sizes"):
            self.output_partition_sizes = self.output_sizes
        else:
            self.output_partition_sizes = [output_size]
393

394
395
396
397
398
399
400
401
402
403
        super().__init__(
            input_size,
            output_size,
            skip_bias_add,
            params_dtype,
            quant_config,
            prefix=prefix,
            return_bias=return_bias,
            disable_tp=disable_tp,
        )
404

405
406
        # All the linear layer supports quant method.
        assert self.quant_method is not None
407
408
409
410
411
412
413
414
415
        self.quant_method.create_weights(
            self,
            self.input_size,
            self.output_partition_sizes,
            self.input_size,
            self.output_size,
            self.params_dtype,
            weight_loader=self.weight_loader,
        )
416

417
418
        if bias:
            self.bias = Parameter(
419
420
421
422
423
424
425
426
427
                torch.empty(self.output_size, dtype=self.params_dtype)
            )
            set_weight_attrs(
                self.bias,
                {
                    "output_dim": 0,
                    "weight_loader": self.weight_loader,
                },
            )
428
429
        else:
            self.register_parameter("bias", None)
430
        self.is_quantization = not isinstance(self.quant_method, UnquantizedLinearMethod)
431

432
433
434
    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).
435
436
437
438
439
440
441
442
443
444
445
        # 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)

446
447
448
        if len(loaded_weight.shape) == 0:
            loaded_weight = loaded_weight.reshape(1)

449
        if envs.VLLM_USE_NN and not self.is_quantization:
450
451
            loaded_weight = loaded_weight.t()
            
452
453
        assert param.size() == loaded_weight.size(), (
            f"Tried to load weights of size {loaded_weight.size()}"
454
455
            f"to a parameter of size {param.size()}"
        )
456
457
        param.data.copy_(loaded_weight)

458
    def forward(
459
        self,
zhuwenwen's avatar
zhuwenwen committed
460
        x: torch.Tensor,
461
    ) -> torch.Tensor | tuple[torch.Tensor, Parameter | None]:
zhuwenwen's avatar
zhuwenwen committed
462
463
        bias = self.bias if not self.skip_bias_add else None
        assert self.quant_method is not None
464

zhuwenwen's avatar
zhuwenwen committed
465
        output = self.quant_method.apply(self, x, bias)
466

zhuwenwen's avatar
zhuwenwen committed
467
468
        if not self.return_bias:
            return output
469
        output_bias = self.bias if self.skip_bias_add else None
zhuwenwen's avatar
zhuwenwen committed
470
        return output, output_bias
471

472
473
474
475
476
477
    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

478

479
# --8<-- [start:column_parallel_linear]
480
@CustomOp.register("column_parallel_linear")
481
class ColumnParallelLinear(LinearBase):
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
    """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.
498
        quant_config: Quantization configure.
499
        prefix: The name of the layer in the state dict, including all parents
500
                        (e.g. model.layers.0.qkv_proj)
501
502
        return_bias: If true, return bias together with outputs in forward pass.
        disable_tp: If true, weights matrix won't be sharded through tp rank.
503
504
    """

505
506
    # --8<-- [end:column_parallel_linear]

507
508
509
510
511
512
513
    def __init__(
        self,
        input_size: int,
        output_size: int,
        bias: bool = True,
        gather_output: bool = False,
        skip_bias_add: bool = False,
514
515
        params_dtype: torch.dtype | None = None,
        quant_config: QuantizationConfig | None = None,
516
517
518
        prefix: str = "",
        *,
        return_bias: bool = True,
519
        disable_tp: bool = False,
520
    ):
521
        # Divide the weight matrix along the last dimension.
522
523
        self.tp_rank = get_tensor_model_parallel_rank() if not disable_tp else 0
        self.tp_size = get_tensor_model_parallel_world_size() if not disable_tp else 1
524
525
        self.input_size_per_partition = input_size
        self.output_size_per_partition = divide(output_size, self.tp_size)
526
527
528
529
        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 = [
530
                divide(output_size, self.tp_size) for output_size in self.output_sizes
531
532
            ]

533
534
535
536
537
538
539
540
541
542
        super().__init__(
            input_size,
            output_size,
            skip_bias_add,
            params_dtype,
            quant_config,
            prefix,
            return_bias=return_bias,
            disable_tp=disable_tp,
        )
543

544
        self._maybe_allow_fp8_block_shape_mismatch()
545
546
547
        self.gather_output = gather_output

        assert self.quant_method is not None
548
549
        self.quant_method.create_weights(
            layer=self,
550
            input_size_per_partition=self.input_size_per_partition,
551
552
553
554
            output_partition_sizes=self.output_partition_sizes,
            input_size=self.input_size,
            output_size=self.output_size,
            params_dtype=self.params_dtype,
555
            weight_loader=(
556
557
558
559
560
                self.weight_loader_v2
                if self.quant_method.__class__.__name__ in WEIGHT_LOADER_V2_SUPPORTED
                else self.weight_loader
            ),
        )
561
562
        if bias:
            self.bias = Parameter(
563
564
565
566
567
568
569
570
571
                torch.empty(self.output_size_per_partition, dtype=params_dtype)
            )
            set_weight_attrs(
                self.bias,
                {
                    "output_dim": 0,
                    "weight_loader": self.weight_loader,
                },
            )
572
573
        else:
            self.register_parameter("bias", None)
574
        self.update_param_tp_status()
575
        self.is_quantization = not isinstance(self.quant_method, UnquantizedLinearMethod)
576

577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
    def _maybe_allow_fp8_block_shape_mismatch(self) -> None:
        quant_config = getattr(self, "quant_config", None)
        weight_block = getattr(quant_config, "weight_block_size", None)
        if (
            weight_block is None
            or len(weight_block) < 1
            or len(self.output_partition_sizes) <= 1
        ):
            return

        try:
            block_n = int(weight_block[0])
        except (ValueError, TypeError):
            return

        if block_n <= 0:
            return

        if any(size % block_n != 0 for size in self.output_partition_sizes):
            self.allow_fp8_block_shape_mismatch = True
            logger.debug(
                "Allowing FP8 block shape mismatch for %s (block_n=%d, partitions=%s)",
                getattr(self, "prefix", "<unknown>"),
                block_n,
                self.output_partition_sizes,
            )

604
605
    def weight_loader(self, param: Parameter, loaded_weight: torch.Tensor):
        output_dim = getattr(param, "output_dim", None)
606

607
608
609
610
611
612
        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

613
614
615
616
617
618
619
620
        # 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):
621
622
            final_shape = list(loaded_weight.shape)
            if output_dim is not None:
623
                assert final_shape[output_dim] % self.tp_size == 0
624
                final_shape[output_dim] = final_shape[output_dim] // self.tp_size
625
            param.materialize(final_shape, dtype=loaded_weight.dtype)
626

627
        param_data = param.data
628
        if output_dim is not None and not is_sharded_weight:
629
            if not envs.VLLM_USE_NN or len(param_data.shape)==1 or self.is_quantization:
630
631
632
                shard_size = param_data.shape[output_dim] 
            else:
                shard_size = param_data.shape[int(not(output_dim))]
633
            start_idx = self.tp_rank * shard_size
634
            loaded_weight = loaded_weight.narrow(output_dim, start_idx, shard_size)
635

636
637
638
639
        # 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)
640

641
        if envs.VLLM_USE_NN and not self.is_quantization:
642
            loaded_weight = loaded_weight.t()
643
644
645
        assert param_data.shape == loaded_weight.shape
        param_data.copy_(loaded_weight)

646
    def weight_loader_v2(self, param: BasevLLMParameter, loaded_weight: torch.Tensor):
647
648
649
650
651
        # 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)
652
        param.load_column_parallel_weight(loaded_weight=loaded_weight, is_quantization=self.is_quantization)
653

654
    def forward(
655
656
        self,
        input_,
657
658
        *,
        iqis: tuple[torch.Tensor, torch.Tensor] | None = None
659
    ) -> torch.Tensor | tuple[torch.Tensor, Parameter | None]:
zhuwenwen's avatar
zhuwenwen committed
660
661
662
663
        bias = self.bias if not self.skip_bias_add else None

        # Matrix multiply.
        assert self.quant_method is not None
664
665
666
667
        if iqis is not None:
            output_parallel = self.quant_method.apply(self, input_, bias, input_quant_args=iqis)
        else:
            output_parallel = self.quant_method.apply(self, input_, bias)
zhuwenwen's avatar
zhuwenwen committed
668
669
670
671

        if self.gather_output and self.tp_size > 1:
            # All-gather across the partitions.
            output = tensor_model_parallel_all_gather(output_parallel)
672
        else:
zhuwenwen's avatar
zhuwenwen committed
673
            output = output_parallel
674

zhuwenwen's avatar
zhuwenwen committed
675
676
        if not self.return_bias:
            return output
677
        output_bias = self.bias if self.skip_bias_add else None
zhuwenwen's avatar
zhuwenwen committed
678
        return output, output_bias
679

680
681
682
683
    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}"
684
        s += f", tp_size={self.tp_size}"
685
686
687
        s += f", gather_output={self.gather_output}"
        return s

688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706

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.
707
        quant_config: Quantization configure.
708
709
        prefix: The name of the layer in the state dict, including all parents
                        (e.g. model.layers.0.qkv_proj)
710
        return_bias: If true, return bias together with outputs in forward pass.
711
712
        disable_tp: If true, all weights matrix won't be sharded, this layer
                    will be treated as a "Replicated" MergedLinear.
713
    """
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
    def forward(
        self,
        input_,
        *, iqis: tuple[torch.Tensor, torch.Tensor] | None = None

    ) -> torch.Tensor | tuple[torch.Tensor, Parameter | None]:
        bias = self.bias if not self.skip_bias_add else None

        # Matrix multiply.
        assert self.quant_method is not None
        if envs.USE_FUSED_RMS_QUANT and iqis is not None:
            output_parallel = self.quant_method.apply(self, input_, bias, input_quant_args=iqis)
        else:
            output_parallel = self.quant_method.apply(self, input_, bias)

        if self.gather_output and self.tp_size > 1:
            # All-gather across the partitions.
            output = tensor_model_parallel_all_gather(output_parallel)
        else:
            output = output_parallel

        if not self.return_bias:
            return output
        output_bias = self.bias if self.skip_bias_add else None
        return output, output_bias
739

740
741
742
743
744
745
746
    def __init__(
        self,
        input_size: int,
        output_sizes: list[int],
        bias: bool = True,
        gather_output: bool = False,
        skip_bias_add: bool = False,
747
748
        params_dtype: torch.dtype | None = None,
        quant_config: QuantizationConfig | None = None,
749
750
751
        prefix: str = "",
        *,
        return_bias: bool = True,
752
        disable_tp: bool = False,
753
    ):
754
        self.output_sizes = output_sizes
755
756
        self.tp_size = get_tensor_model_parallel_world_size() if not disable_tp else 1
        self.tp_rank = get_tensor_model_parallel_rank() if not disable_tp else 0
757

758
759
760
761
762
763
764
765
766
767
768
769
770
        assert all(output_size % self.tp_size == 0 for output_size in output_sizes)
        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,
            quant_config=quant_config,
            prefix=prefix,
            return_bias=return_bias,
            disable_tp=disable_tp,
        )
771
        self.is_quantization = not isinstance(self.quant_method, UnquantizedLinearMethod)
James Fleming's avatar
James Fleming committed
772

773
774
775
776
    def weight_loader(
        self,
        param: Parameter,
        loaded_weight: torch.Tensor,
777
        loaded_shard_id: int | None = None,
778
    ):
779
780
781
782
783
        # 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:
784
785
786
787
788
            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 = {
789
                    i: loaded_weight.item() for i, _ in enumerate(self.output_sizes)
790
                }
791
792
            return

793
794
        if is_gguf_weight:
            output_dim = getattr(param, "output_dim", None)
795
796
            shard_size = loaded_weight.size(output_dim) // self.tp_size
            start_idx = self.tp_rank * shard_size
797

798
            if loaded_shard_id is not None:
799
                loaded_weight = loaded_weight.narrow(output_dim, start_idx, shard_size)
800
801
802
803
                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
804

805
806
        param_data = param.data
        output_dim = getattr(param, "output_dim", None)
807
808
        # Special case for per-tensor scale to load scalar into fused array.
        needs_scalar_to_array = getattr(param, "needs_scalar_to_array", False)
809

810
        if loaded_shard_id is None:
811
812
            # Loaded weight is already fused on disk (mlp).
            # (e.g., Phi-3's gate_up_proj).
813
            if output_dim is None:
814
                if needs_scalar_to_array:
815
                    param_data, loaded_weight = adjust_scalar_to_fused_array(
816
817
                        param_data, loaded_weight, 0
                    )
818

819
820
821
822
                assert param_data.shape == loaded_weight.shape
                param_data.copy_(loaded_weight)
                return
            current_shard_offset = 0
823
            use_bitsandbytes_4bit = getattr(param, "use_bitsandbytes_4bit", False)
824
            shard_offsets: list[tuple[int, int, int]] = []
825
826
827
828
829
            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:
830
                # Special case for Quantization.
831
832
833
                # If quantized, we need to adjust the offset and size to account
                # for the packing.
                if packed_dim == output_dim:
834
835
                    shard_size = shard_size // param.packed_factor
                    shard_offset = shard_offset // param.packed_factor
836
                    # Special case for Marlin.
837
                    shard_size, shard_offset = adjust_marlin_shard(
838
839
                        param, shard_size, shard_offset
                    )
840

841
                shard_size, shard_offset = adjust_bitblas_shard(
842
843
                    param, shard_size, shard_offset
                )
844

845
                if use_bitsandbytes_4bit:
846
847
848
849
850
851
852
                    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(
853
854
                        param, orig_offsets, str(shard_id)
                    )
855

856
                loaded_weight_shard = loaded_weight.narrow(
857
858
                    output_dim, shard_offset, shard_size
                )
859
860
861
862
863
                self.weight_loader(param, loaded_weight_shard, shard_id)
            return

        assert loaded_shard_id < len(self.output_sizes)
        if output_dim is not None:
864
865
866
867
868
869
870
871
872
873
874
875
            shard_offset = sum(self.output_sizes[:loaded_shard_id])
            shard_size = self.output_sizes[loaded_shard_id]

            if isinstance(param, BlockQuantScaleParameter):
                weight_block_size = getattr(self, "weight_block_size", None)
                shard_size, shard_offset = adjust_block_scale_shard(
                    weight_block_size, shard_size, shard_offset
                )

            shard_offset //= self.tp_size
            shard_size //= self.tp_size

876
            # Special case for quantization.
877
878
879
880
            # 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:
881
882
                shard_size = shard_size // param.packed_factor
                shard_offset = shard_offset // param.packed_factor
883
                # Special case for Marlin.
884
                shard_size, shard_offset = adjust_marlin_shard(
885
886
                    param, shard_size, shard_offset
                )
887
            shard_size, shard_offset = adjust_bitblas_shard(
888
889
                param, shard_size, shard_offset
            )
gaoqiong's avatar
gaoqiong committed
890

891
            use_bitsandbytes_4bit = getattr(param, "use_bitsandbytes_4bit", False)
892
893
894
895
896
            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

897
            if use_bitsandbytes_4bit:
898
                shard_size = loaded_weight.shape[output_dim]
899
                shard_offset = loaded_weight.shape[output_dim] * loaded_shard_id
900

901
            if not envs.VLLM_USE_NN or self.is_quantization or (envs.VLLM_USE_NN and param_data.dim()==1):
902
903
904
                param_data = param_data.narrow(output_dim, shard_offset, shard_size)
            else:
                param_data = param_data.narrow(int(not(output_dim)), shard_offset, shard_size)
905
            start_idx = self.tp_rank * shard_size
906
            if not is_sharded_weight:
907
                loaded_weight = loaded_weight.narrow(output_dim, start_idx, shard_size)
908
909
910
        # Special case for per-tensor scales in fused case.
        elif needs_scalar_to_array:
            param_data, loaded_weight = adjust_scalar_to_fused_array(
911
912
                param_data, loaded_weight, loaded_shard_id
            )
913

914
        else:
CHU Tianxiang's avatar
CHU Tianxiang committed
915
916
917
918
919
            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 "
920
921
                    "the same for all partitions."
                )
922

923
        if envs.VLLM_USE_NN and not self.is_quantization:
924
925
            loaded_weight = loaded_weight.t()
            
gaoqiong's avatar
gaoqiong committed
926
927
        assert param_data.shape == loaded_weight.shape
        param_data.copy_(loaded_weight)
928

929
930
931
    def _load_fused_module_from_checkpoint(
        self, param: BasevLLMParameter, loaded_weight: torch.Tensor
    ):
932
933
934
        """
        Handle special case for models where MLP layers are already
        fused on disk. In this case, we have no shard id. This function
935
        determines the shard id by splitting these layers and then calls
936
937
938
939
940
941
942
        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
943
        shard_offsets: list[tuple[int, int, int]] = []
944
945
946
947
948
949
950
951
        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.
952
953
954
955
956
957
958
959
960
961
962
            if (
                isinstance(param, (PackedColumnParameter, PackedvLLMParameter))
                and param.packed_dim == param.output_dim
            ):
                shard_size, shard_offset = param.adjust_shard_indexes_for_packing(
                    shard_size=shard_size, shard_offset=shard_offset
                )

            loaded_weight_shard = loaded_weight.narrow(
                param.output_dim, shard_offset, shard_size
            )
963
964
            self.weight_loader_v2(param, loaded_weight_shard, shard_id)

965
966
967
968
    def weight_loader_v2(
        self,
        param: BasevLLMParameter,
        loaded_weight: torch.Tensor,
969
        loaded_shard_id: int | None = None,
970
    ):
971
        if loaded_shard_id is None:
972
            if isinstance(param, PerTensorScaleParameter):
973
                param.load_merged_column_weight(loaded_weight=loaded_weight, shard_id=0)
974
                return
975
            elif type(param) in (RowvLLMParameter, BasevLLMParameter):
976
                param.load_merged_column_weight(loaded_weight=loaded_weight)
977
                return
978
            # TODO: @dsikka - move to parameter.py
979
980
981
982
983
            self._load_fused_module_from_checkpoint(param, loaded_weight)
            return

        assert loaded_shard_id < len(self.output_sizes)

984
985
986
        shard_offset = sum(self.output_sizes[:loaded_shard_id])
        shard_size = self.output_sizes[loaded_shard_id]

987
        if isinstance(param, BlockQuantScaleParameter):
988
989
990
            weight_block_size = getattr(self, "weight_block_size", None)
            shard_size, shard_offset = adjust_block_scale_shard(
                weight_block_size, shard_size, shard_offset
991
            )
992
993
994

        shard_offset //= self.tp_size
        shard_size //= self.tp_size
995

996
997
998
999
1000
1001
        param.load_merged_column_weight(
            loaded_weight=loaded_weight,
            shard_id=loaded_shard_id,
            shard_offset=shard_offset,
            shard_size=shard_size,
            tp_rank=self.tp_rank,
1002
            is_quantization=self.is_quantization
1003
        )
1004

1005
1006
1007
1008
1009
1010
1011
1012
1013
1014
1015
1016
1017
1018
1019
1020
1021
1022
1023
1024
1025
1026

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.
1027
        quant_config: Quantization configure.
1028
1029
        prefix: The name of the layer in the state dict, including all parents
                        (e.g. model.layers.0.qkv_proj)
1030
        return_bias: If true, return bias together with outputs in forward pass.
1031
        disable_tp: If true, weights matrix won't be sharded through tp rank.
1032
1033
    """

1034
1035
1036
1037
1038
    def __init__(
        self,
        hidden_size: int,
        head_size: int,
        total_num_heads: int,
1039
        total_num_kv_heads: int | None = None,
1040
1041
        bias: bool = True,
        skip_bias_add: bool = False,
1042
1043
        params_dtype: torch.dtype | None = None,
        quant_config: QuantizationConfig | None = None,
1044
1045
1046
        prefix: str = "",
        *,
        return_bias: bool = True,
1047
        disable_tp: bool = False,
1048
        v_head_size: int | None = None,
1049
    ):
1050
1051
        self.hidden_size = hidden_size
        self.head_size = head_size
1052
        self.v_head_size = v_head_size if v_head_size is not None else head_size
1053
1054
1055
1056
1057
        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.
1058
        tp_size = get_tensor_model_parallel_world_size() if not disable_tp else 1
1059
1060
1061
        self.num_heads = divide(self.total_num_heads, tp_size)
        if tp_size >= self.total_num_kv_heads:
            self.num_kv_heads = 1
1062
            self.num_kv_head_replicas = divide(tp_size, self.total_num_kv_heads)
1063
1064
1065
1066
        else:
            self.num_kv_heads = divide(self.total_num_kv_heads, tp_size)
            self.num_kv_head_replicas = 1
        input_size = self.hidden_size
1067
        output_size = (
1068
1069
1070
1071
            self.num_heads * self.head_size
            + self.num_kv_heads * self.head_size
            + self.num_kv_heads * self.v_head_size
        ) * tp_size
1072
1073
1074
        self.output_sizes = [
            self.num_heads * self.head_size * tp_size,  # q_proj
            self.num_kv_heads * self.head_size * tp_size,  # k_proj
1075
            self.num_kv_heads * self.v_head_size * tp_size,  # v_proj
James Fleming's avatar
James Fleming committed
1076
        ]
gaoqiong's avatar
gaoqiong committed
1077

1078
1079
1080
1081
1082
1083
1084
1085
1086
1087
1088
1089
        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,
            quant_config=quant_config,
            prefix=prefix,
            return_bias=return_bias,
            disable_tp=disable_tp,
        )
1090
        self.is_quantization = not isinstance(self.quant_method, UnquantizedLinearMethod)
1091

1092
1093
1094
1095
1096
    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,
1097
1098
            "total": (self.num_heads + self.num_kv_heads) * self.head_size
            + self.num_kv_heads * self.v_head_size,
1099
1100
1101
1102
1103
1104
1105
        }
        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,
1106
            "v": self.num_kv_heads * self.v_head_size,
1107
1108
1109
        }
        return shard_size_mapping.get(loaded_shard_id)

1110
1111
1112
    def _load_fused_module_from_checkpoint(
        self, param: BasevLLMParameter, loaded_weight: torch.Tensor
    ):
1113
        """
1114
        Handle special case for models where QKV layers are already
1115
        fused on disk. In this case, we have no shard id. This function
1116
        determines the shard id by splitting these layers and then calls
1117
1118
1119
1120
1121
1122
1123
1124
        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),
1125
1126
1127
1128
1129
1130
1131
1132
            (
                "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,
1133
                self.total_num_kv_heads * self.v_head_size,
1134
            ),
1135
1136
1137
1138
1139
1140
        ]

        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.
1141
1142
1143
1144
1145
1146
1147
1148
1149
1150
1151
            if (
                isinstance(param, (PackedColumnParameter, PackedvLLMParameter))
                and param.packed_dim == param.output_dim
            ):
                shard_size, shard_offset = param.adjust_shard_indexes_for_packing(
                    shard_size=shard_size, shard_offset=shard_offset
                )

            loaded_weight_shard = loaded_weight.narrow(
                param.output_dim, shard_offset, shard_size
            )
1152
1153
            self.weight_loader_v2(param, loaded_weight_shard, shard_id)

1154
1155
1156
1157
    def weight_loader_v2(
        self,
        param: BasevLLMParameter,
        loaded_weight: torch.Tensor,
1158
        loaded_shard_id: str | None = None,
1159
    ):
1160
        if loaded_shard_id is None:  # special case for certain models
1161
            if isinstance(param, PerTensorScaleParameter):
1162
1163
1164
                param.load_qkv_weight(
                    loaded_weight=loaded_weight, shard_id=0, tp_rank=self.tp_rank
                )
1165
                return
1166
            elif type(param) in (RowvLLMParameter, BasevLLMParameter):
1167
                param.load_qkv_weight(loaded_weight=loaded_weight, tp_rank=self.tp_rank)
1168
                return
1169
            # TODO: @dsikka - move to parameter.py
1170
1171
1172
1173
1174
1175
1176
1177
            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)

1178
        if isinstance(param, BlockQuantScaleParameter):
1179
1180
1181
1182
            weight_block_size = getattr(self, "weight_block_size", None)
            shard_size, shard_offset = adjust_block_scale_shard(
                weight_block_size, shard_size, shard_offset
            )
1183

1184
1185
1186
1187
1188
1189
1190
        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,
            tp_rank=self.tp_rank,
1191
            is_quantization=self.is_quantization, 
1192
        )
1193

1194
1195
1196
1197
    def weight_loader(
        self,
        param: Parameter,
        loaded_weight: torch.Tensor,
1198
        loaded_shard_id: str | None = None,
1199
    ):
1200
1201
1202
1203
        # 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)
1204
        if is_gguf_weight_type:
1205
            idx_map = {"q": 0, "k": 1, "v": 2}
1206
1207
1208
1209
            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:
1210
                param.shard_weight_type = {k: loaded_weight.item() for k in idx_map}
1211
1212
            return

1213
1214
        if is_gguf_weight:
            output_dim = getattr(param, "output_dim", None)
1215
1216
            shard_size = loaded_weight.size(output_dim) // self.tp_size
            start_idx = self.tp_rank * shard_size
1217

1218
            if loaded_shard_id is not None:
1219
                loaded_weight = loaded_weight.narrow(output_dim, start_idx, shard_size)
1220
1221
1222
1223
                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
1224

1225
1226
        param_data = param.data
        output_dim = getattr(param, "output_dim", None)
1227

1228
1229
        # Special case for per-tensor scales in fused case.
        needs_scalar_to_array = getattr(param, "needs_scalar_to_array", False)
1230

1231
        if loaded_shard_id is None:
1232
1233
            # Loaded weight is already fused on disk (qkv).
            # (e.g., Phi-3's qkv_proj).
1234
            if output_dim is None:
1235
                if needs_scalar_to_array:
1236
                    param_data, loaded_weight = adjust_scalar_to_fused_array(
1237
1238
                        param_data, loaded_weight, 0
                    )
1239

1240
1241
1242
1243
1244
1245
                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),
1246
1247
1248
1249
1250
1251
1252
1253
                (
                    "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,
1254
                    self.total_num_kv_heads * self.v_head_size,
1255
                ),
1256
            ]
1257
            use_bitsandbytes_4bit = getattr(param, "use_bitsandbytes_4bit", False)
1258

1259
1260
            packed_dim = getattr(param, "packed_dim", None)
            for shard_id, shard_offset, shard_size in shard_offsets:
1261
                # Special case for Quantized Weights.
1262
1263
1264
                # If quantized, we need to adjust the offset and size to account
                # for the packing.
                if packed_dim == output_dim:
1265
1266
                    shard_size = shard_size // param.packed_factor
                    shard_offset = shard_offset // param.packed_factor
1267

1268
                    # Special case for Marlin.
1269
                    shard_size, shard_offset = adjust_marlin_shard(
1270
1271
                        param, shard_size, shard_offset
                    )
1272

1273
1274
1275
                if use_bitsandbytes_4bit:
                    orig_qkv_offsets = {
                        "q": (0, self.total_num_heads * self.head_size),
1276
1277
1278
1279
1280
1281
1282
                        "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,
1283
                            self.total_num_kv_heads * self.v_head_size,
1284
1285
                        ),
                        "total": (
1286
1287
1288
                            (self.total_num_heads + self.total_num_kv_heads)
                            * self.head_size
                            + self.total_num_kv_heads * self.v_head_size,
1289
1290
                            0,
                        ),
1291
1292
1293
                    }

                    shard_size, shard_offset = adjust_bitsandbytes_4bit_shard(
1294
1295
                        param, orig_qkv_offsets, shard_id
                    )
1296

1297
                loaded_weight_shard = loaded_weight.narrow(
1298
1299
                    output_dim, shard_offset, shard_size
                )
1300
1301
1302
1303
                self.weight_loader(param, loaded_weight_shard, shard_id)
            return

        assert loaded_shard_id in ["q", "k", "v"]
1304
1305

        # If output dim is defined, use the default loading process.
1306
1307
1308
1309
1310
1311
1312
1313
        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":
1314
                shard_offset = (self.num_heads + self.num_kv_heads) * self.head_size
1315
                shard_size = self.num_kv_heads * self.v_head_size
1316
1317
1318
1319
1320
1321
1322

            if isinstance(param, BlockQuantScaleParameter):
                weight_block_size = getattr(self, "weight_block_size", None)
                shard_size, shard_offset = adjust_block_scale_shard(
                    weight_block_size, shard_size, shard_offset
                )

1323
            # Special case for Quantized Weights.
1324
1325
1326
1327
            # 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:
1328
1329
                shard_size = shard_size // param.packed_factor
                shard_offset = shard_offset // param.packed_factor
1330

1331
                # Special case for Marlin.
1332
                shard_size, shard_offset = adjust_marlin_shard(
1333
1334
                    param, shard_size, shard_offset
                )
1335

1336
            use_bitsandbytes_4bit = getattr(param, "use_bitsandbytes_4bit", False)
1337
1338
1339
1340
1341
            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

1342
            if use_bitsandbytes_4bit:
1343
1344
                orig_qkv_offsets = {
                    "q": (0, self.num_heads * self.head_size),
1345
1346
1347
1348
1349
1350
                    "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,
1351
                        self.num_kv_heads * self.v_head_size,
1352
1353
                    ),
                    "total": (
1354
1355
                        (self.num_heads + self.num_kv_heads) * self.head_size
                        + self.num_kv_heads * self.v_head_size,
1356
1357
                        0,
                    ),
1358
                }
1359
                shard_size, shard_offset = adjust_bitsandbytes_4bit_shard(
1360
1361
                    param, orig_qkv_offsets, loaded_shard_id
                )
gaoqiong's avatar
gaoqiong committed
1362

1363
            if not envs.VLLM_USE_NN or len(param_data.shape)==1 or self.is_quantization:
1364
                param_data = param_data.narrow(output_dim, shard_offset, shard_size)
1365
1366
1367
            else:
                param_data = param_data.narrow(int(not(output_dim)), shard_offset,
                                               shard_size)
zhuwenwen's avatar
zhuwenwen committed
1368
            if loaded_shard_id == "q":
1369
                shard_rank = self.tp_rank
1370
            else:
1371
1372
                shard_rank = self.tp_rank // self.num_kv_head_replicas
            start_idx = shard_rank * shard_size
1373

1374
            if not is_sharded_weight:
1375
                loaded_weight = loaded_weight.narrow(output_dim, start_idx, shard_size)
1376

1377
1378
1379
        # Special case for per-tensor scales in fused case.
        elif needs_scalar_to_array:
            param_data, loaded_weight = adjust_scalar_to_fused_array(
1380
1381
                param_data, loaded_weight, loaded_shard_id
            )
1382
        else:
CHU Tianxiang's avatar
CHU Tianxiang committed
1383
1384
1385
1386
1387
            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 "
1388
1389
                    "for all partitions."
                )
gaoqiong's avatar
gaoqiong committed
1390

1391
        if envs.VLLM_USE_NN and not self.is_quantization:
1392
            loaded_weight = loaded_weight.t()
gaoqiong's avatar
gaoqiong committed
1393
1394
        assert param_data.shape == loaded_weight.shape
        param_data.copy_(loaded_weight)
1395
1396


1397
# --8<-- [start:row_parallel_linear]
1398
@CustomOp.register("row_parallel_linear")
1399
class RowParallelLinear(LinearBase):
1400
1401
1402
1403
1404
1405
1406
1407
1408
1409
1410
1411
1412
1413
1414
1415
1416
1417
1418
1419
1420
1421
    """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.
1422
1423
1424
        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
1425
        quant_config: Quantization configure.
1426
1427
1428
        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.
1429
        disable_tp: If true, weights matrix won't be sharded through tp rank.
1430
1431
    """

1432
1433
    # --8<-- [end:row_parallel_linear]

1434
1435
1436
1437
1438
1439
1440
    def __init__(
        self,
        input_size: int,
        output_size: int,
        bias: bool = True,
        input_is_parallel: bool = True,
        skip_bias_add: bool = False,
1441
        params_dtype: torch.dtype | None = None,
1442
        reduce_results: bool = True,
1443
        quant_config: QuantizationConfig | None = None,
1444
1445
1446
        prefix: str = "",
        *,
        return_bias: bool = True,
1447
        disable_tp: bool = False,
1448
    ):
1449
        # Divide the weight matrix along the first dimension.
1450
1451
        self.tp_rank = get_tensor_model_parallel_rank() if not disable_tp else 0
        self.tp_size = get_tensor_model_parallel_world_size() if not disable_tp else 1
1452
1453
1454
1455
        self.input_size_per_partition = divide(input_size, self.tp_size)
        self.output_size_per_partition = output_size
        self.output_partition_sizes = [output_size]

1456
1457
1458
1459
1460
1461
1462
1463
1464
1465
        super().__init__(
            input_size,
            output_size,
            skip_bias_add,
            params_dtype,
            quant_config,
            prefix,
            return_bias=return_bias,
            disable_tp=disable_tp,
        )
1466

1467
1468
1469
        self.input_is_parallel = input_is_parallel
        self.reduce_results = reduce_results

1470
        assert self.quant_method is not None
1471
1472
1473
        self.quant_method.create_weights(
            layer=self,
            input_size_per_partition=self.input_size_per_partition,
1474
            output_partition_sizes=self.output_partition_sizes,
1475
1476
1477
            input_size=self.input_size,
            output_size=self.output_size,
            params_dtype=self.params_dtype,
1478
            weight_loader=(
1479
1480
1481
1482
1483
                self.weight_loader_v2
                if self.quant_method.__class__.__name__ in WEIGHT_LOADER_V2_SUPPORTED
                else self.weight_loader
            ),
        )
1484
        if not reduce_results and (bias and not skip_bias_add):
1485
1486
1487
1488
            raise ValueError(
                "When not reduce the results, adding bias to the "
                "results can lead to incorrect results"
            )
1489
1490

        if bias:
1491
1492
1493
1494
1495
1496
1497
1498
            self.bias = Parameter(torch.empty(self.output_size, dtype=params_dtype))
            set_weight_attrs(
                self.bias,
                {
                    "output_dim": 0,
                    "weight_loader": self.weight_loader,
                },
            )
1499
1500
        else:
            self.register_parameter("bias", None)
1501
        self.update_param_tp_status()
1502
        self.is_quantization = not isinstance(self.quant_method, UnquantizedLinearMethod)
1503
1504
1505

    def weight_loader(self, param: Parameter, loaded_weight: torch.Tensor):
        input_dim = getattr(param, "input_dim", None)
1506
        use_bitsandbytes_4bit = getattr(param, "use_bitsandbytes_4bit", False)
1507
1508
1509
1510
        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
1511
1512
1513
1514
1515
1516
1517
1518
1519

        # 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):
1520
1521
            weight_shape = list(loaded_weight.shape)
            if input_dim:
1522
                weight_shape[input_dim] = weight_shape[input_dim] // self.tp_size
1523
            param.materialize(tuple(weight_shape), dtype=loaded_weight.dtype)
1524

1525
        param_data = param.data
1526
        if input_dim is not None and not is_sharded_weight:
1527
            if not envs.VLLM_USE_NN or self.is_quantization:
1528
1529
1530
                shard_size = param_data.shape[input_dim]
            else:
                shard_size = param_data.shape[int(not(input_dim))]
1531
            start_idx = self.tp_rank * shard_size
1532
            loaded_weight = loaded_weight.narrow(input_dim, start_idx, shard_size)
1533

1534
1535
1536
        # 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:
1537
1538
            loaded_weight = loaded_weight.reshape(1)

1539
        if envs.VLLM_USE_NN and not self.is_quantization:
1540
            loaded_weight = loaded_weight.t()
1541
1542
1543
        assert param_data.shape == loaded_weight.shape
        param_data.copy_(loaded_weight)

1544
    def weight_loader_v2(self, param: BasevLLMParameter, loaded_weight: torch.Tensor):
1545
1546
1547
1548
1549
1550
        # 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)

1551
        param.load_row_parallel_weight(loaded_weight=loaded_weight, is_quantization=self.is_quantization)
1552

1553
    def forward(
1554
1555
        self,
        input_,
1556
1557
        *,
        iqis: tuple[torch.Tensor, torch.Tensor] | None = None
1558
    ) -> torch.Tensor | tuple[torch.Tensor, Parameter | None]:
1559
1560
1561
1562
        if self.input_is_parallel:
            input_parallel = input_
        else:
            splitted_input = split_tensor_along_last_dim(
1563
1564
                input_, num_partitions=self.tp_size
            )
1565
            input_parallel = splitted_input[self.tp_rank].contiguous()
1566
1567

        # Matrix multiply.
1568
        assert self.quant_method is not None
1569
1570
1571
        # 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
1572
1573
1574
1575
        if iqis is not None:
            output_parallel = self.quant_method.apply(self, input_parallel, bias_, input_quant_args=iqis)
        else:
            output_parallel = self.quant_method.apply(self, input_parallel, bias_)
1576

1577
        if self.reduce_results and self.tp_size > 1:
zhuwenwen's avatar
zhuwenwen committed
1578
            output = tensor_model_parallel_all_reduce(output_parallel)
1579
        else:
1580
1581
            output = output_parallel

1582
1583
        if not self.return_bias:
            return output
1584
        output_bias = self.bias if self.skip_bias_add else None
1585
        return output, output_bias
1586
1587

    def extra_repr(self) -> str:
1588
        s = f"in_features={self.input_size_per_partition}"
1589
1590
1591
1592
1593
        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