Skip to content

vllm.lora.punica_wrapper.punica_tpu

PunicaWrapperTPU

Bases: PunicaWrapperBase

PunicaWrapperTPU is designed to manage and provide metadata for the punica kernel. The main function is to maintain the state information for Multi-LoRA, and to provide the interface for the pytorch punica ops.

Source code in vllm/lora/punica_wrapper/punica_tpu.py 
 22
 23
 24
 25
 26
 27
 28
 29
 30
 31
 32
 33
 34
 35
 36
 37
 38
 39
 40
 41
 42
 43
 44
 45
 46
 47
 48
 49
 50
 51
 52
 53
 54
 55
 56
 57
 58
 59
 60
 61
 62
 63
 64
 65
 66
 67
 68
 69
 70
 71
 72
 73
 74
 75
 76
 77
 78
 79
 80
 81
 82
 83
 84
 85
 86
 87
 88
 89
 90
 91
 92
 93
 94
 95
 96
 97
 98
 99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
class PunicaWrapperTPU(PunicaWrapperBase):
    """
    PunicaWrapperTPU is designed to manage and provide metadata for the punica
    kernel. The main function is to maintain the state information for
    Multi-LoRA, and to provide the interface for the pytorch punica ops.
    """

    def __init__(self, max_num_batched_tokens: int, max_batches: int,
                 device: Union[torch.device, str], **kwargs):
        PunicaWrapperBase.__init__(self, max_num_batched_tokens, max_batches,
                                   device)

        # PunicaWrapperBase defines some tensors with dtype=torch.int64, which
        # isn't supported by the TPU. So convert those tensors to int32.
        # Not all of them are used by the TPU so only convert the useful ones.
        self._token_lora_indices = self._token_lora_indices.to(
            dtype=torch.int32)
        self._sampler_indices = self._sampler_indices.to(dtype=torch.int32)
        self._sampler_indices_padded = self._sampler_indices_padded.to(
            dtype=torch.int32)

        torch.ops.xla.dynamo_set_buffer_donor_(self._token_lora_indices, True)
        torch.ops.xla.dynamo_set_buffer_donor_(self._sampler_indices, True)
        torch.ops.xla.dynamo_set_buffer_donor_(self._sampler_indices_padded,
                                               True)
        torch.ops.xla.dynamo_set_buffer_donor_(self._embeddings_indices, True)
        torch.ops.xla.dynamo_set_buffer_donor_(self._long_lora_indices, True)
        torch.ops.xla.dynamo_set_buffer_donor_(self._lora_indices_per_batch,
                                               True)

        torch._dynamo.mark_dynamic(self._token_lora_indices, 0)
        torch._dynamo.mark_dynamic(self._embeddings_indices, 1)
        torch._dynamo.mark_dynamic(self._sampler_indices_padded, 0)

    def _get_token_lora_indices(self, x: torch.Tensor) -> torch.IntTensor:
        return torch.narrow(self._token_lora_indices, 0, 0, x.size(0))

    @property
    def embeddings_indices(self) -> torch.Tensor:
        """
        This property provides access to the indices used for lora embeddings,
        specifically for VocabParallelEmbeddingWithLoRA.
        """
        return self._embeddings_indices[:]

    @property
    def sampler_indices_padded(self) -> torch.Tensor:
        """
        This property provides access to padded sampler indices.
        """
        return self._sampler_indices_padded[:]

    def shrink(
        self,
        x: torch.Tensor,
        w_t_all: torch.Tensor,
        scale: float,
    ):
        return bgmv_shrink(x, w_t_all, self._get_token_lora_indices(x), scale)

    def expand(self, y: torch.Tensor, x: torch.Tensor, w_t_all: torch.Tensor,
               add_inputs: bool):
        return bgmv_expand(x, w_t_all, y, self._get_token_lora_indices(x),
                           add_inputs)

    def expand_slice(self, y: torch.Tensor, x: torch.Tensor,
                     w_t_all: torch.Tensor, y_offset: int, y_slice_size: int,
                     add_inputs: bool) -> torch.Tensor:
        return bgmv_expand_slice(x, w_t_all, y,
                                 self._get_token_lora_indices(x), y_offset,
                                 y_slice_size, add_inputs)

    def add_shrink(self, y: Union[tuple[torch.Tensor, ...], torch.Tensor],
                   x: torch.Tensor, lora_a_stacked: tuple[torch.Tensor, ...],
                   scale: float, **kwargs) -> Optional[torch.Tensor]:
        """
        Performs GEMM for multiple slices of lora_a.

        Semantics:
        for i in range(len(lora_a_stacked)):
            y[i] += (x @ lora_a_stacked[i]) * scale

        Args:
            y (Union[tuple[torch.Tensor, ...], torch.Tensor]): Output tensors
            x (torch.Tensor): Input tensor
            lora_a_stacked (tuple[torch.Tensor, ...]): lora_a's weights
            scale (float): Scaling factor for the operation
        """

        torch.ops.xla.dynamo_set_buffer_donor_(y, True)
        x = x.view(-1, x.shape[-1])

        for slice_idx in range(len(lora_a_stacked)):
            lora_s = lora_a_stacked[slice_idx]
            y_s = self.shrink(x, lora_s, scale)
            y[slice_idx, :, :] = y_s  # type: ignore[index]
        return y

    def add_expand(self,
                   y: torch.Tensor,
                   x: Union[tuple[torch.Tensor, ...], torch.Tensor],
                   lora_b_stacked: tuple[torch.Tensor, ...],
                   lora_bias_stacked: Optional[tuple[torch.Tensor, ...]],
                   output_slices: tuple[int, ...],
                   offset_start: int = 0,
                   add_inputs=True,
                   **kwargs) -> torch.Tensor:
        """
        Performs GEMM and bias addition for multiple slices of lora_b.

        Semantics:
            for i in range(len(lora_b_stacked)):
                slice = output_slices[i]
                y[:, offset:offset+slice] += x[i] @ lora_b_stacked[i] +
                    lora_bias_stacked[i]
                offset += slice

        Args:
            y (torch.Tensor): Output tensor.
            x (Union[tuple[torch.Tensor, ...], torch.Tensor]): Input tensors
            lora_b_stacked (tuple[torch.Tensor, ...]): lora_b's weight
            lora_bias_stacked (Optional[tuple[torch.Tensor, ...]]):
                bias's weight
            output_slices (tuple[int, ...]): Every slice's size
            add_inputs (bool):  Defaults to True.
        """
        y_org = y
        y = y.view(-1, y.shape[-1])
        offset_left = 0

        if lora_bias_stacked is not None:
            y = self._apply_bias(self._get_token_lora_indices(y), y,
                                 output_slices, lora_bias_stacked)
        for slice_idx in range(len(lora_b_stacked)):
            y = self.expand_slice(y,
                                  x[slice_idx],
                                  lora_b_stacked[slice_idx],
                                  offset_left,
                                  output_slices[slice_idx],
                                  add_inputs=add_inputs)
            offset_left += output_slices[slice_idx]
        return y.view_as(y_org)

    def add_lora_embedding(self,
                           y: torch.Tensor,
                           x: torch.Tensor,
                           lora_b_stacked: torch.Tensor,
                           add_inputs: bool = True,
                           **kwargs) -> torch.Tensor:
        """
        Applies lora  specifically for VocabParallelEmbeddingWithLoRA.

        Semantics:
            y += x @ lora_b_stacked

        Args:
            y (torch.Tensor): Output tensor.
            x (torch.Tensor): Input tensor.
            lora_b_stacked (torch.Tensor): lora_b's weights.
            add_inputs (bool): Default to True.
        """

        # Embedding layer only needs the expand op
        return self.expand(y, x, lora_b_stacked, add_inputs)

    def add_lora_linear(self,
                        y: torch.Tensor,
                        x: torch.Tensor,
                        lora_a_stacked: tuple[torch.Tensor, ...],
                        lora_b_stacked: tuple[torch.Tensor, ...],
                        lora_bias_stacked: Optional[tuple[torch.Tensor, ...]],
                        scale: float,
                        output_slices: tuple[int, ...],
                        *,
                        buffer: Optional[tuple[torch.Tensor, ...]] = None,
                        **kwargs) -> torch.Tensor:
        """
        Applicable to linear-related lora.

        Semantics:
            for i in range(len(lora_a_stacked)):
                y[i] += (
                    x[i].unsqueeze(0)
                    @ lora_a_stacked[indices[i], layer_idx, :, :]
                    @ lora_b_stacked[indices[i], layer_idx, :, :]
                    * scale
                    ).squeeze(0)+lora_bias_stacked[i]

        Args:
            y (torch.Tensor): Output tensor. Will not be changed in-place.
            x (torch.Tensor): Input tensor (T, E)
            lora_a_stacked (tuple[torch.Tensor, ...]): lora_a's weight.
            lora_b_stacked (tuple[torch.Tensor, ...]): lora_b's weight.
            lora_bias_stacked (Optional[tuple[torch.Tensor, ...]]): lora's bias.
            scale (float): Scaling factor.
            output_slices (tuple[int, ...]): Every slice's size.
            buffer (Optional[tuple[torch.Tensor, ...]]): Defaults to None.
        """

        assert len(lora_a_stacked) == len(lora_b_stacked) == len(output_slices)
        if lora_bias_stacked is not None:
            assert len(lora_bias_stacked) == len(output_slices)
            y = self._apply_bias(self._get_token_lora_indices(y), y,
                                 output_slices, lora_bias_stacked)

        if buffer is None:
            r = lora_b_stacked[0].size(-1)
            T = x.size(0)
            buffer = torch.zeros(
                (len(output_slices), T, r),
                dtype=x.dtype,
                device=x.device,
            )
        buffer = self.add_shrink(buffer, x, lora_a_stacked, scale, **kwargs)
        return self.add_expand(y,
                               buffer,
                               lora_b_stacked,
                               None,
                               output_slices,
                               add_inputs=True,
                               **kwargs)

    def add_lora_logits(self,
                        y: torch.Tensor,
                        x: torch.Tensor,
                        lora_a_stacked: torch.Tensor,
                        lora_b_stacked: torch.Tensor,
                        scale,
                        *,
                        buffer: Optional[torch.Tensor] = None,
                        **kwargs) -> torch.Tensor:
        """
        Applies lora specifically for LogitsProcessorWithLoRA.

        Semantics:
            buffer = (x @ lora_a_stacked) * scale
            y += buffer @ lora_b_stacked

        Args:
            y (torch.Tensor): Output tensor.
            x (torch.Tensor): Input tensor.
            lora_a_stacked (torch.Tensor): lora_a's weights.
            lora_b_stacked (torch.Tensor):lora_b's weights.
            scale (float): Scaling factor.
            buffer (Optional[torch.Tensor]):Default to None.
        """
        y_org = y
        y = y.view(-1, y.shape[-1])
        x = x.view(-1, x.shape[-1])

        sampler_indices = torch.narrow(self._sampler_indices, 0, 0, x.size(0))
        buffer = bgmv_shrink(x, lora_a_stacked, sampler_indices, scale)
        y = bgmv_expand(buffer,
                        lora_b_stacked,
                        y,
                        sampler_indices,
                        add_inputs=True)
        return y.view_as(y_org)

    def _apply_bias(
        self,
        indices: torch.Tensor,
        output: torch.Tensor,
        output_slices: tuple[int, ...],
        lora_bias_stacked: tuple[Optional[torch.Tensor], ...],
    ):
        """Applies bias to output

        Input shapes:
            lora_bias_stacked:      3 element tuple of (num_loras, output_dim)
            indices:           (batch_size)
            output:            (batch_size, q_slice_size + 2*kv_slice_size)
            output_slices:     n-1 element tuple of (slice_size...),
                            where n is number of slices
        """
        org_output = output
        output = output.view(-1, output.shape[-1])
        indices = indices.view(-1)

        offset_left = 0
        for slice_idx, slice in enumerate(output_slices):
            bias = lora_bias_stacked[slice_idx]
            if bias is not None:
                bias = bias.view(-1, bias.shape[-1])
                bias = bias[indices]
                bias = torch.where(indices[:, None] == -1, 0, bias)

                bias = F.pad(bias, (offset_left, output.shape[1] -
                                    (offset_left + slice), 0, 0))

                output += bias
            offset_left += slice

        return output.view_as(org_output)

    # This performs the same tensor ops as the base method, except it does them
    # on the CPU then transfers the results to the TPU
    def _update_base_metadata(
        self,
        mapping: "LoRAMapping",
        lora_index_to_id: list[Optional[int]],
        max_loras: int,
        vocab_size: int,
        extra_vocab_size: int,
        long_lora_context: Optional["LongContextLoRAContext"] = None,
    ):
        # Make sure we don't accidentally collect outside operations
        xm.mark_step()

        # Pad the prompt mapping to avoid running into recompiles on the TPU
        # TODO: Should this happen inside mapping internally? If so how can we
        # avoid having backend specific LoRAMapping classes?
        mapping.prompt_mapping = self._pad_prompt_mapping(
            mapping.prompt_mapping)

        (
            base_indices,
            sampler_indices,
            sampler_indices_padded,
            embeddings_indices,
            long_lora_offsets_tensor,
            indices_len,
        ) = convert_mapping(
            mapping,
            lora_index_to_id,
            max_loras,
            vocab_size,
            extra_vocab_size,
            "cpu",
            long_lora_context,
        )
        self._token_lora_indices = self._pad_to_shape(
            base_indices, self._token_lora_indices.shape,
            dims=1).to(self.device)
        self._sampler_indices = self._pad_to_shape(sampler_indices,
                                                   self._sampler_indices.shape,
                                                   dims=1).to(self.device)
        self._sampler_indices_padded = self._pad_to_shape(
            sampler_indices_padded, self._sampler_indices_padded.shape,
            dims=1).to(self.device)
        self._embeddings_indices = self._pad_to_shape(
            embeddings_indices, self._embeddings_indices.shape,
            dims=2).to(self.device)
        if long_lora_offsets_tensor is not None:
            self._long_lora_indices = self._pad_to_shape(
                long_lora_offsets_tensor,
                self._long_lora_indices.shape,
                dims=1).to(self.device)
        else:
            zeroed = torch.zeros_like(self._long_lora_indices.cpu(),
                                      dtype=torch.int32)
            self._long_lora_indices = zeroed.to(self.device)
        self.indices_len[:] = indices_len

    def _update_prefill_metadata(self,
                                 token_lora_tensor: torch.Tensor) -> None:
        self.batch_size = 1
        self._lora_indices_per_batch[:self.
                                     batch_size] = token_lora_tensor[:self.
                                                                     batch_size]

    def _pad_prompt_mapping(
            self, prompt_mapping: tuple[int, ...]) -> tuple[int, ...]:
        num_reqs = len(prompt_mapping)

        # From vllm/v1/worker/tpu_model_runner:51, but need to avoid a circular
        # import
        MIN_NUM_SEQS = 8

        padded_num_reqs = max(2**math.ceil(math.log2(num_reqs)), MIN_NUM_SEQS)
        pad_len = padded_num_reqs - num_reqs

        padding = [-1] * pad_len
        return tuple(list(prompt_mapping) + padding)

    def _pad_to_shape(self, src, target_shape, dims=1):
        if dims == 1:
            pad_len = target_shape[0] - src.shape[0]
            return F.pad(src, (0, pad_len), value=0).to(torch.int32)
        else:
            pad_rows = target_shape[0] - src.shape[0]
            pad_cols = target_shape[1] - src.shape[1]
            return F.pad(src, (0, pad_cols, 0, pad_rows),
                         value=0).to(torch.int32)

_sampler_indices instance-attribute 

_sampler_indices = to(dtype=int32)

_sampler_indices_padded instance-attribute 

_sampler_indices_padded = to(dtype=int32)

_token_lora_indices instance-attribute 

_token_lora_indices = to(dtype=int32)

embeddings_indices property 

embeddings_indices: Tensor

This property provides access to the indices used for lora embeddings, specifically for VocabParallelEmbeddingWithLoRA.

sampler_indices_padded property 

sampler_indices_padded: Tensor

This property provides access to padded sampler indices.

__init__ 

__init__(
    max_num_batched_tokens: int,
    max_batches: int,
    device: Union[device, str],
    **kwargs,
)
Source code in vllm/lora/punica_wrapper/punica_tpu.py 
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
def __init__(self, max_num_batched_tokens: int, max_batches: int,
             device: Union[torch.device, str], **kwargs):
    PunicaWrapperBase.__init__(self, max_num_batched_tokens, max_batches,
                               device)

    # PunicaWrapperBase defines some tensors with dtype=torch.int64, which
    # isn't supported by the TPU. So convert those tensors to int32.
    # Not all of them are used by the TPU so only convert the useful ones.
    self._token_lora_indices = self._token_lora_indices.to(
        dtype=torch.int32)
    self._sampler_indices = self._sampler_indices.to(dtype=torch.int32)
    self._sampler_indices_padded = self._sampler_indices_padded.to(
        dtype=torch.int32)

    torch.ops.xla.dynamo_set_buffer_donor_(self._token_lora_indices, True)
    torch.ops.xla.dynamo_set_buffer_donor_(self._sampler_indices, True)
    torch.ops.xla.dynamo_set_buffer_donor_(self._sampler_indices_padded,
                                           True)
    torch.ops.xla.dynamo_set_buffer_donor_(self._embeddings_indices, True)
    torch.ops.xla.dynamo_set_buffer_donor_(self._long_lora_indices, True)
    torch.ops.xla.dynamo_set_buffer_donor_(self._lora_indices_per_batch,
                                           True)

    torch._dynamo.mark_dynamic(self._token_lora_indices, 0)
    torch._dynamo.mark_dynamic(self._embeddings_indices, 1)
    torch._dynamo.mark_dynamic(self._sampler_indices_padded, 0)

_apply_bias 

_apply_bias(
    indices: Tensor,
    output: Tensor,
    output_slices: tuple[int, ...],
    lora_bias_stacked: tuple[Optional[Tensor], ...],
)

Applies bias to output

Input shapes

lora_bias_stacked: 3 element tuple of (num_loras, output_dim) indices: (batch_size) output: (batch_size, q_slice_size + 2*kv_slice_size) output_slices: n-1 element tuple of (slice_size...), where n is number of slices

Source code in vllm/lora/punica_wrapper/punica_tpu.py 
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
def _apply_bias(
    self,
    indices: torch.Tensor,
    output: torch.Tensor,
    output_slices: tuple[int, ...],
    lora_bias_stacked: tuple[Optional[torch.Tensor], ...],
):
    """Applies bias to output

    Input shapes:
        lora_bias_stacked:      3 element tuple of (num_loras, output_dim)
        indices:           (batch_size)
        output:            (batch_size, q_slice_size + 2*kv_slice_size)
        output_slices:     n-1 element tuple of (slice_size...),
                        where n is number of slices
    """
    org_output = output
    output = output.view(-1, output.shape[-1])
    indices = indices.view(-1)

    offset_left = 0
    for slice_idx, slice in enumerate(output_slices):
        bias = lora_bias_stacked[slice_idx]
        if bias is not None:
            bias = bias.view(-1, bias.shape[-1])
            bias = bias[indices]
            bias = torch.where(indices[:, None] == -1, 0, bias)

            bias = F.pad(bias, (offset_left, output.shape[1] -
                                (offset_left + slice), 0, 0))

            output += bias
        offset_left += slice

    return output.view_as(org_output)

_get_token_lora_indices 

_get_token_lora_indices(x: Tensor) -> IntTensor
Source code in vllm/lora/punica_wrapper/punica_tpu.py 
56
57
def _get_token_lora_indices(self, x: torch.Tensor) -> torch.IntTensor:
    return torch.narrow(self._token_lora_indices, 0, 0, x.size(0))

_pad_prompt_mapping 

_pad_prompt_mapping(
    prompt_mapping: tuple[int, ...],
) -> tuple[int, ...]
Source code in vllm/lora/punica_wrapper/punica_tpu.py 
383
384
385
386
387
388
389
390
391
392
393
394
395
def _pad_prompt_mapping(
        self, prompt_mapping: tuple[int, ...]) -> tuple[int, ...]:
    num_reqs = len(prompt_mapping)

    # From vllm/v1/worker/tpu_model_runner:51, but need to avoid a circular
    # import
    MIN_NUM_SEQS = 8

    padded_num_reqs = max(2**math.ceil(math.log2(num_reqs)), MIN_NUM_SEQS)
    pad_len = padded_num_reqs - num_reqs

    padding = [-1] * pad_len
    return tuple(list(prompt_mapping) + padding)

_pad_to_shape 

_pad_to_shape(src, target_shape, dims=1)
Source code in vllm/lora/punica_wrapper/punica_tpu.py 
397
398
399
400
401
402
403
404
405
def _pad_to_shape(self, src, target_shape, dims=1):
    if dims == 1:
        pad_len = target_shape[0] - src.shape[0]
        return F.pad(src, (0, pad_len), value=0).to(torch.int32)
    else:
        pad_rows = target_shape[0] - src.shape[0]
        pad_cols = target_shape[1] - src.shape[1]
        return F.pad(src, (0, pad_cols, 0, pad_rows),
                     value=0).to(torch.int32)

_update_base_metadata 

_update_base_metadata(
    mapping: LoRAMapping,
    lora_index_to_id: list[Optional[int]],
    max_loras: int,
    vocab_size: int,
    extra_vocab_size: int,
    long_lora_context: Optional[
        LongContextLoRAContext
    ] = None,
)
Source code in vllm/lora/punica_wrapper/punica_tpu.py 
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
def _update_base_metadata(
    self,
    mapping: "LoRAMapping",
    lora_index_to_id: list[Optional[int]],
    max_loras: int,
    vocab_size: int,
    extra_vocab_size: int,
    long_lora_context: Optional["LongContextLoRAContext"] = None,
):
    # Make sure we don't accidentally collect outside operations
    xm.mark_step()

    # Pad the prompt mapping to avoid running into recompiles on the TPU
    # TODO: Should this happen inside mapping internally? If so how can we
    # avoid having backend specific LoRAMapping classes?
    mapping.prompt_mapping = self._pad_prompt_mapping(
        mapping.prompt_mapping)

    (
        base_indices,
        sampler_indices,
        sampler_indices_padded,
        embeddings_indices,
        long_lora_offsets_tensor,
        indices_len,
    ) = convert_mapping(
        mapping,
        lora_index_to_id,
        max_loras,
        vocab_size,
        extra_vocab_size,
        "cpu",
        long_lora_context,
    )
    self._token_lora_indices = self._pad_to_shape(
        base_indices, self._token_lora_indices.shape,
        dims=1).to(self.device)
    self._sampler_indices = self._pad_to_shape(sampler_indices,
                                               self._sampler_indices.shape,
                                               dims=1).to(self.device)
    self._sampler_indices_padded = self._pad_to_shape(
        sampler_indices_padded, self._sampler_indices_padded.shape,
        dims=1).to(self.device)
    self._embeddings_indices = self._pad_to_shape(
        embeddings_indices, self._embeddings_indices.shape,
        dims=2).to(self.device)
    if long_lora_offsets_tensor is not None:
        self._long_lora_indices = self._pad_to_shape(
            long_lora_offsets_tensor,
            self._long_lora_indices.shape,
            dims=1).to(self.device)
    else:
        zeroed = torch.zeros_like(self._long_lora_indices.cpu(),
                                  dtype=torch.int32)
        self._long_lora_indices = zeroed.to(self.device)
    self.indices_len[:] = indices_len

_update_prefill_metadata 

_update_prefill_metadata(token_lora_tensor: Tensor) -> None
Source code in vllm/lora/punica_wrapper/punica_tpu.py 
376
377
378
379
380
381
def _update_prefill_metadata(self,
                             token_lora_tensor: torch.Tensor) -> None:
    self.batch_size = 1
    self._lora_indices_per_batch[:self.
                                 batch_size] = token_lora_tensor[:self.
                                                                 batch_size]

add_expand 

add_expand(
    y: Tensor,
    x: Union[tuple[Tensor, ...], Tensor],
    lora_b_stacked: tuple[Tensor, ...],
    lora_bias_stacked: Optional[tuple[Tensor, ...]],
    output_slices: tuple[int, ...],
    offset_start: int = 0,
    add_inputs=True,
    **kwargs,
) -> Tensor

Performs GEMM and bias addition for multiple slices of lora_b.

Semantics

for i in range(len(lora_b_stacked)): slice = output_slices[i] y[:, offset:offset+slice] += x[i] @ lora_b_stacked[i] + lora_bias_stacked[i] offset += slice

Parameters:

Name Type Description Default
y Tensor

Output tensor.

 required
x Union[tuple[Tensor, ...], Tensor]

Input tensors

 required
lora_b_stacked tuple[Tensor, ...]

lora_b's weight

 required
lora_bias_stacked Optional[tuple[Tensor, ...]]

bias's weight

 required
output_slices tuple[int, ...]

Every slice's size

 required
add_inputs bool

Defaults to True.

 True
Source code in vllm/lora/punica_wrapper/punica_tpu.py 
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
def add_expand(self,
               y: torch.Tensor,
               x: Union[tuple[torch.Tensor, ...], torch.Tensor],
               lora_b_stacked: tuple[torch.Tensor, ...],
               lora_bias_stacked: Optional[tuple[torch.Tensor, ...]],
               output_slices: tuple[int, ...],
               offset_start: int = 0,
               add_inputs=True,
               **kwargs) -> torch.Tensor:
    """
    Performs GEMM and bias addition for multiple slices of lora_b.

    Semantics:
        for i in range(len(lora_b_stacked)):
            slice = output_slices[i]
            y[:, offset:offset+slice] += x[i] @ lora_b_stacked[i] +
                lora_bias_stacked[i]
            offset += slice

    Args:
        y (torch.Tensor): Output tensor.
        x (Union[tuple[torch.Tensor, ...], torch.Tensor]): Input tensors
        lora_b_stacked (tuple[torch.Tensor, ...]): lora_b's weight
        lora_bias_stacked (Optional[tuple[torch.Tensor, ...]]):
            bias's weight
        output_slices (tuple[int, ...]): Every slice's size
        add_inputs (bool):  Defaults to True.
    """
    y_org = y
    y = y.view(-1, y.shape[-1])
    offset_left = 0

    if lora_bias_stacked is not None:
        y = self._apply_bias(self._get_token_lora_indices(y), y,
                             output_slices, lora_bias_stacked)
    for slice_idx in range(len(lora_b_stacked)):
        y = self.expand_slice(y,
                              x[slice_idx],
                              lora_b_stacked[slice_idx],
                              offset_left,
                              output_slices[slice_idx],
                              add_inputs=add_inputs)
        offset_left += output_slices[slice_idx]
    return y.view_as(y_org)

add_lora_embedding 

add_lora_embedding(
    y: Tensor,
    x: Tensor,
    lora_b_stacked: Tensor,
    add_inputs: bool = True,
    **kwargs,
) -> Tensor

Applies lora specifically for VocabParallelEmbeddingWithLoRA.

Semantics

y += x @ lora_b_stacked

Parameters:

Name Type Description Default
y Tensor

Output tensor.

 required
x Tensor

Input tensor.

 required
lora_b_stacked Tensor

lora_b's weights.

 required
add_inputs bool

Default to True.

 True
Source code in vllm/lora/punica_wrapper/punica_tpu.py 
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
def add_lora_embedding(self,
                       y: torch.Tensor,
                       x: torch.Tensor,
                       lora_b_stacked: torch.Tensor,
                       add_inputs: bool = True,
                       **kwargs) -> torch.Tensor:
    """
    Applies lora  specifically for VocabParallelEmbeddingWithLoRA.

    Semantics:
        y += x @ lora_b_stacked

    Args:
        y (torch.Tensor): Output tensor.
        x (torch.Tensor): Input tensor.
        lora_b_stacked (torch.Tensor): lora_b's weights.
        add_inputs (bool): Default to True.
    """

    # Embedding layer only needs the expand op
    return self.expand(y, x, lora_b_stacked, add_inputs)

add_lora_linear 

add_lora_linear(
    y: Tensor,
    x: Tensor,
    lora_a_stacked: tuple[Tensor, ...],
    lora_b_stacked: tuple[Tensor, ...],
    lora_bias_stacked: Optional[tuple[Tensor, ...]],
    scale: float,
    output_slices: tuple[int, ...],
    *,
    buffer: Optional[tuple[Tensor, ...]] = None,
    **kwargs,
) -> Tensor

Applicable to linear-related lora.

Semantics

for i in range(len(lora_a_stacked)): y[i] += ( x[i].unsqueeze(0) @ lora_a_stacked[indices[i], layer_idx, :, :] @ lora_b_stacked[indices[i], layer_idx, :, :] * scale ).squeeze(0)+lora_bias_stacked[i]

Parameters:

Name Type Description Default
y Tensor

Output tensor. Will not be changed in-place.

 required
x Tensor

Input tensor (T, E)

 required
lora_a_stacked tuple[Tensor, ...]

lora_a's weight.

 required
lora_b_stacked tuple[Tensor, ...]

lora_b's weight.

 required
lora_bias_stacked Optional[tuple[Tensor, ...]]

lora's bias.

 required
scale float

Scaling factor.

 required
output_slices tuple[int, ...]

Every slice's size.

 required
buffer Optional[tuple[Tensor, ...]]

Defaults to None.

 None
Source code in vllm/lora/punica_wrapper/punica_tpu.py 
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
def add_lora_linear(self,
                    y: torch.Tensor,
                    x: torch.Tensor,
                    lora_a_stacked: tuple[torch.Tensor, ...],
                    lora_b_stacked: tuple[torch.Tensor, ...],
                    lora_bias_stacked: Optional[tuple[torch.Tensor, ...]],
                    scale: float,
                    output_slices: tuple[int, ...],
                    *,
                    buffer: Optional[tuple[torch.Tensor, ...]] = None,
                    **kwargs) -> torch.Tensor:
    """
    Applicable to linear-related lora.

    Semantics:
        for i in range(len(lora_a_stacked)):
            y[i] += (
                x[i].unsqueeze(0)
                @ lora_a_stacked[indices[i], layer_idx, :, :]
                @ lora_b_stacked[indices[i], layer_idx, :, :]
                * scale
                ).squeeze(0)+lora_bias_stacked[i]

    Args:
        y (torch.Tensor): Output tensor. Will not be changed in-place.
        x (torch.Tensor): Input tensor (T, E)
        lora_a_stacked (tuple[torch.Tensor, ...]): lora_a's weight.
        lora_b_stacked (tuple[torch.Tensor, ...]): lora_b's weight.
        lora_bias_stacked (Optional[tuple[torch.Tensor, ...]]): lora's bias.
        scale (float): Scaling factor.
        output_slices (tuple[int, ...]): Every slice's size.
        buffer (Optional[tuple[torch.Tensor, ...]]): Defaults to None.
    """

    assert len(lora_a_stacked) == len(lora_b_stacked) == len(output_slices)
    if lora_bias_stacked is not None:
        assert len(lora_bias_stacked) == len(output_slices)
        y = self._apply_bias(self._get_token_lora_indices(y), y,
                             output_slices, lora_bias_stacked)

    if buffer is None:
        r = lora_b_stacked[0].size(-1)
        T = x.size(0)
        buffer = torch.zeros(
            (len(output_slices), T, r),
            dtype=x.dtype,
            device=x.device,
        )
    buffer = self.add_shrink(buffer, x, lora_a_stacked, scale, **kwargs)
    return self.add_expand(y,
                           buffer,
                           lora_b_stacked,
                           None,
                           output_slices,
                           add_inputs=True,
                           **kwargs)

add_lora_logits 

add_lora_logits(
    y: Tensor,
    x: Tensor,
    lora_a_stacked: Tensor,
    lora_b_stacked: Tensor,
    scale,
    *,
    buffer: Optional[Tensor] = None,
    **kwargs,
) -> Tensor

Applies lora specifically for LogitsProcessorWithLoRA.

Semantics

buffer = (x @ lora_a_stacked) * scale y += buffer @ lora_b_stacked

Parameters:

Name Type Description Default
y Tensor

Output tensor.

 required
x Tensor

Input tensor.

 required
lora_a_stacked Tensor

lora_a's weights.

 required
lora_b_stacked Tensor

lora_b's weights.

 required
scale float

Scaling factor.

 required
buffer Optional[Tensor]

Default to None.

 None
Source code in vllm/lora/punica_wrapper/punica_tpu.py 
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
def add_lora_logits(self,
                    y: torch.Tensor,
                    x: torch.Tensor,
                    lora_a_stacked: torch.Tensor,
                    lora_b_stacked: torch.Tensor,
                    scale,
                    *,
                    buffer: Optional[torch.Tensor] = None,
                    **kwargs) -> torch.Tensor:
    """
    Applies lora specifically for LogitsProcessorWithLoRA.

    Semantics:
        buffer = (x @ lora_a_stacked) * scale
        y += buffer @ lora_b_stacked

    Args:
        y (torch.Tensor): Output tensor.
        x (torch.Tensor): Input tensor.
        lora_a_stacked (torch.Tensor): lora_a's weights.
        lora_b_stacked (torch.Tensor):lora_b's weights.
        scale (float): Scaling factor.
        buffer (Optional[torch.Tensor]):Default to None.
    """
    y_org = y
    y = y.view(-1, y.shape[-1])
    x = x.view(-1, x.shape[-1])

    sampler_indices = torch.narrow(self._sampler_indices, 0, 0, x.size(0))
    buffer = bgmv_shrink(x, lora_a_stacked, sampler_indices, scale)
    y = bgmv_expand(buffer,
                    lora_b_stacked,
                    y,
                    sampler_indices,
                    add_inputs=True)
    return y.view_as(y_org)

add_shrink 

add_shrink(
    y: Union[tuple[Tensor, ...], Tensor],
    x: Tensor,
    lora_a_stacked: tuple[Tensor, ...],
    scale: float,
    **kwargs,
) -> Optional[Tensor]

Performs GEMM for multiple slices of lora_a.

Semantics: for i in range(len(lora_a_stacked)): y[i] += (x @ lora_a_stacked[i]) * scale

Parameters:

Name Type Description Default
y Union[tuple[Tensor, ...], Tensor]

Output tensors

 required
x Tensor

Input tensor

 required
lora_a_stacked tuple[Tensor, ...]

lora_a's weights

 required
scale float

Scaling factor for the operation

 required
Source code in vllm/lora/punica_wrapper/punica_tpu.py 
 94
 95
 96
 97
 98
 99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
def add_shrink(self, y: Union[tuple[torch.Tensor, ...], torch.Tensor],
               x: torch.Tensor, lora_a_stacked: tuple[torch.Tensor, ...],
               scale: float, **kwargs) -> Optional[torch.Tensor]:
    """
    Performs GEMM for multiple slices of lora_a.

    Semantics:
    for i in range(len(lora_a_stacked)):
        y[i] += (x @ lora_a_stacked[i]) * scale

    Args:
        y (Union[tuple[torch.Tensor, ...], torch.Tensor]): Output tensors
        x (torch.Tensor): Input tensor
        lora_a_stacked (tuple[torch.Tensor, ...]): lora_a's weights
        scale (float): Scaling factor for the operation
    """

    torch.ops.xla.dynamo_set_buffer_donor_(y, True)
    x = x.view(-1, x.shape[-1])

    for slice_idx in range(len(lora_a_stacked)):
        lora_s = lora_a_stacked[slice_idx]
        y_s = self.shrink(x, lora_s, scale)
        y[slice_idx, :, :] = y_s  # type: ignore[index]
    return y

expand 

expand(
    y: Tensor, x: Tensor, w_t_all: Tensor, add_inputs: bool
)
Source code in vllm/lora/punica_wrapper/punica_tpu.py 
82
83
84
85
def expand(self, y: torch.Tensor, x: torch.Tensor, w_t_all: torch.Tensor,
           add_inputs: bool):
    return bgmv_expand(x, w_t_all, y, self._get_token_lora_indices(x),
                       add_inputs)

expand_slice 

expand_slice(
    y: Tensor,
    x: Tensor,
    w_t_all: Tensor,
    y_offset: int,
    y_slice_size: int,
    add_inputs: bool,
) -> Tensor
Source code in vllm/lora/punica_wrapper/punica_tpu.py 
87
88
89
90
91
92
def expand_slice(self, y: torch.Tensor, x: torch.Tensor,
                 w_t_all: torch.Tensor, y_offset: int, y_slice_size: int,
                 add_inputs: bool) -> torch.Tensor:
    return bgmv_expand_slice(x, w_t_all, y,
                             self._get_token_lora_indices(x), y_offset,
                             y_slice_size, add_inputs)

shrink 

shrink(x: Tensor, w_t_all: Tensor, scale: float)
Source code in vllm/lora/punica_wrapper/punica_tpu.py 
74
75
76
77
78
79
80
def shrink(
    self,
    x: torch.Tensor,
    w_t_all: torch.Tensor,
    scale: float,
):
    return bgmv_shrink(x, w_t_all, self._get_token_lora_indices(x), scale)