跳到内容

vllm_gaudi.ops.hpu_rotary_embedding

HPUDeepseekScalingRotaryEmbedding

基类: DeepseekScalingRotaryEmbedding

源文件在 vllm_gaudi/ops/hpu_rotary_embedding.py 
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
@DeepseekScalingRotaryEmbedding.register_oot
class HPUDeepseekScalingRotaryEmbedding(DeepseekScalingRotaryEmbedding):

    def prepare_cos_sin(self,
                        positions: torch.Tensor,
                        offsets: Optional[torch.Tensor] = None,
                        recompute_cos_sin: bool = False):
        self.recompute_cos_sin = recompute_cos_sin
        if offsets is not None:
            offsets = offsets.view(positions.shape[0], -1)
            positions = positions + offsets
        positions = positions.flatten()
        num_tokens = positions.shape[0]
        cos_sin = self.cos_sin_cache.index_select(0, positions).view(num_tokens, 1, -1)
        cos, sin = cos_sin.chunk(2, dim=-1)
        if self.is_neox_style:
            cos = torch.cat((cos, cos), dim=-1)
            sin = torch.cat((sin, sin), dim=-1)
        else:
            sin = torch.repeat_interleave(sin, 2, dim=-1, output_size=cos_sin.shape[-1])
            cos = torch.repeat_interleave(cos, 2, dim=-1, output_size=cos_sin.shape[-1])
        self.register_buffer("cos", cos, persistent=False)
        self.register_buffer("sin", sin, persistent=False)

    def forward(
        self,
        positions: torch.Tensor,
        query: torch.Tensor,
        key: torch.Tensor,
        offsets: Optional[torch.Tensor] = None,
    ) -> tuple[torch.Tensor, torch.Tensor]:
        from habana_frameworks.torch.hpex.kernels import (RotaryPosEmbeddingMode, apply_rotary_pos_emb)

        # Prepare cos-sin caches for long-context + LoRA with offsets for every
        # forward, since the offset information wasn't available previously
        if not hasattr(self, "sin") or self.recompute_cos_sin:
            self.prepare_cos_sin(positions, offsets, recompute_cos_sin=True)
        if hasattr(self, "scaling_factors") or hasattr(self, "scaling_factor") or self.sin is None:
            self.prepare_cos_sin(positions, offsets)
        num_tokens = positions.shape[0] * positions.shape[1]

        # deepseek_v2 MLA attention did an unsqueeze on key due to assumption on
        # GPU with (num_tokens, hidden_size)
        if key.dim() == 4:
            key = key.squeeze(1)
        # HPU RoPE kernel requires hidden dimension for cos and sin to be equal
        # to query hidden dimension, so the original tensors need to be
        # expanded
        # GPT-NeoX kernel requires position_ids = None, offset, mode = BLOCKWISE
        # and expansion of cos/sin tensors via concatenation
        # GPT-J kernel requires position_ids = None, offset = 0, mode = PAIRWISE
        # and expansion of cos/sin tensors via repeat_interleave
        rope_mode: RotaryPosEmbeddingMode
        rope_mode = RotaryPosEmbeddingMode.BLOCKWISE if self.is_neox_style else RotaryPosEmbeddingMode.PAIRWISE
        sin = self.sin
        cos = self.cos
        query_shape = query.shape
        key_shape = key.shape
        query = query.view(num_tokens, -1, self.head_size)
        key = key.view(num_tokens, -1, self.head_size)

        if self.head_size == self.rotary_dim:
            # Avoid unnecessary slicing and concatenation
            query = apply_rotary_pos_emb(query, cos, sin, None, 0, rope_mode)
            key = apply_rotary_pos_emb(key, cos, sin, None, 0, rope_mode)
            return query.reshape(query_shape), key.reshape(key_shape)

        query_rot = query[..., :self.rotary_dim]
        query_pass = query[..., self.rotary_dim:]
        query_rot = apply_rotary_pos_emb(query_rot, cos, sin, None, 0, rope_mode)
        query = torch.cat((query_rot, query_pass), dim=-1).reshape(query_shape)

        key_rot = key[..., :self.rotary_dim]
        key_pass = key[..., self.rotary_dim:]
        key_rot = apply_rotary_pos_emb(key_rot, cos, sin, None, 0, rope_mode)
        key = torch.cat((key_rot, key_pass), dim=-1).reshape(key_shape)
        return query, key

    def forward_oot(
        self,
        positions: torch.Tensor,
        query: torch.Tensor,
        key: torch.Tensor,
        offsets: Optional[torch.Tensor] = None,
    ) -> tuple[torch.Tensor, torch.Tensor]:
        # This method is not used in DeepseekScalingRotaryEmbedding
        return self.forward(positions, query, key, offsets)

forward 

forward(
    positions: Tensor,
    query: Tensor,
    key: Tensor,
    offsets: Optional[Tensor] = None,
) -> tuple[Tensor, Tensor]
源文件在 vllm_gaudi/ops/hpu_rotary_embedding.py 
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
def forward(
    self,
    positions: torch.Tensor,
    query: torch.Tensor,
    key: torch.Tensor,
    offsets: Optional[torch.Tensor] = None,
) -> tuple[torch.Tensor, torch.Tensor]:
    from habana_frameworks.torch.hpex.kernels import (RotaryPosEmbeddingMode, apply_rotary_pos_emb)

    # Prepare cos-sin caches for long-context + LoRA with offsets for every
    # forward, since the offset information wasn't available previously
    if not hasattr(self, "sin") or self.recompute_cos_sin:
        self.prepare_cos_sin(positions, offsets, recompute_cos_sin=True)
    if hasattr(self, "scaling_factors") or hasattr(self, "scaling_factor") or self.sin is None:
        self.prepare_cos_sin(positions, offsets)
    num_tokens = positions.shape[0] * positions.shape[1]

    # deepseek_v2 MLA attention did an unsqueeze on key due to assumption on
    # GPU with (num_tokens, hidden_size)
    if key.dim() == 4:
        key = key.squeeze(1)
    # HPU RoPE kernel requires hidden dimension for cos and sin to be equal
    # to query hidden dimension, so the original tensors need to be
    # expanded
    # GPT-NeoX kernel requires position_ids = None, offset, mode = BLOCKWISE
    # and expansion of cos/sin tensors via concatenation
    # GPT-J kernel requires position_ids = None, offset = 0, mode = PAIRWISE
    # and expansion of cos/sin tensors via repeat_interleave
    rope_mode: RotaryPosEmbeddingMode
    rope_mode = RotaryPosEmbeddingMode.BLOCKWISE if self.is_neox_style else RotaryPosEmbeddingMode.PAIRWISE
    sin = self.sin
    cos = self.cos
    query_shape = query.shape
    key_shape = key.shape
    query = query.view(num_tokens, -1, self.head_size)
    key = key.view(num_tokens, -1, self.head_size)

    if self.head_size == self.rotary_dim:
        # Avoid unnecessary slicing and concatenation
        query = apply_rotary_pos_emb(query, cos, sin, None, 0, rope_mode)
        key = apply_rotary_pos_emb(key, cos, sin, None, 0, rope_mode)
        return query.reshape(query_shape), key.reshape(key_shape)

    query_rot = query[..., :self.rotary_dim]
    query_pass = query[..., self.rotary_dim:]
    query_rot = apply_rotary_pos_emb(query_rot, cos, sin, None, 0, rope_mode)
    query = torch.cat((query_rot, query_pass), dim=-1).reshape(query_shape)

    key_rot = key[..., :self.rotary_dim]
    key_pass = key[..., self.rotary_dim:]
    key_rot = apply_rotary_pos_emb(key_rot, cos, sin, None, 0, rope_mode)
    key = torch.cat((key_rot, key_pass), dim=-1).reshape(key_shape)
    return query, key

forward_oot 

forward_oot(
    positions: Tensor,
    query: Tensor,
    key: Tensor,
    offsets: Optional[Tensor] = None,
) -> tuple[Tensor, Tensor]
源文件在 vllm_gaudi/ops/hpu_rotary_embedding.py 
388
389
390
391
392
393
394
395
396
def forward_oot(
    self,
    positions: torch.Tensor,
    query: torch.Tensor,
    key: torch.Tensor,
    offsets: Optional[torch.Tensor] = None,
) -> tuple[torch.Tensor, torch.Tensor]:
    # This method is not used in DeepseekScalingRotaryEmbedding
    return self.forward(positions, query, key, offsets)

prepare_cos_sin 

prepare_cos_sin(
    positions: Tensor,
    offsets: Optional[Tensor] = None,
    recompute_cos_sin: bool = False,
)
源文件在 vllm_gaudi/ops/hpu_rotary_embedding.py 
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
def prepare_cos_sin(self,
                    positions: torch.Tensor,
                    offsets: Optional[torch.Tensor] = None,
                    recompute_cos_sin: bool = False):
    self.recompute_cos_sin = recompute_cos_sin
    if offsets is not None:
        offsets = offsets.view(positions.shape[0], -1)
        positions = positions + offsets
    positions = positions.flatten()
    num_tokens = positions.shape[0]
    cos_sin = self.cos_sin_cache.index_select(0, positions).view(num_tokens, 1, -1)
    cos, sin = cos_sin.chunk(2, dim=-1)
    if self.is_neox_style:
        cos = torch.cat((cos, cos), dim=-1)
        sin = torch.cat((sin, sin), dim=-1)
    else:
        sin = torch.repeat_interleave(sin, 2, dim=-1, output_size=cos_sin.shape[-1])
        cos = torch.repeat_interleave(cos, 2, dim=-1, output_size=cos_sin.shape[-1])
    self.register_buffer("cos", cos, persistent=False)
    self.register_buffer("sin", sin, persistent=False)

HPUDynamicNTKScalingRotaryEmbedding

基类: DynamicNTKScalingRotaryEmbedding

源文件在 vllm_gaudi/ops/hpu_rotary_embedding.py 
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
@DynamicNTKScalingRotaryEmbedding.register_oot
class HPUDynamicNTKScalingRotaryEmbedding(DynamicNTKScalingRotaryEmbedding):

    def prepare_cos_sin(self,
                        positions: torch.Tensor,
                        offsets: Optional[torch.Tensor] = None,
                        recompute_cos_sin: bool = False):
        self.recompute_cos_sin = recompute_cos_sin
        if offsets is not None:
            offsets = offsets.view(positions.shape[0], -1)
            positions = positions + offsets
        positions = positions.flatten()
        num_tokens = positions.shape[0]
        cos_sin = self.cos_sin_cache.index_select(0, positions).view(num_tokens, 1, -1)
        cos, sin = cos_sin.chunk(2, dim=-1)
        if self.is_neox_style:
            cos = torch.cat((cos, cos), dim=-1)
            sin = torch.cat((sin, sin), dim=-1)
        else:
            sin = torch.repeat_interleave(sin, 2, dim=-1, output_size=cos_sin.shape[-1])
            cos = torch.repeat_interleave(cos, 2, dim=-1, output_size=cos_sin.shape[-1])
        self.register_buffer("cos", cos, persistent=False)
        self.register_buffer("sin", sin, persistent=False)

    def forward_oot(
        self,
        positions: torch.Tensor,
        query: torch.Tensor,
        key: torch.Tensor,
        offsets: Optional[torch.Tensor] = None,
    ) -> tuple[torch.Tensor, torch.Tensor]:
        from habana_frameworks.torch.hpex.kernels import (RotaryPosEmbeddingMode, apply_rotary_pos_emb)

        # Prepare cos-sin caches for long-context + LoRA with offsets for every
        # forward, since the offset information wasn't available previously
        if not hasattr(self, "sin") or self.recompute_cos_sin:
            self.prepare_cos_sin(positions, offsets, recompute_cos_sin=True)
        if hasattr(self, "scaling_factors") or hasattr(self, "scaling_factor") or self.sin is None:
            self.prepare_cos_sin(positions, offsets)
        num_tokens = positions.shape[0] * positions.shape[1]
        # HPU RoPE kernel requires hidden dimension for cos and sin to be equal
        # to query hidden dimension, so the original tensors need to be
        # expanded
        # GPT-NeoX kernel requires position_ids = None, offset, mode = BLOCKWISE
        # and expansion of cos/sin tensors via concatenation
        # GPT-J kernel requires position_ids = None, offset = 0, mode = PAIRWISE
        # and expansion of cos/sin tensors via repeat_interleave
        rope_mode: RotaryPosEmbeddingMode
        rope_mode = RotaryPosEmbeddingMode.BLOCKWISE if self.is_neox_style else RotaryPosEmbeddingMode.PAIRWISE
        sin = self.sin
        cos = self.cos
        query_shape = query.shape
        key_shape = key.shape
        query = query.view(num_tokens, -1, self.head_size)
        key = key.view(num_tokens, -1, self.head_size)

        if self.head_size == self.rotary_dim:
            # Avoid unnecessary slicing and concatenation
            query = apply_rotary_pos_emb(query, cos, sin, None, 0, rope_mode)
            key = apply_rotary_pos_emb(key, cos, sin, None, 0, rope_mode)
            return query.reshape(query_shape), key.reshape(key_shape)

        query_rot = query[..., :self.rotary_dim]
        query_pass = query[..., self.rotary_dim:]
        query_rot = apply_rotary_pos_emb(query_rot, cos, sin, None, 0, rope_mode)
        query = torch.cat((query_rot, query_pass), dim=-1).reshape(query_shape)

        key_rot = key[..., :self.rotary_dim]
        key_pass = key[..., self.rotary_dim:]
        key_rot = apply_rotary_pos_emb(key_rot, cos, sin, None, 0, rope_mode)
        key = torch.cat((key_rot, key_pass), dim=-1).reshape(key_shape)
        return query, key

forward_oot 

forward_oot(
    positions: Tensor,
    query: Tensor,
    key: Tensor,
    offsets: Optional[Tensor] = None,
) -> tuple[Tensor, Tensor]
源文件在 vllm_gaudi/ops/hpu_rotary_embedding.py 
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
def forward_oot(
    self,
    positions: torch.Tensor,
    query: torch.Tensor,
    key: torch.Tensor,
    offsets: Optional[torch.Tensor] = None,
) -> tuple[torch.Tensor, torch.Tensor]:
    from habana_frameworks.torch.hpex.kernels import (RotaryPosEmbeddingMode, apply_rotary_pos_emb)

    # Prepare cos-sin caches for long-context + LoRA with offsets for every
    # forward, since the offset information wasn't available previously
    if not hasattr(self, "sin") or self.recompute_cos_sin:
        self.prepare_cos_sin(positions, offsets, recompute_cos_sin=True)
    if hasattr(self, "scaling_factors") or hasattr(self, "scaling_factor") or self.sin is None:
        self.prepare_cos_sin(positions, offsets)
    num_tokens = positions.shape[0] * positions.shape[1]
    # HPU RoPE kernel requires hidden dimension for cos and sin to be equal
    # to query hidden dimension, so the original tensors need to be
    # expanded
    # GPT-NeoX kernel requires position_ids = None, offset, mode = BLOCKWISE
    # and expansion of cos/sin tensors via concatenation
    # GPT-J kernel requires position_ids = None, offset = 0, mode = PAIRWISE
    # and expansion of cos/sin tensors via repeat_interleave
    rope_mode: RotaryPosEmbeddingMode
    rope_mode = RotaryPosEmbeddingMode.BLOCKWISE if self.is_neox_style else RotaryPosEmbeddingMode.PAIRWISE
    sin = self.sin
    cos = self.cos
    query_shape = query.shape
    key_shape = key.shape
    query = query.view(num_tokens, -1, self.head_size)
    key = key.view(num_tokens, -1, self.head_size)

    if self.head_size == self.rotary_dim:
        # Avoid unnecessary slicing and concatenation
        query = apply_rotary_pos_emb(query, cos, sin, None, 0, rope_mode)
        key = apply_rotary_pos_emb(key, cos, sin, None, 0, rope_mode)
        return query.reshape(query_shape), key.reshape(key_shape)

    query_rot = query[..., :self.rotary_dim]
    query_pass = query[..., self.rotary_dim:]
    query_rot = apply_rotary_pos_emb(query_rot, cos, sin, None, 0, rope_mode)
    query = torch.cat((query_rot, query_pass), dim=-1).reshape(query_shape)

    key_rot = key[..., :self.rotary_dim]
    key_pass = key[..., self.rotary_dim:]
    key_rot = apply_rotary_pos_emb(key_rot, cos, sin, None, 0, rope_mode)
    key = torch.cat((key_rot, key_pass), dim=-1).reshape(key_shape)
    return query, key

prepare_cos_sin 

prepare_cos_sin(
    positions: Tensor,
    offsets: Optional[Tensor] = None,
    recompute_cos_sin: bool = False,
)
源文件在 vllm_gaudi/ops/hpu_rotary_embedding.py 
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
def prepare_cos_sin(self,
                    positions: torch.Tensor,
                    offsets: Optional[torch.Tensor] = None,
                    recompute_cos_sin: bool = False):
    self.recompute_cos_sin = recompute_cos_sin
    if offsets is not None:
        offsets = offsets.view(positions.shape[0], -1)
        positions = positions + offsets
    positions = positions.flatten()
    num_tokens = positions.shape[0]
    cos_sin = self.cos_sin_cache.index_select(0, positions).view(num_tokens, 1, -1)
    cos, sin = cos_sin.chunk(2, dim=-1)
    if self.is_neox_style:
        cos = torch.cat((cos, cos), dim=-1)
        sin = torch.cat((sin, sin), dim=-1)
    else:
        sin = torch.repeat_interleave(sin, 2, dim=-1, output_size=cos_sin.shape[-1])
        cos = torch.repeat_interleave(cos, 2, dim=-1, output_size=cos_sin.shape[-1])
    self.register_buffer("cos", cos, persistent=False)
    self.register_buffer("sin", sin, persistent=False)

HPULinearScalingRotaryEmbedding

基类: LinearScalingRotaryEmbedding

源文件在 vllm_gaudi/ops/hpu_rotary_embedding.py 
 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
@LinearScalingRotaryEmbedding.register_oot
class HPULinearScalingRotaryEmbedding(LinearScalingRotaryEmbedding):

    def prepare_cos_sin(self,
                        positions: torch.Tensor,
                        offsets: Optional[torch.Tensor] = None,
                        recompute_cos_sin: bool = False):
        self.recompute_cos_sin = recompute_cos_sin
        if offsets is not None:
            offsets = offsets.view(positions.shape[0], -1)
            positions = positions + offsets
        positions = positions.flatten()
        num_tokens = positions.shape[0]
        cos_sin = self.cos_sin_cache.index_select(0, positions).view(num_tokens, 1, -1)
        cos, sin = cos_sin.chunk(2, dim=-1)
        if self.is_neox_style:
            cos = torch.cat((cos, cos), dim=-1)
            sin = torch.cat((sin, sin), dim=-1)
        else:
            sin = torch.repeat_interleave(sin, 2, dim=-1, output_size=cos_sin.shape[-1])
            cos = torch.repeat_interleave(cos, 2, dim=-1, output_size=cos_sin.shape[-1])
        self.register_buffer("cos", cos, persistent=False)
        self.register_buffer("sin", sin, persistent=False)

    def forward_oot(
        self,
        positions: torch.Tensor,
        query: torch.Tensor,
        key: torch.Tensor,
        offsets: Optional[torch.Tensor] = None,
    ) -> tuple[torch.Tensor, torch.Tensor]:
        from habana_frameworks.torch.hpex.kernels import (RotaryPosEmbeddingMode, apply_rotary_pos_emb)

        # Prepare cos-sin caches for long-context + LoRA with offsets for every
        # forward, since the offset information wasn't available previously
        if not hasattr(self, "sin") or self.recompute_cos_sin:
            self.prepare_cos_sin(positions, offsets, recompute_cos_sin=True)
        if hasattr(self, "scaling_factors") or hasattr(self, "scaling_factor") or self.sin is None:
            self.prepare_cos_sin(positions, offsets)
        num_tokens = positions.shape[0] * positions.shape[1]
        # HPU RoPE kernel requires hidden dimension for cos and sin to be equal
        # to query hidden dimension, so the original tensors need to be
        # expanded
        # GPT-NeoX kernel requires position_ids = None, offset, mode = BLOCKWISE
        # and expansion of cos/sin tensors via concatenation
        # GPT-J kernel requires position_ids = None, offset = 0, mode = PAIRWISE
        # and expansion of cos/sin tensors via repeat_interleave
        rope_mode: RotaryPosEmbeddingMode
        rope_mode = RotaryPosEmbeddingMode.BLOCKWISE if self.is_neox_style else RotaryPosEmbeddingMode.PAIRWISE
        sin = self.sin
        cos = self.cos
        query_shape = query.shape
        key_shape = key.shape
        query = query.view(num_tokens, -1, self.head_size)
        key = key.view(num_tokens, -1, self.head_size)

        if self.head_size == self.rotary_dim:
            # Avoid unnecessary slicing and concatenation
            query = apply_rotary_pos_emb(query, cos, sin, None, 0, rope_mode)
            key = apply_rotary_pos_emb(key, cos, sin, None, 0, rope_mode)
            return query.reshape(query_shape), key.reshape(key_shape)

        query_rot = query[..., :self.rotary_dim]
        query_pass = query[..., self.rotary_dim:]
        query_rot = apply_rotary_pos_emb(query_rot, cos, sin, None, 0, rope_mode)
        query = torch.cat((query_rot, query_pass), dim=-1).reshape(query_shape)

        key_rot = key[..., :self.rotary_dim]
        key_pass = key[..., self.rotary_dim:]
        key_rot = apply_rotary_pos_emb(key_rot, cos, sin, None, 0, rope_mode)
        key = torch.cat((key_rot, key_pass), dim=-1).reshape(key_shape)
        return query, key

forward_oot 

forward_oot(
    positions: Tensor,
    query: Tensor,
    key: Tensor,
    offsets: Optional[Tensor] = None,
) -> tuple[Tensor, Tensor]
源文件在 vllm_gaudi/ops/hpu_rotary_embedding.py 
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
def forward_oot(
    self,
    positions: torch.Tensor,
    query: torch.Tensor,
    key: torch.Tensor,
    offsets: Optional[torch.Tensor] = None,
) -> tuple[torch.Tensor, torch.Tensor]:
    from habana_frameworks.torch.hpex.kernels import (RotaryPosEmbeddingMode, apply_rotary_pos_emb)

    # Prepare cos-sin caches for long-context + LoRA with offsets for every
    # forward, since the offset information wasn't available previously
    if not hasattr(self, "sin") or self.recompute_cos_sin:
        self.prepare_cos_sin(positions, offsets, recompute_cos_sin=True)
    if hasattr(self, "scaling_factors") or hasattr(self, "scaling_factor") or self.sin is None:
        self.prepare_cos_sin(positions, offsets)
    num_tokens = positions.shape[0] * positions.shape[1]
    # HPU RoPE kernel requires hidden dimension for cos and sin to be equal
    # to query hidden dimension, so the original tensors need to be
    # expanded
    # GPT-NeoX kernel requires position_ids = None, offset, mode = BLOCKWISE
    # and expansion of cos/sin tensors via concatenation
    # GPT-J kernel requires position_ids = None, offset = 0, mode = PAIRWISE
    # and expansion of cos/sin tensors via repeat_interleave
    rope_mode: RotaryPosEmbeddingMode
    rope_mode = RotaryPosEmbeddingMode.BLOCKWISE if self.is_neox_style else RotaryPosEmbeddingMode.PAIRWISE
    sin = self.sin
    cos = self.cos
    query_shape = query.shape
    key_shape = key.shape
    query = query.view(num_tokens, -1, self.head_size)
    key = key.view(num_tokens, -1, self.head_size)

    if self.head_size == self.rotary_dim:
        # Avoid unnecessary slicing and concatenation
        query = apply_rotary_pos_emb(query, cos, sin, None, 0, rope_mode)
        key = apply_rotary_pos_emb(key, cos, sin, None, 0, rope_mode)
        return query.reshape(query_shape), key.reshape(key_shape)

    query_rot = query[..., :self.rotary_dim]
    query_pass = query[..., self.rotary_dim:]
    query_rot = apply_rotary_pos_emb(query_rot, cos, sin, None, 0, rope_mode)
    query = torch.cat((query_rot, query_pass), dim=-1).reshape(query_shape)

    key_rot = key[..., :self.rotary_dim]
    key_pass = key[..., self.rotary_dim:]
    key_rot = apply_rotary_pos_emb(key_rot, cos, sin, None, 0, rope_mode)
    key = torch.cat((key_rot, key_pass), dim=-1).reshape(key_shape)
    return query, key

prepare_cos_sin 

prepare_cos_sin(
    positions: Tensor,
    offsets: Optional[Tensor] = None,
    recompute_cos_sin: bool = False,
)
源文件在 vllm_gaudi/ops/hpu_rotary_embedding.py 
 91
 92
 93
 94
 95
 96
 97
 98
 99
100
101
102
103
104
105
106
107
108
109
110
def prepare_cos_sin(self,
                    positions: torch.Tensor,
                    offsets: Optional[torch.Tensor] = None,
                    recompute_cos_sin: bool = False):
    self.recompute_cos_sin = recompute_cos_sin
    if offsets is not None:
        offsets = offsets.view(positions.shape[0], -1)
        positions = positions + offsets
    positions = positions.flatten()
    num_tokens = positions.shape[0]
    cos_sin = self.cos_sin_cache.index_select(0, positions).view(num_tokens, 1, -1)
    cos, sin = cos_sin.chunk(2, dim=-1)
    if self.is_neox_style:
        cos = torch.cat((cos, cos), dim=-1)
        sin = torch.cat((sin, sin), dim=-1)
    else:
        sin = torch.repeat_interleave(sin, 2, dim=-1, output_size=cos_sin.shape[-1])
        cos = torch.repeat_interleave(cos, 2, dim=-1, output_size=cos_sin.shape[-1])
    self.register_buffer("cos", cos, persistent=False)
    self.register_buffer("sin", sin, persistent=False)

HPULlama3RotaryEmbedding

基类: Llama3RotaryEmbedding

源文件在 vllm_gaudi/ops/hpu_rotary_embedding.py 
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
@Llama3RotaryEmbedding.register_oot
class HPULlama3RotaryEmbedding(Llama3RotaryEmbedding):

    def prepare_cos_sin(self,
                        positions: torch.Tensor,
                        offsets: Optional[torch.Tensor] = None,
                        recompute_cos_sin: bool = False):
        self.recompute_cos_sin = recompute_cos_sin
        if offsets is not None:
            offsets = offsets.view(positions.shape[0], -1)
            positions = positions + offsets
        positions = positions.flatten()
        num_tokens = positions.shape[0]
        cos_sin = self.cos_sin_cache.index_select(0, positions).view(num_tokens, 1, -1)
        cos, sin = cos_sin.chunk(2, dim=-1)
        if self.is_neox_style:
            cos = torch.cat((cos, cos), dim=-1)
            sin = torch.cat((sin, sin), dim=-1)
        else:
            sin = torch.repeat_interleave(sin, 2, dim=-1, output_size=cos_sin.shape[-1])
            cos = torch.repeat_interleave(cos, 2, dim=-1, output_size=cos_sin.shape[-1])
        self.register_buffer("cos", cos, persistent=False)
        self.register_buffer("sin", sin, persistent=False)

    def forward_oot(
        self,
        positions: torch.Tensor,
        query: torch.Tensor,
        key: torch.Tensor,
        offsets: Optional[torch.Tensor] = None,
    ) -> tuple[torch.Tensor, torch.Tensor]:
        from habana_frameworks.torch.hpex.kernels import (RotaryPosEmbeddingMode, apply_rotary_pos_emb)

        # Prepare cos-sin caches for long-context + LoRA with offsets for every
        # forward, since the offset information wasn't available previously
        if not hasattr(self, "sin") or self.recompute_cos_sin:
            self.prepare_cos_sin(positions, offsets, recompute_cos_sin=True)
        if hasattr(self, "scaling_factors") or hasattr(self, "scaling_factor") or self.sin is None:
            self.prepare_cos_sin(positions, offsets)
        num_tokens = positions.shape[0] * positions.shape[1]
        # HPU RoPE kernel requires hidden dimension for cos and sin to be equal
        # to query hidden dimension, so the original tensors need to be
        # expanded
        # GPT-NeoX kernel requires position_ids = None, offset, mode = BLOCKWISE
        # and expansion of cos/sin tensors via concatenation
        # GPT-J kernel requires position_ids = None, offset = 0, mode = PAIRWISE
        # and expansion of cos/sin tensors via repeat_interleave
        rope_mode: RotaryPosEmbeddingMode
        rope_mode = RotaryPosEmbeddingMode.BLOCKWISE if self.is_neox_style else RotaryPosEmbeddingMode.PAIRWISE
        sin = self.sin
        cos = self.cos
        query_shape = query.shape
        key_shape = key.shape
        query = query.view(num_tokens, -1, self.head_size)
        key = key.view(num_tokens, -1, self.head_size)

        if self.head_size == self.rotary_dim:
            # Avoid unnecessary slicing and concatenation
            query = apply_rotary_pos_emb(query, cos, sin, None, 0, rope_mode)
            key = apply_rotary_pos_emb(key, cos, sin, None, 0, rope_mode)
            return query.reshape(query_shape), key.reshape(key_shape)

        query_rot = query[..., :self.rotary_dim]
        query_pass = query[..., self.rotary_dim:]
        query_rot = apply_rotary_pos_emb(query_rot, cos, sin, None, 0, rope_mode)
        query = torch.cat((query_rot, query_pass), dim=-1).reshape(query_shape)

        key_rot = key[..., :self.rotary_dim]
        key_pass = key[..., self.rotary_dim:]
        key_rot = apply_rotary_pos_emb(key_rot, cos, sin, None, 0, rope_mode)
        key = torch.cat((key_rot, key_pass), dim=-1).reshape(key_shape)
        return query, key

forward_oot 

forward_oot(
    positions: Tensor,
    query: Tensor,
    key: Tensor,
    offsets: Optional[Tensor] = None,
) -> tuple[Tensor, Tensor]
源文件在 vllm_gaudi/ops/hpu_rotary_embedding.py 
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
def forward_oot(
    self,
    positions: torch.Tensor,
    query: torch.Tensor,
    key: torch.Tensor,
    offsets: Optional[torch.Tensor] = None,
) -> tuple[torch.Tensor, torch.Tensor]:
    from habana_frameworks.torch.hpex.kernels import (RotaryPosEmbeddingMode, apply_rotary_pos_emb)

    # Prepare cos-sin caches for long-context + LoRA with offsets for every
    # forward, since the offset information wasn't available previously
    if not hasattr(self, "sin") or self.recompute_cos_sin:
        self.prepare_cos_sin(positions, offsets, recompute_cos_sin=True)
    if hasattr(self, "scaling_factors") or hasattr(self, "scaling_factor") or self.sin is None:
        self.prepare_cos_sin(positions, offsets)
    num_tokens = positions.shape[0] * positions.shape[1]
    # HPU RoPE kernel requires hidden dimension for cos and sin to be equal
    # to query hidden dimension, so the original tensors need to be
    # expanded
    # GPT-NeoX kernel requires position_ids = None, offset, mode = BLOCKWISE
    # and expansion of cos/sin tensors via concatenation
    # GPT-J kernel requires position_ids = None, offset = 0, mode = PAIRWISE
    # and expansion of cos/sin tensors via repeat_interleave
    rope_mode: RotaryPosEmbeddingMode
    rope_mode = RotaryPosEmbeddingMode.BLOCKWISE if self.is_neox_style else RotaryPosEmbeddingMode.PAIRWISE
    sin = self.sin
    cos = self.cos
    query_shape = query.shape
    key_shape = key.shape
    query = query.view(num_tokens, -1, self.head_size)
    key = key.view(num_tokens, -1, self.head_size)

    if self.head_size == self.rotary_dim:
        # Avoid unnecessary slicing and concatenation
        query = apply_rotary_pos_emb(query, cos, sin, None, 0, rope_mode)
        key = apply_rotary_pos_emb(key, cos, sin, None, 0, rope_mode)
        return query.reshape(query_shape), key.reshape(key_shape)

    query_rot = query[..., :self.rotary_dim]
    query_pass = query[..., self.rotary_dim:]
    query_rot = apply_rotary_pos_emb(query_rot, cos, sin, None, 0, rope_mode)
    query = torch.cat((query_rot, query_pass), dim=-1).reshape(query_shape)

    key_rot = key[..., :self.rotary_dim]
    key_pass = key[..., self.rotary_dim:]
    key_rot = apply_rotary_pos_emb(key_rot, cos, sin, None, 0, rope_mode)
    key = torch.cat((key_rot, key_pass), dim=-1).reshape(key_shape)
    return query, key

prepare_cos_sin 

prepare_cos_sin(
    positions: Tensor,
    offsets: Optional[Tensor] = None,
    recompute_cos_sin: bool = False,
)
源文件在 vllm_gaudi/ops/hpu_rotary_embedding.py 
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
def prepare_cos_sin(self,
                    positions: torch.Tensor,
                    offsets: Optional[torch.Tensor] = None,
                    recompute_cos_sin: bool = False):
    self.recompute_cos_sin = recompute_cos_sin
    if offsets is not None:
        offsets = offsets.view(positions.shape[0], -1)
        positions = positions + offsets
    positions = positions.flatten()
    num_tokens = positions.shape[0]
    cos_sin = self.cos_sin_cache.index_select(0, positions).view(num_tokens, 1, -1)
    cos, sin = cos_sin.chunk(2, dim=-1)
    if self.is_neox_style:
        cos = torch.cat((cos, cos), dim=-1)
        sin = torch.cat((sin, sin), dim=-1)
    else:
        sin = torch.repeat_interleave(sin, 2, dim=-1, output_size=cos_sin.shape[-1])
        cos = torch.repeat_interleave(cos, 2, dim=-1, output_size=cos_sin.shape[-1])
    self.register_buffer("cos", cos, persistent=False)
    self.register_buffer("sin", sin, persistent=False)

HPULlama4VisionRotaryEmbedding

基类: Llama4VisionRotaryEmbedding

源文件在 vllm_gaudi/ops/hpu_rotary_embedding.py 
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
@Llama4VisionRotaryEmbedding.register_oot
class HPULlama4VisionRotaryEmbedding(Llama4VisionRotaryEmbedding):

    def __init__(
        self,
        head_size: int,
        rotary_dim: int,
        max_position_embeddings: int,
        base: float,
        is_neox_style: bool,
        dtype: torch.dtype,
    ) -> None:
        # original model use dtype as complex
        # for HPU, need to change to 'float'
        dtype = torch.float
        super().__init__(head_size, rotary_dim, max_position_embeddings, base, is_neox_style, dtype)

    def _compute_cos_sin_cache(self) -> torch.Tensor:
        inv_freq = self._compute_inv_freq(self.base)

        # self.max_position_embeddings here is number of image patches
        # i.e. (image_size // patch_size) ** 2
        num_patches = self.max_position_embeddings
        img_idx = torch.arange(num_patches,
                    dtype=torch.int32) \
                    .reshape(num_patches, 1)
        img_idx = torch.cat([img_idx, img_idx[:1]], dim=0)
        img_idx[-1, -1] = -2  # set to ID_CLS_TOKEN
        num_patches_single_dim = int(math.sqrt(num_patches))
        frequencies_x = img_idx % num_patches_single_dim
        frequencies_y = img_idx // num_patches_single_dim
        freqs_x = ((frequencies_x + 1)[..., None] * inv_freq[None, None, :]).repeat_interleave(2, dim=-1)
        freqs_y = ((frequencies_y + 1)[..., None] * inv_freq[None, None, :]).repeat_interleave(2, dim=-1)
        freqs = torch.cat([freqs_x, freqs_y], dim=-1).float().contiguous()[..., ::2]
        freqs = freqs.masked_fill(img_idx.reshape(-1, 1, 1) < 0, 0)

        # Compute cosine and sine for each angle.
        cos_vals = torch.cos(freqs)
        sin_vals = torch.sin(freqs)

        cache = torch.concat([cos_vals, sin_vals], dim=-1)
        return cache

    def forward_oot(  # type: ignore[override]
            self,
            query: torch.Tensor,
            key: torch.Tensor,
    ) -> tuple[torch.Tensor, torch.Tensor]:
        # Ensure the cache is on the right device.
        self.cos_sin_cache: torch.Tensor = self.cos_sin_cache.to(query.device)
        cos_cache, sin_cache = self.cos_sin_cache.chunk(2, dim=-1)
        # shape: [577, 1, 44]

        query_2d = query.float().reshape(*query.shape[:-1], -1, 2)
        key_2d = key.float().reshape(*key.shape[:-1], -1, 2)
        # e.g., [17, 577, 8, 44, 2]

        # Reshape cos_cache and sin_cache to broadcast properly.
        # We want them to have shape [1, 577, 1, 44] to match the
        # query dimensions (except for the last two dims).
        cos_cache = cos_cache.view(1, cos_cache.shape[0], 1, cos_cache.shape[-1])
        sin_cache = sin_cache.view(1, sin_cache.shape[0], 1, sin_cache.shape[-1])
        # e.g., [1, 577, 1, 44]

        # Separate the real and imaginary parts.
        q_real, q_imag = query_2d.unbind(-1)  # each: [17, 577, 8, 44]
        k_real, k_imag = key_2d.unbind(-1)  # each: [17, 577, 8, 44]

        # Manually apply the complex multiplication (rotation) using
        # the trigonometric identities.
        # For a complex multiplication: (a+ib)*(c+id) = (ac - bd) + i(ad + bc)
        q_rotated_real = q_real * cos_cache - q_imag * sin_cache
        q_rotated_imag = q_real * sin_cache + q_imag * cos_cache

        k_rotated_real = k_real * cos_cache - k_imag * sin_cache
        k_rotated_imag = k_real * sin_cache + k_imag * cos_cache

        # Re-stack the rotated components into a last dimension of size 2.
        q_rotated = torch.stack([q_rotated_real, q_rotated_imag], dim=-1)  # shape: [17, 577, 8, 44, 2]
        k_rotated = torch.stack([k_rotated_real, k_rotated_imag], dim=-1)  # shape: [17, 577, 8, 44, 2]

        # Flatten the last two dimensions to match the original output shape.
        # Flatten back to the desired shape
        # (e.g., collapse the last two dimensions).
        query_out = q_rotated.flatten(3)
        key_out = k_rotated.flatten(3)

        return query_out.type_as(query), key_out.type_as(key)

__init__ 

__init__(
    head_size: int,
    rotary_dim: int,
    max_position_embeddings: int,
    base: float,
    is_neox_style: bool,
    dtype: dtype,
) -> None
源文件在 vllm_gaudi/ops/hpu_rotary_embedding.py 
540
541
542
543
544
545
546
547
548
549
550
551
552
def __init__(
    self,
    head_size: int,
    rotary_dim: int,
    max_position_embeddings: int,
    base: float,
    is_neox_style: bool,
    dtype: torch.dtype,
) -> None:
    # original model use dtype as complex
    # for HPU, need to change to 'float'
    dtype = torch.float
    super().__init__(head_size, rotary_dim, max_position_embeddings, base, is_neox_style, dtype)

_compute_cos_sin_cache 

_compute_cos_sin_cache() -> Tensor
源文件在 vllm_gaudi/ops/hpu_rotary_embedding.py 
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
def _compute_cos_sin_cache(self) -> torch.Tensor:
    inv_freq = self._compute_inv_freq(self.base)

    # self.max_position_embeddings here is number of image patches
    # i.e. (image_size // patch_size) ** 2
    num_patches = self.max_position_embeddings
    img_idx = torch.arange(num_patches,
                dtype=torch.int32) \
                .reshape(num_patches, 1)
    img_idx = torch.cat([img_idx, img_idx[:1]], dim=0)
    img_idx[-1, -1] = -2  # set to ID_CLS_TOKEN
    num_patches_single_dim = int(math.sqrt(num_patches))
    frequencies_x = img_idx % num_patches_single_dim
    frequencies_y = img_idx // num_patches_single_dim
    freqs_x = ((frequencies_x + 1)[..., None] * inv_freq[None, None, :]).repeat_interleave(2, dim=-1)
    freqs_y = ((frequencies_y + 1)[..., None] * inv_freq[None, None, :]).repeat_interleave(2, dim=-1)
    freqs = torch.cat([freqs_x, freqs_y], dim=-1).float().contiguous()[..., ::2]
    freqs = freqs.masked_fill(img_idx.reshape(-1, 1, 1) < 0, 0)

    # Compute cosine and sine for each angle.
    cos_vals = torch.cos(freqs)
    sin_vals = torch.sin(freqs)

    cache = torch.concat([cos_vals, sin_vals], dim=-1)
    return cache

forward_oot 

forward_oot(
    query: Tensor, key: Tensor
) -> tuple[Tensor, Tensor]
源文件在 vllm_gaudi/ops/hpu_rotary_embedding.py 
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
def forward_oot(  # type: ignore[override]
        self,
        query: torch.Tensor,
        key: torch.Tensor,
) -> tuple[torch.Tensor, torch.Tensor]:
    # Ensure the cache is on the right device.
    self.cos_sin_cache: torch.Tensor = self.cos_sin_cache.to(query.device)
    cos_cache, sin_cache = self.cos_sin_cache.chunk(2, dim=-1)
    # shape: [577, 1, 44]

    query_2d = query.float().reshape(*query.shape[:-1], -1, 2)
    key_2d = key.float().reshape(*key.shape[:-1], -1, 2)
    # e.g., [17, 577, 8, 44, 2]

    # Reshape cos_cache and sin_cache to broadcast properly.
    # We want them to have shape [1, 577, 1, 44] to match the
    # query dimensions (except for the last two dims).
    cos_cache = cos_cache.view(1, cos_cache.shape[0], 1, cos_cache.shape[-1])
    sin_cache = sin_cache.view(1, sin_cache.shape[0], 1, sin_cache.shape[-1])
    # e.g., [1, 577, 1, 44]

    # Separate the real and imaginary parts.
    q_real, q_imag = query_2d.unbind(-1)  # each: [17, 577, 8, 44]
    k_real, k_imag = key_2d.unbind(-1)  # each: [17, 577, 8, 44]

    # Manually apply the complex multiplication (rotation) using
    # the trigonometric identities.
    # For a complex multiplication: (a+ib)*(c+id) = (ac - bd) + i(ad + bc)
    q_rotated_real = q_real * cos_cache - q_imag * sin_cache
    q_rotated_imag = q_real * sin_cache + q_imag * cos_cache

    k_rotated_real = k_real * cos_cache - k_imag * sin_cache
    k_rotated_imag = k_real * sin_cache + k_imag * cos_cache

    # Re-stack the rotated components into a last dimension of size 2.
    q_rotated = torch.stack([q_rotated_real, q_rotated_imag], dim=-1)  # shape: [17, 577, 8, 44, 2]
    k_rotated = torch.stack([k_rotated_real, k_rotated_imag], dim=-1)  # shape: [17, 577, 8, 44, 2]

    # Flatten the last two dimensions to match the original output shape.
    # Flatten back to the desired shape
    # (e.g., collapse the last two dimensions).
    query_out = q_rotated.flatten(3)
    key_out = k_rotated.flatten(3)

    return query_out.type_as(query), key_out.type_as(key)

HPUMRotaryEmbedding

基类: MRotaryEmbedding

源文件在 vllm_gaudi/ops/hpu_rotary_embedding.py 
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
@MRotaryEmbedding.register_oot
class HPUMRotaryEmbedding(MRotaryEmbedding):

    def prepare_cos_sin(self,
                        positions: torch.Tensor,
                        offsets: Optional[torch.Tensor] = None,
                        recompute_cos_sin: bool = False):
        self.recompute_cos_sin = recompute_cos_sin
        if offsets is not None:
            offsets = offsets.view(positions.shape[0], -1)
            positions = positions + offsets
        positions = positions.flatten()
        num_tokens = positions.shape[0]
        cos_sin = self.cos_sin_cache.index_select(0, positions).view(num_tokens, 1, -1)
        cos, sin = cos_sin.chunk(2, dim=-1)
        if self.is_neox_style:
            cos = torch.cat((cos, cos), dim=-1)
            sin = torch.cat((sin, sin), dim=-1)
        else:
            sin = torch.repeat_interleave(sin, 2, dim=-1, output_size=cos_sin.shape[-1])
            cos = torch.repeat_interleave(cos, 2, dim=-1, output_size=cos_sin.shape[-1])
        self.register_buffer("cos", cos, persistent=False)
        self.register_buffer("sin", sin, persistent=False)

    def forward_oot(
        self,
        positions: torch.Tensor,
        query: torch.Tensor,
        key: torch.Tensor,
        offsets: Optional[torch.Tensor] = None,
    ) -> tuple[torch.Tensor, torch.Tensor]:
        from habana_frameworks.torch.hpex.kernels import (RotaryPosEmbeddingMode, apply_rotary_pos_emb)

        num_tokens = positions.shape[-1]
        cos_sin = self.cos_sin_cache[positions]
        cos, sin = cos_sin.chunk(2, dim=-1)
        if positions.ndim == 2:
            assert self.mrope_section

            cos = torch.cat([m[i] for i, m in enumerate(cos.split(self.mrope_section, dim=-1))], dim=-1)
            sin = torch.cat([m[i] for i, m in enumerate(sin.split(self.mrope_section, dim=-1))], dim=-1)
        if self.is_neox_style:
            cos = torch.cat((cos, cos), dim=-1).unsqueeze(-2)
            sin = torch.cat((sin, sin), dim=-1).unsqueeze(-2)
        else:
            sin = torch.repeat_interleave(sin, 2, dim=-1, output_size=cos_sin.shape[-1]).unsqueeze(-2)
            cos = torch.repeat_interleave(cos, 2, dim=-1, output_size=cos_sin.shape[-1]).unsqueeze(-2)
        # HPU RoPE kernel requires hidden dimension for cos and sin to be equal
        # to query hidden dimension, so the original tensors need to be
        # expanded
        # GPT-NeoX kernel requires position_ids = None, offset, mode = BLOCKWISE
        # and expansion of cos/sin tensors via concatenation
        # GPT-J kernel requires position_ids = None, offset = 0, mode = PAIRWISE
        # and expansion of cos/sin tensors via repeat_interleave
        rope_mode: RotaryPosEmbeddingMode
        rope_mode = RotaryPosEmbeddingMode.BLOCKWISE if self.is_neox_style else RotaryPosEmbeddingMode.PAIRWISE
        query_shape = query.shape
        key_shape = key.shape
        query = query.view(num_tokens, -1, self.head_size)
        key = key.view(num_tokens, -1, self.head_size)

        if self.head_size == self.rotary_dim:
            # Avoid unnecessary slicing and concatenation
            query = apply_rotary_pos_emb(query, cos, sin, None, 0, rope_mode)
            key = apply_rotary_pos_emb(key, cos, sin, None, 0, rope_mode)
            return query.reshape(query_shape), key.reshape(key_shape)

        query_rot = query[..., :self.rotary_dim]
        query_pass = query[..., self.rotary_dim:]
        query_rot = apply_rotary_pos_emb(query_rot, cos, sin, None, 0, rope_mode)
        query = torch.cat((query_rot, query_pass), dim=-1).reshape(query_shape)

        key_rot = key[..., :self.rotary_dim]
        key_pass = key[..., self.rotary_dim:]
        key_rot = apply_rotary_pos_emb(key_rot, cos, sin, None, 0, rope_mode)
        key = torch.cat((key_rot, key_pass), dim=-1).reshape(key_shape)
        return query, key

forward_oot 

forward_oot(
    positions: Tensor,
    query: Tensor,
    key: Tensor,
    offsets: Optional[Tensor] = None,
) -> tuple[Tensor, Tensor]
源文件在 vllm_gaudi/ops/hpu_rotary_embedding.py 
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
def forward_oot(
    self,
    positions: torch.Tensor,
    query: torch.Tensor,
    key: torch.Tensor,
    offsets: Optional[torch.Tensor] = None,
) -> tuple[torch.Tensor, torch.Tensor]:
    from habana_frameworks.torch.hpex.kernels import (RotaryPosEmbeddingMode, apply_rotary_pos_emb)

    num_tokens = positions.shape[-1]
    cos_sin = self.cos_sin_cache[positions]
    cos, sin = cos_sin.chunk(2, dim=-1)
    if positions.ndim == 2:
        assert self.mrope_section

        cos = torch.cat([m[i] for i, m in enumerate(cos.split(self.mrope_section, dim=-1))], dim=-1)
        sin = torch.cat([m[i] for i, m in enumerate(sin.split(self.mrope_section, dim=-1))], dim=-1)
    if self.is_neox_style:
        cos = torch.cat((cos, cos), dim=-1).unsqueeze(-2)
        sin = torch.cat((sin, sin), dim=-1).unsqueeze(-2)
    else:
        sin = torch.repeat_interleave(sin, 2, dim=-1, output_size=cos_sin.shape[-1]).unsqueeze(-2)
        cos = torch.repeat_interleave(cos, 2, dim=-1, output_size=cos_sin.shape[-1]).unsqueeze(-2)
    # HPU RoPE kernel requires hidden dimension for cos and sin to be equal
    # to query hidden dimension, so the original tensors need to be
    # expanded
    # GPT-NeoX kernel requires position_ids = None, offset, mode = BLOCKWISE
    # and expansion of cos/sin tensors via concatenation
    # GPT-J kernel requires position_ids = None, offset = 0, mode = PAIRWISE
    # and expansion of cos/sin tensors via repeat_interleave
    rope_mode: RotaryPosEmbeddingMode
    rope_mode = RotaryPosEmbeddingMode.BLOCKWISE if self.is_neox_style else RotaryPosEmbeddingMode.PAIRWISE
    query_shape = query.shape
    key_shape = key.shape
    query = query.view(num_tokens, -1, self.head_size)
    key = key.view(num_tokens, -1, self.head_size)

    if self.head_size == self.rotary_dim:
        # Avoid unnecessary slicing and concatenation
        query = apply_rotary_pos_emb(query, cos, sin, None, 0, rope_mode)
        key = apply_rotary_pos_emb(key, cos, sin, None, 0, rope_mode)
        return query.reshape(query_shape), key.reshape(key_shape)

    query_rot = query[..., :self.rotary_dim]
    query_pass = query[..., self.rotary_dim:]
    query_rot = apply_rotary_pos_emb(query_rot, cos, sin, None, 0, rope_mode)
    query = torch.cat((query_rot, query_pass), dim=-1).reshape(query_shape)

    key_rot = key[..., :self.rotary_dim]
    key_pass = key[..., self.rotary_dim:]
    key_rot = apply_rotary_pos_emb(key_rot, cos, sin, None, 0, rope_mode)
    key = torch.cat((key_rot, key_pass), dim=-1).reshape(key_shape)
    return query, key

prepare_cos_sin 

prepare_cos_sin(
    positions: Tensor,
    offsets: Optional[Tensor] = None,
    recompute_cos_sin: bool = False,
)
源文件在 vllm_gaudi/ops/hpu_rotary_embedding.py 
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
def prepare_cos_sin(self,
                    positions: torch.Tensor,
                    offsets: Optional[torch.Tensor] = None,
                    recompute_cos_sin: bool = False):
    self.recompute_cos_sin = recompute_cos_sin
    if offsets is not None:
        offsets = offsets.view(positions.shape[0], -1)
        positions = positions + offsets
    positions = positions.flatten()
    num_tokens = positions.shape[0]
    cos_sin = self.cos_sin_cache.index_select(0, positions).view(num_tokens, 1, -1)
    cos, sin = cos_sin.chunk(2, dim=-1)
    if self.is_neox_style:
        cos = torch.cat((cos, cos), dim=-1)
        sin = torch.cat((sin, sin), dim=-1)
    else:
        sin = torch.repeat_interleave(sin, 2, dim=-1, output_size=cos_sin.shape[-1])
        cos = torch.repeat_interleave(cos, 2, dim=-1, output_size=cos_sin.shape[-1])
    self.register_buffer("cos", cos, persistent=False)
    self.register_buffer("sin", sin, persistent=False)

HPUPhi3LongRoPEScaledRotaryEmbedding

基类: Phi3LongRoPEScaledRotaryEmbedding

源文件在 vllm_gaudi/ops/hpu_rotary_embedding.py 
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
@CustomOp.register_oot(name='Phi3LongRoPEScaledRotaryEmbedding')
class HPUPhi3LongRoPEScaledRotaryEmbedding(Phi3LongRoPEScaledRotaryEmbedding):

    def prepare_cos_sin(self,
                        positions: torch.Tensor,
                        offsets: Optional[torch.Tensor] = None,
                        recompute_cos_sin: bool = False):
        self.recompute_cos_sin = recompute_cos_sin
        if offsets is not None:
            offsets = offsets.view(positions.shape[0], -1)
            positions = positions + offsets
        positions = positions.flatten()
        num_tokens = positions.shape[0]
        cos_sin = self.long_short_cos_sin_cache.index_select(0, positions).view(num_tokens, 1, -1)
        cos, sin = cos_sin.chunk(2, dim=-1)
        cos = torch.cat((cos, cos), dim=-1)
        sin = torch.cat((sin, sin), dim=-1)

        self.register_buffer("cos", cos, persistent=False)
        self.register_buffer("sin", sin, persistent=False)

    def forward_oot(
        self,
        positions: torch.Tensor,
        query: torch.Tensor,
        key: torch.Tensor,
        offsets: Optional[torch.Tensor] = None,
    ) -> tuple[torch.Tensor, torch.Tensor]:
        from habana_frameworks.torch.hpex.kernels import (RotaryPosEmbeddingMode, apply_rotary_pos_emb)

        rope_mode: RotaryPosEmbeddingMode
        rope_mode = RotaryPosEmbeddingMode.BLOCKWISE

        if hasattr(self, "scaling_factors") or self.sin is None:
            self.prepare_cos_sin(positions, offsets)
        if self.recompute_cos_sin:
            self.prepare_cos_sin(positions, offsets, recompute_cos_sin=True)

        sin = self.sin
        cos = self.cos

        if offsets is not None:
            offsets = offsets.view(positions.shape[0], -1)
            positions = positions + offsets
        positions = positions.flatten()
        num_tokens = positions.shape[0]

        query_shape = query.shape
        query = query.view(num_tokens, -1, self.head_size)
        query_rot = query[..., :self.rotary_dim]
        query_pass = query[..., self.rotary_dim:]
        query_rot = apply_rotary_pos_emb(query_rot, cos, sin, None, 0, rope_mode)
        query = torch.cat((query_rot, query_pass), dim=-1).reshape(query_shape)

        key_shape = key.shape
        key = key.view(num_tokens, -1, self.head_size)
        key_rot = key[..., :self.rotary_dim]
        key_pass = key[..., self.rotary_dim:]
        key_rot = apply_rotary_pos_emb(key_rot, cos, sin, None, 0, rope_mode)
        key = torch.cat((key_rot, key_pass), dim=-1).reshape(key_shape)

        return query, key

forward_oot 

forward_oot(
    positions: Tensor,
    query: Tensor,
    key: Tensor,
    offsets: Optional[Tensor] = None,
) -> tuple[Tensor, Tensor]
源文件在 vllm_gaudi/ops/hpu_rotary_embedding.py 
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
def forward_oot(
    self,
    positions: torch.Tensor,
    query: torch.Tensor,
    key: torch.Tensor,
    offsets: Optional[torch.Tensor] = None,
) -> tuple[torch.Tensor, torch.Tensor]:
    from habana_frameworks.torch.hpex.kernels import (RotaryPosEmbeddingMode, apply_rotary_pos_emb)

    rope_mode: RotaryPosEmbeddingMode
    rope_mode = RotaryPosEmbeddingMode.BLOCKWISE

    if hasattr(self, "scaling_factors") or self.sin is None:
        self.prepare_cos_sin(positions, offsets)
    if self.recompute_cos_sin:
        self.prepare_cos_sin(positions, offsets, recompute_cos_sin=True)

    sin = self.sin
    cos = self.cos

    if offsets is not None:
        offsets = offsets.view(positions.shape[0], -1)
        positions = positions + offsets
    positions = positions.flatten()
    num_tokens = positions.shape[0]

    query_shape = query.shape
    query = query.view(num_tokens, -1, self.head_size)
    query_rot = query[..., :self.rotary_dim]
    query_pass = query[..., self.rotary_dim:]
    query_rot = apply_rotary_pos_emb(query_rot, cos, sin, None, 0, rope_mode)
    query = torch.cat((query_rot, query_pass), dim=-1).reshape(query_shape)

    key_shape = key.shape
    key = key.view(num_tokens, -1, self.head_size)
    key_rot = key[..., :self.rotary_dim]
    key_pass = key[..., self.rotary_dim:]
    key_rot = apply_rotary_pos_emb(key_rot, cos, sin, None, 0, rope_mode)
    key = torch.cat((key_rot, key_pass), dim=-1).reshape(key_shape)

    return query, key

prepare_cos_sin 

prepare_cos_sin(
    positions: Tensor,
    offsets: Optional[Tensor] = None,
    recompute_cos_sin: bool = False,
)
源文件在 vllm_gaudi/ops/hpu_rotary_embedding.py 
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
def prepare_cos_sin(self,
                    positions: torch.Tensor,
                    offsets: Optional[torch.Tensor] = None,
                    recompute_cos_sin: bool = False):
    self.recompute_cos_sin = recompute_cos_sin
    if offsets is not None:
        offsets = offsets.view(positions.shape[0], -1)
        positions = positions + offsets
    positions = positions.flatten()
    num_tokens = positions.shape[0]
    cos_sin = self.long_short_cos_sin_cache.index_select(0, positions).view(num_tokens, 1, -1)
    cos, sin = cos_sin.chunk(2, dim=-1)
    cos = torch.cat((cos, cos), dim=-1)
    sin = torch.cat((sin, sin), dim=-1)

    self.register_buffer("cos", cos, persistent=False)
    self.register_buffer("sin", sin, persistent=False)

HPURotaryEmbedding

基类: RotaryEmbedding

原始旋转位置嵌入。

源文件在 vllm_gaudi/ops/hpu_rotary_embedding.py 
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
@RotaryEmbedding.register_oot
class HPURotaryEmbedding(RotaryEmbedding):
    """Original rotary positional embedding."""

    def prepare_cos_sin(self,
                        positions: torch.Tensor,
                        offsets: Optional[torch.Tensor] = None,
                        recompute_cos_sin: bool = False):
        self.recompute_cos_sin = recompute_cos_sin
        if offsets is not None:
            offsets = offsets.view(positions.shape[0], -1)
            positions = positions + offsets
        positions = positions.flatten()
        num_tokens = positions.shape[0]
        cos_sin = self.cos_sin_cache.index_select(0, positions).view(num_tokens, 1, -1)
        cos, sin = cos_sin.chunk(2, dim=-1)
        if self.is_neox_style:
            cos = torch.cat((cos, cos), dim=-1)
            sin = torch.cat((sin, sin), dim=-1)
        else:
            sin = torch.repeat_interleave(sin, 2, dim=-1, output_size=cos_sin.shape[-1])
            cos = torch.repeat_interleave(cos, 2, dim=-1, output_size=cos_sin.shape[-1])
        self.register_buffer("cos", cos, persistent=False)
        self.register_buffer("sin", sin, persistent=False)

    def forward_oot(
        self,
        positions: torch.Tensor,
        query: torch.Tensor,
        key: torch.Tensor,
        offsets: Optional[torch.Tensor] = None,
    ) -> tuple[torch.Tensor, torch.Tensor]:
        from habana_frameworks.torch.hpex.kernels import (RotaryPosEmbeddingMode, apply_rotary_pos_emb)

        # Prepare cos-sin caches for long-context + LoRA with offsets for every
        # forward, since the offset information wasn't available previously
        if not hasattr(self, "sin") or self.recompute_cos_sin:
            self.prepare_cos_sin(positions, offsets, recompute_cos_sin=True)
        if hasattr(self, "scaling_factors") or hasattr(self, "scaling_factor") or self.sin is None:
            self.prepare_cos_sin(positions, offsets)
        num_tokens = positions.numel()
        # HPU RoPE kernel requires hidden dimension for cos and sin to be equal
        # to query hidden dimension, so the original tensors need to be
        # expanded
        # GPT-NeoX kernel requires position_ids = None, offset, mode = BLOCKWISE
        # and expansion of cos/sin tensors via concatenation
        # GPT-J kernel requires position_ids = None, offset = 0, mode = PAIRWISE
        # and expansion of cos/sin tensors via repeat_interleave
        rope_mode: RotaryPosEmbeddingMode
        rope_mode = RotaryPosEmbeddingMode.BLOCKWISE if self.is_neox_style else RotaryPosEmbeddingMode.PAIRWISE
        sin = self.sin
        cos = self.cos
        query_shape = query.shape
        key_shape = key.shape
        query = query.view(num_tokens, -1, self.head_size)
        key = key.view(num_tokens, -1, self.head_size)

        if self.head_size == self.rotary_dim:
            # Avoid unnecessary slicing and concatenation
            query = apply_rotary_pos_emb(query, cos, sin, None, 0, rope_mode)
            key = apply_rotary_pos_emb(key, cos, sin, None, 0, rope_mode)
            return query.reshape(query_shape), key.reshape(key_shape)

        query_rot = query[..., :self.rotary_dim]
        query_pass = query[..., self.rotary_dim:]
        query_rot = apply_rotary_pos_emb(query_rot, cos, sin, None, 0, rope_mode)
        query = torch.cat((query_rot, query_pass), dim=-1).reshape(query_shape)

        key_rot = key[..., :self.rotary_dim]
        key_pass = key[..., self.rotary_dim:]
        key_rot = apply_rotary_pos_emb(key_rot, cos, sin, None, 0, rope_mode)
        key = torch.cat((key_rot, key_pass), dim=-1).reshape(key_shape)
        return query, key

forward_oot 

forward_oot(
    positions: Tensor,
    query: Tensor,
    key: Tensor,
    offsets: Optional[Tensor] = None,
) -> tuple[Tensor, Tensor]
源文件在 vllm_gaudi/ops/hpu_rotary_embedding.py 
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
def forward_oot(
    self,
    positions: torch.Tensor,
    query: torch.Tensor,
    key: torch.Tensor,
    offsets: Optional[torch.Tensor] = None,
) -> tuple[torch.Tensor, torch.Tensor]:
    from habana_frameworks.torch.hpex.kernels import (RotaryPosEmbeddingMode, apply_rotary_pos_emb)

    # Prepare cos-sin caches for long-context + LoRA with offsets for every
    # forward, since the offset information wasn't available previously
    if not hasattr(self, "sin") or self.recompute_cos_sin:
        self.prepare_cos_sin(positions, offsets, recompute_cos_sin=True)
    if hasattr(self, "scaling_factors") or hasattr(self, "scaling_factor") or self.sin is None:
        self.prepare_cos_sin(positions, offsets)
    num_tokens = positions.numel()
    # HPU RoPE kernel requires hidden dimension for cos and sin to be equal
    # to query hidden dimension, so the original tensors need to be
    # expanded
    # GPT-NeoX kernel requires position_ids = None, offset, mode = BLOCKWISE
    # and expansion of cos/sin tensors via concatenation
    # GPT-J kernel requires position_ids = None, offset = 0, mode = PAIRWISE
    # and expansion of cos/sin tensors via repeat_interleave
    rope_mode: RotaryPosEmbeddingMode
    rope_mode = RotaryPosEmbeddingMode.BLOCKWISE if self.is_neox_style else RotaryPosEmbeddingMode.PAIRWISE
    sin = self.sin
    cos = self.cos
    query_shape = query.shape
    key_shape = key.shape
    query = query.view(num_tokens, -1, self.head_size)
    key = key.view(num_tokens, -1, self.head_size)

    if self.head_size == self.rotary_dim:
        # Avoid unnecessary slicing and concatenation
        query = apply_rotary_pos_emb(query, cos, sin, None, 0, rope_mode)
        key = apply_rotary_pos_emb(key, cos, sin, None, 0, rope_mode)
        return query.reshape(query_shape), key.reshape(key_shape)

    query_rot = query[..., :self.rotary_dim]
    query_pass = query[..., self.rotary_dim:]
    query_rot = apply_rotary_pos_emb(query_rot, cos, sin, None, 0, rope_mode)
    query = torch.cat((query_rot, query_pass), dim=-1).reshape(query_shape)

    key_rot = key[..., :self.rotary_dim]
    key_pass = key[..., self.rotary_dim:]
    key_rot = apply_rotary_pos_emb(key_rot, cos, sin, None, 0, rope_mode)
    key = torch.cat((key_rot, key_pass), dim=-1).reshape(key_shape)
    return query, key

prepare_cos_sin 

prepare_cos_sin(
    positions: Tensor,
    offsets: Optional[Tensor] = None,
    recompute_cos_sin: bool = False,
)
源文件在 vllm_gaudi/ops/hpu_rotary_embedding.py 
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
def prepare_cos_sin(self,
                    positions: torch.Tensor,
                    offsets: Optional[torch.Tensor] = None,
                    recompute_cos_sin: bool = False):
    self.recompute_cos_sin = recompute_cos_sin
    if offsets is not None:
        offsets = offsets.view(positions.shape[0], -1)
        positions = positions + offsets
    positions = positions.flatten()
    num_tokens = positions.shape[0]
    cos_sin = self.cos_sin_cache.index_select(0, positions).view(num_tokens, 1, -1)
    cos, sin = cos_sin.chunk(2, dim=-1)
    if self.is_neox_style:
        cos = torch.cat((cos, cos), dim=-1)
        sin = torch.cat((sin, sin), dim=-1)
    else:
        sin = torch.repeat_interleave(sin, 2, dim=-1, output_size=cos_sin.shape[-1])
        cos = torch.repeat_interleave(cos, 2, dim=-1, output_size=cos_sin.shape[-1])
    self.register_buffer("cos", cos, persistent=False)
    self.register_buffer("sin", sin, persistent=False)

HPUYaRNScalingRotaryEmbedding

基类: YaRNScalingRotaryEmbedding

源文件在 vllm_gaudi/ops/hpu_rotary_embedding.py 
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
@YaRNScalingRotaryEmbedding.register_oot
class HPUYaRNScalingRotaryEmbedding(YaRNScalingRotaryEmbedding):

    def prepare_cos_sin(self,
                        positions: torch.Tensor,
                        offsets: Optional[torch.Tensor] = None,
                        recompute_cos_sin: bool = False):
        self.recompute_cos_sin = recompute_cos_sin
        if offsets is not None:
            offsets = offsets.view(positions.shape[0], -1)
            positions = positions + offsets
        positions = positions.flatten()
        num_tokens = positions.shape[0]
        cos_sin = self.cos_sin_cache.index_select(0, positions).view(num_tokens, 1, -1)
        cos, sin = cos_sin.chunk(2, dim=-1)
        if self.is_neox_style:
            cos = torch.cat((cos, cos), dim=-1)
            sin = torch.cat((sin, sin), dim=-1)
        else:
            sin = torch.repeat_interleave(sin, 2, dim=-1, output_size=cos_sin.shape[-1])
            cos = torch.repeat_interleave(cos, 2, dim=-1, output_size=cos_sin.shape[-1])
        self.register_buffer("cos", cos, persistent=False)
        self.register_buffer("sin", sin, persistent=False)

    def forward_oot(
        self,
        positions: torch.Tensor,
        query: torch.Tensor,
        key: torch.Tensor,
        offsets: Optional[torch.Tensor] = None,
    ) -> tuple[torch.Tensor, torch.Tensor]:
        from habana_frameworks.torch.hpex.kernels import (RotaryPosEmbeddingMode, apply_rotary_pos_emb)

        # Prepare cos-sin caches for long-context + LoRA with offsets for every
        # forward, since the offset information wasn't available previously
        if not hasattr(self, "sin") or self.recompute_cos_sin:
            self.prepare_cos_sin(positions, offsets, recompute_cos_sin=True)
        if hasattr(self, "scaling_factors") or hasattr(self, "scaling_factor") or self.sin is None:
            self.prepare_cos_sin(positions, offsets)
        num_tokens = positions.shape[0] * positions.shape[1]
        # HPU RoPE kernel requires hidden dimension for cos and sin to be equal
        # to query hidden dimension, so the original tensors need to be
        # expanded
        # GPT-NeoX kernel requires position_ids = None, offset, mode = BLOCKWISE
        # and expansion of cos/sin tensors via concatenation
        # GPT-J kernel requires position_ids = None, offset = 0, mode = PAIRWISE
        # and expansion of cos/sin tensors via repeat_interleave
        rope_mode: RotaryPosEmbeddingMode
        rope_mode = RotaryPosEmbeddingMode.BLOCKWISE if self.is_neox_style else RotaryPosEmbeddingMode.PAIRWISE
        sin = self.sin
        cos = self.cos
        query_shape = query.shape
        key_shape = key.shape
        query = query.view(num_tokens, -1, self.head_size)
        key = key.view(num_tokens, -1, self.head_size)

        if self.head_size == self.rotary_dim:
            # Avoid unnecessary slicing and concatenation
            query = apply_rotary_pos_emb(query, cos, sin, None, 0, rope_mode)
            key = apply_rotary_pos_emb(key, cos, sin, None, 0, rope_mode)
            return query.reshape(query_shape), key.reshape(key_shape)

        query_rot = query[..., :self.rotary_dim]
        query_pass = query[..., self.rotary_dim:]
        query_rot = apply_rotary_pos_emb(query_rot, cos, sin, None, 0, rope_mode)
        query = torch.cat((query_rot, query_pass), dim=-1).reshape(query_shape)

        key_rot = key[..., :self.rotary_dim]
        key_pass = key[..., self.rotary_dim:]
        key_rot = apply_rotary_pos_emb(key_rot, cos, sin, None, 0, rope_mode)
        key = torch.cat((key_rot, key_pass), dim=-1).reshape(key_shape)
        return query, key

forward_oot 

forward_oot(
    positions: Tensor,
    query: Tensor,
    key: Tensor,
    offsets: Optional[Tensor] = None,
) -> tuple[Tensor, Tensor]
源文件在 vllm_gaudi/ops/hpu_rotary_embedding.py 
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
def forward_oot(
    self,
    positions: torch.Tensor,
    query: torch.Tensor,
    key: torch.Tensor,
    offsets: Optional[torch.Tensor] = None,
) -> tuple[torch.Tensor, torch.Tensor]:
    from habana_frameworks.torch.hpex.kernels import (RotaryPosEmbeddingMode, apply_rotary_pos_emb)

    # Prepare cos-sin caches for long-context + LoRA with offsets for every
    # forward, since the offset information wasn't available previously
    if not hasattr(self, "sin") or self.recompute_cos_sin:
        self.prepare_cos_sin(positions, offsets, recompute_cos_sin=True)
    if hasattr(self, "scaling_factors") or hasattr(self, "scaling_factor") or self.sin is None:
        self.prepare_cos_sin(positions, offsets)
    num_tokens = positions.shape[0] * positions.shape[1]
    # HPU RoPE kernel requires hidden dimension for cos and sin to be equal
    # to query hidden dimension, so the original tensors need to be
    # expanded
    # GPT-NeoX kernel requires position_ids = None, offset, mode = BLOCKWISE
    # and expansion of cos/sin tensors via concatenation
    # GPT-J kernel requires position_ids = None, offset = 0, mode = PAIRWISE
    # and expansion of cos/sin tensors via repeat_interleave
    rope_mode: RotaryPosEmbeddingMode
    rope_mode = RotaryPosEmbeddingMode.BLOCKWISE if self.is_neox_style else RotaryPosEmbeddingMode.PAIRWISE
    sin = self.sin
    cos = self.cos
    query_shape = query.shape
    key_shape = key.shape
    query = query.view(num_tokens, -1, self.head_size)
    key = key.view(num_tokens, -1, self.head_size)

    if self.head_size == self.rotary_dim:
        # Avoid unnecessary slicing and concatenation
        query = apply_rotary_pos_emb(query, cos, sin, None, 0, rope_mode)
        key = apply_rotary_pos_emb(key, cos, sin, None, 0, rope_mode)
        return query.reshape(query_shape), key.reshape(key_shape)

    query_rot = query[..., :self.rotary_dim]
    query_pass = query[..., self.rotary_dim:]
    query_rot = apply_rotary_pos_emb(query_rot, cos, sin, None, 0, rope_mode)
    query = torch.cat((query_rot, query_pass), dim=-1).reshape(query_shape)

    key_rot = key[..., :self.rotary_dim]
    key_pass = key[..., self.rotary_dim:]
    key_rot = apply_rotary_pos_emb(key_rot, cos, sin, None, 0, rope_mode)
    key = torch.cat((key_rot, key_pass), dim=-1).reshape(key_shape)
    return query, key

prepare_cos_sin 

prepare_cos_sin(
    positions: Tensor,
    offsets: Optional[Tensor] = None,
    recompute_cos_sin: bool = False,
)
源文件在 vllm_gaudi/ops/hpu_rotary_embedding.py 
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
def prepare_cos_sin(self,
                    positions: torch.Tensor,
                    offsets: Optional[torch.Tensor] = None,
                    recompute_cos_sin: bool = False):
    self.recompute_cos_sin = recompute_cos_sin
    if offsets is not None:
        offsets = offsets.view(positions.shape[0], -1)
        positions = positions + offsets
    positions = positions.flatten()
    num_tokens = positions.shape[0]
    cos_sin = self.cos_sin_cache.index_select(0, positions).view(num_tokens, 1, -1)
    cos, sin = cos_sin.chunk(2, dim=-1)
    if self.is_neox_style:
        cos = torch.cat((cos, cos), dim=-1)
        sin = torch.cat((sin, sin), dim=-1)
    else:
        sin = torch.repeat_interleave(sin, 2, dim=-1, output_size=cos_sin.shape[-1])
        cos = torch.repeat_interleave(cos, 2, dim=-1, output_size=cos_sin.shape[-1])
    self.register_buffer("cos", cos, persistent=False)
    self.register_buffer("sin", sin, persistent=False)