Skip to content

vllm.model_executor.models.whisper

logger module-attribute 

logger = init_logger(__name__)

WhisperAttention

Bases: Module

Source code in vllm/model_executor/models/whisper.py 
 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
class WhisperAttention(nn.Module):

    def __init__(
        self,
        embed_dim: int,
        num_heads: int,
        bias: bool = True,
        attn_type: AttentionType = AttentionType.DECODER,
        cache_config: Optional[CacheConfig] = None,
        quant_config: Optional[QuantizationConfig] = None,
        prefix: str = "",
    ):
        super().__init__()
        self.embed_dim = embed_dim
        tp_size = get_tensor_model_parallel_world_size()
        self.total_num_heads = num_heads
        assert self.total_num_heads % tp_size == 0
        self.num_heads = self.total_num_heads // tp_size
        if self.total_num_heads >= tp_size:
            # Number of heads is greater than TP size, so we partition
            # the KV heads across multiple tensor parallel GPUs.
            assert self.total_num_heads % tp_size == 0
        else:
            # Number of heads is less than TP size, so we replicate
            # the KV heads across multiple tensor parallel GPUs.
            assert tp_size % self.total_num_heads == 0
        self.num_kv_heads = max(1, self.total_num_heads // tp_size)
        self.head_dim = self.embed_dim // self.total_num_heads
        self.q_size = self.num_heads * self.head_dim
        self.kv_size = self.num_kv_heads * self.head_dim
        self.attn_type = attn_type

        if (self.head_dim * num_heads) != self.embed_dim:
            raise ValueError(
                f"embed_dim must be divisible by num_heads (got `embed_dim`: "
                f"{self.embed_dim} and `num_heads`: {num_heads}).")
        self.scaling = self.head_dim**-0.5

        self._init_qkv(embed_dim, bias, quant_config, prefix=prefix)
        self.out_proj = RowParallelLinear(
            input_size=embed_dim,
            output_size=embed_dim,
            bias=bias,
            quant_config=quant_config,
            prefix=f"{prefix}.out_proj",
        )
        self.attn = Attention(
            self.num_heads,
            self.head_dim,
            self.scaling,
            num_kv_heads=self.num_kv_heads,
            cache_config=cache_config,
            quant_config=quant_config,
            prefix=f"{prefix}.attn",
            attn_type=self.attn_type,
        )

    def _init_qkv(
        self,
        embed_dim: int,
        bias: bool = True,
        quant_config: Optional[QuantizationConfig] = None,
        prefix: str = "",
    ) -> None:
        self.qkv_proj = QKVParallelLinear(
            hidden_size=embed_dim,
            head_size=self.head_dim,
            total_num_heads=self.total_num_heads,
            total_num_kv_heads=self.total_num_heads,
            bias=bias,
            quant_config=quant_config,
            prefix=f"{prefix}.qkv_proj",
        )

    def forward(
        self,
        hidden_states: torch.Tensor,
    ):
        qkv, _ = self.qkv_proj(hidden_states)
        q, k, v = qkv.split([self.q_size, self.kv_size, self.kv_size], dim=-1)

        attn_output = self.attn(q, k, v)

        output, _ = self.out_proj(attn_output)

        return output

attn instance-attribute 

attn = Attention(
    num_heads,
    head_dim,
    scaling,
    num_kv_heads=num_kv_heads,
    cache_config=cache_config,
    quant_config=quant_config,
    prefix=f"{prefix}.attn",
    attn_type=attn_type,
)

attn_type instance-attribute 

attn_type = attn_type

embed_dim instance-attribute 

embed_dim = embed_dim

head_dim instance-attribute 

head_dim = embed_dim // total_num_heads

kv_size instance-attribute 

kv_size = num_kv_heads * head_dim

num_heads instance-attribute 

num_heads = total_num_heads // tp_size

num_kv_heads instance-attribute 

num_kv_heads = max(1, total_num_heads // tp_size)

out_proj instance-attribute 

out_proj = RowParallelLinear(
    input_size=embed_dim,
    output_size=embed_dim,
    bias=bias,
    quant_config=quant_config,
    prefix=f"{prefix}.out_proj",
)

q_size instance-attribute 

q_size = num_heads * head_dim

scaling instance-attribute 

scaling = head_dim ** -0.5

total_num_heads instance-attribute 

total_num_heads = num_heads

__init__ 

__init__(
    embed_dim: int,
    num_heads: int,
    bias: bool = True,
    attn_type: AttentionType = DECODER,
    cache_config: Optional[CacheConfig] = None,
    quant_config: Optional[QuantizationConfig] = None,
    prefix: str = "",
)
Source code in vllm/model_executor/models/whisper.py 
 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
def __init__(
    self,
    embed_dim: int,
    num_heads: int,
    bias: bool = True,
    attn_type: AttentionType = AttentionType.DECODER,
    cache_config: Optional[CacheConfig] = None,
    quant_config: Optional[QuantizationConfig] = None,
    prefix: str = "",
):
    super().__init__()
    self.embed_dim = embed_dim
    tp_size = get_tensor_model_parallel_world_size()
    self.total_num_heads = num_heads
    assert self.total_num_heads % tp_size == 0
    self.num_heads = self.total_num_heads // tp_size
    if self.total_num_heads >= tp_size:
        # Number of heads is greater than TP size, so we partition
        # the KV heads across multiple tensor parallel GPUs.
        assert self.total_num_heads % tp_size == 0
    else:
        # Number of heads is less than TP size, so we replicate
        # the KV heads across multiple tensor parallel GPUs.
        assert tp_size % self.total_num_heads == 0
    self.num_kv_heads = max(1, self.total_num_heads // tp_size)
    self.head_dim = self.embed_dim // self.total_num_heads
    self.q_size = self.num_heads * self.head_dim
    self.kv_size = self.num_kv_heads * self.head_dim
    self.attn_type = attn_type

    if (self.head_dim * num_heads) != self.embed_dim:
        raise ValueError(
            f"embed_dim must be divisible by num_heads (got `embed_dim`: "
            f"{self.embed_dim} and `num_heads`: {num_heads}).")
    self.scaling = self.head_dim**-0.5

    self._init_qkv(embed_dim, bias, quant_config, prefix=prefix)
    self.out_proj = RowParallelLinear(
        input_size=embed_dim,
        output_size=embed_dim,
        bias=bias,
        quant_config=quant_config,
        prefix=f"{prefix}.out_proj",
    )
    self.attn = Attention(
        self.num_heads,
        self.head_dim,
        self.scaling,
        num_kv_heads=self.num_kv_heads,
        cache_config=cache_config,
        quant_config=quant_config,
        prefix=f"{prefix}.attn",
        attn_type=self.attn_type,
    )

_init_qkv 

_init_qkv(
    embed_dim: int,
    bias: bool = True,
    quant_config: Optional[QuantizationConfig] = None,
    prefix: str = "",
) -> None
Source code in vllm/model_executor/models/whisper.py 
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
def _init_qkv(
    self,
    embed_dim: int,
    bias: bool = True,
    quant_config: Optional[QuantizationConfig] = None,
    prefix: str = "",
) -> None:
    self.qkv_proj = QKVParallelLinear(
        hidden_size=embed_dim,
        head_size=self.head_dim,
        total_num_heads=self.total_num_heads,
        total_num_kv_heads=self.total_num_heads,
        bias=bias,
        quant_config=quant_config,
        prefix=f"{prefix}.qkv_proj",
    )

forward 

forward(hidden_states: Tensor)
Source code in vllm/model_executor/models/whisper.py 
132
133
134
135
136
137
138
139
140
141
142
143
def forward(
    self,
    hidden_states: torch.Tensor,
):
    qkv, _ = self.qkv_proj(hidden_states)
    q, k, v = qkv.split([self.q_size, self.kv_size, self.kv_size], dim=-1)

    attn_output = self.attn(q, k, v)

    output, _ = self.out_proj(attn_output)

    return output

WhisperAudioInputs

Bases: TypedDict

Source code in vllm/model_executor/models/whisper.py 
44
45
46
class WhisperAudioInputs(TypedDict):
    input_features: NestedTensors
    """Shape: `(batch_size, 128, M)`"""

input_features instance-attribute 

input_features: NestedTensors

Shape: (batch_size, 128, M)

WhisperCrossAttention

Bases: WhisperAttention

Source code in vllm/model_executor/models/whisper.py 
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
class WhisperCrossAttention(WhisperAttention):

    def __init__(
        self,
        embed_dim: int,
        num_heads: int,
        bias: bool = True,
        cache_config: Optional[CacheConfig] = None,
        quant_config: Optional[QuantizationConfig] = None,
        prefix: str = "",
    ):
        super().__init__(
            embed_dim=embed_dim,
            num_heads=num_heads,
            bias=bias,
            cache_config=cache_config,
            quant_config=quant_config,
            prefix=prefix,
            attn_type=AttentionType.ENCODER_DECODER,
        )

    def _init_qkv(
        self,
        embed_dim: int,
        bias: bool = True,
        quant_config: Optional[QuantizationConfig] = None,
        prefix: str = "",
    ) -> None:
        self.q_proj = ColumnParallelLinear(
            input_size=embed_dim,
            output_size=embed_dim,
            bias=bias,
            quant_config=quant_config,
            prefix=f"{prefix}.q_proj",
        )
        self.kv_proj = QKVParallelLinear(
            hidden_size=embed_dim,
            head_size=self.head_dim,
            total_num_heads=0,
            total_num_kv_heads=self.total_num_heads,
            bias=bias,
            quant_config=quant_config,
            prefix=f"{prefix}.kv_proj",
        )

    def forward(
        self,
        hidden_states: torch.Tensor,
        encoder_hidden_states: Optional[torch.Tensor],
    ):
        q, _ = self.q_proj(hidden_states)

        # Encoder hidden states are only computed once during prefill phase.
        # Afterwards, the keys and values should be available in the kv-cache.
        if encoder_hidden_states is not None:
            kv, _ = self.kv_proj(encoder_hidden_states)
            k, v = kv.split([self.kv_size, self.kv_size], dim=-1)
        else:
            k = v = None

        attn_output = self.attn(q, k, v)

        output, _ = self.out_proj(attn_output)

        return output

__init__ 

__init__(
    embed_dim: int,
    num_heads: int,
    bias: bool = True,
    cache_config: Optional[CacheConfig] = None,
    quant_config: Optional[QuantizationConfig] = None,
    prefix: str = "",
)
Source code in vllm/model_executor/models/whisper.py 
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
def __init__(
    self,
    embed_dim: int,
    num_heads: int,
    bias: bool = True,
    cache_config: Optional[CacheConfig] = None,
    quant_config: Optional[QuantizationConfig] = None,
    prefix: str = "",
):
    super().__init__(
        embed_dim=embed_dim,
        num_heads=num_heads,
        bias=bias,
        cache_config=cache_config,
        quant_config=quant_config,
        prefix=prefix,
        attn_type=AttentionType.ENCODER_DECODER,
    )

_init_qkv 

_init_qkv(
    embed_dim: int,
    bias: bool = True,
    quant_config: Optional[QuantizationConfig] = None,
    prefix: str = "",
) -> None
Source code in vllm/model_executor/models/whisper.py 
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
def _init_qkv(
    self,
    embed_dim: int,
    bias: bool = True,
    quant_config: Optional[QuantizationConfig] = None,
    prefix: str = "",
) -> None:
    self.q_proj = ColumnParallelLinear(
        input_size=embed_dim,
        output_size=embed_dim,
        bias=bias,
        quant_config=quant_config,
        prefix=f"{prefix}.q_proj",
    )
    self.kv_proj = QKVParallelLinear(
        hidden_size=embed_dim,
        head_size=self.head_dim,
        total_num_heads=0,
        total_num_kv_heads=self.total_num_heads,
        bias=bias,
        quant_config=quant_config,
        prefix=f"{prefix}.kv_proj",
    )

forward 

forward(
    hidden_states: Tensor,
    encoder_hidden_states: Optional[Tensor],
)
Source code in vllm/model_executor/models/whisper.py 
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
def forward(
    self,
    hidden_states: torch.Tensor,
    encoder_hidden_states: Optional[torch.Tensor],
):
    q, _ = self.q_proj(hidden_states)

    # Encoder hidden states are only computed once during prefill phase.
    # Afterwards, the keys and values should be available in the kv-cache.
    if encoder_hidden_states is not None:
        kv, _ = self.kv_proj(encoder_hidden_states)
        k, v = kv.split([self.kv_size, self.kv_size], dim=-1)
    else:
        k = v = None

    attn_output = self.attn(q, k, v)

    output, _ = self.out_proj(attn_output)

    return output

WhisperDecoder

Bases: Module

Source code in vllm/model_executor/models/whisper.py 
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
class WhisperDecoder(nn.Module):

    def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
        super().__init__()
        config = vllm_config.model_config.hf_config
        self.layerdrop = config.decoder_layerdrop
        self.padding_idx = config.pad_token_id
        self.max_target_positions = config.max_target_positions
        self.max_source_positions = config.max_source_positions
        self.embed_scale = (math.sqrt(config.d_model)
                            if config.scale_embedding else 1.0)

        self.embed_tokens = nn.Embedding(config.vocab_size, config.d_model,
                                         self.padding_idx)
        self.embed_positions = WhisperPositionalEmbedding(
            self.max_target_positions, config.d_model)
        self.start_layer, self.end_layer, self.layers = make_layers(
            config.decoder_layers,
            lambda prefix: WhisperDecoderLayer(vllm_config=vllm_config,
                                               prefix=f"{prefix}.layers"),
            prefix=f"{prefix}.layers",
        )
        self.layer_norm = nn.LayerNorm(config.d_model)

    def forward(
        self,
        input_ids,
        positions: torch.Tensor,
        encoder_hidden_states: Optional[torch.Tensor],
    ):
        inputs_embeds = self.get_input_embeddings(input_ids)
        positions = self.embed_positions(positions)
        hidden_states = inputs_embeds + positions

        for decoder_layer in self.layers:
            hidden_states = decoder_layer(
                hidden_states,
                encoder_hidden_states=encoder_hidden_states,
            )

        hidden_states = self.layer_norm(hidden_states)
        return hidden_states

    def get_input_embeddings(
        self,
        input_ids: torch.Tensor,
    ) -> torch.Tensor:
        return self.embed_tokens(input_ids)

embed_positions instance-attribute 

embed_positions = WhisperPositionalEmbedding(
    max_target_positions, d_model
)

embed_scale instance-attribute 

embed_scale = sqrt(d_model) if scale_embedding else 1.0

embed_tokens instance-attribute 

embed_tokens = Embedding(vocab_size, d_model, padding_idx)

layer_norm instance-attribute 

layer_norm = LayerNorm(d_model)

layerdrop instance-attribute 

layerdrop = decoder_layerdrop

max_source_positions instance-attribute 

max_source_positions = max_source_positions

max_target_positions instance-attribute 

max_target_positions = max_target_positions

padding_idx instance-attribute 

padding_idx = pad_token_id

__init__ 

__init__(*, vllm_config: VllmConfig, prefix: str = '')
Source code in vllm/model_executor/models/whisper.py 
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
    super().__init__()
    config = vllm_config.model_config.hf_config
    self.layerdrop = config.decoder_layerdrop
    self.padding_idx = config.pad_token_id
    self.max_target_positions = config.max_target_positions
    self.max_source_positions = config.max_source_positions
    self.embed_scale = (math.sqrt(config.d_model)
                        if config.scale_embedding else 1.0)

    self.embed_tokens = nn.Embedding(config.vocab_size, config.d_model,
                                     self.padding_idx)
    self.embed_positions = WhisperPositionalEmbedding(
        self.max_target_positions, config.d_model)
    self.start_layer, self.end_layer, self.layers = make_layers(
        config.decoder_layers,
        lambda prefix: WhisperDecoderLayer(vllm_config=vllm_config,
                                           prefix=f"{prefix}.layers"),
        prefix=f"{prefix}.layers",
    )
    self.layer_norm = nn.LayerNorm(config.d_model)

forward 

forward(
    input_ids,
    positions: Tensor,
    encoder_hidden_states: Optional[Tensor],
)
Source code in vllm/model_executor/models/whisper.py 
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
def forward(
    self,
    input_ids,
    positions: torch.Tensor,
    encoder_hidden_states: Optional[torch.Tensor],
):
    inputs_embeds = self.get_input_embeddings(input_ids)
    positions = self.embed_positions(positions)
    hidden_states = inputs_embeds + positions

    for decoder_layer in self.layers:
        hidden_states = decoder_layer(
            hidden_states,
            encoder_hidden_states=encoder_hidden_states,
        )

    hidden_states = self.layer_norm(hidden_states)
    return hidden_states

get_input_embeddings 

get_input_embeddings(input_ids: Tensor) -> Tensor
Source code in vllm/model_executor/models/whisper.py 
445
446
447
448
449
def get_input_embeddings(
    self,
    input_ids: torch.Tensor,
) -> torch.Tensor:
    return self.embed_tokens(input_ids)

WhisperDecoderLayer

Bases: Module

Source code in vllm/model_executor/models/whisper.py 
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
class WhisperDecoderLayer(nn.Module):

    def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
        super().__init__()
        config = vllm_config.model_config.hf_config
        cache_config = vllm_config.cache_config
        quant_config = vllm_config.quant_config

        self.self_attn = WhisperAttention(
            embed_dim=config.d_model,
            num_heads=config.decoder_attention_heads,
            attn_type=AttentionType.DECODER,
            cache_config=cache_config,
            quant_config=quant_config,
            prefix=f"{prefix}.self_attn",
        )
        self.self_attn_layer_norm = nn.LayerNorm(config.d_model)
        self.encoder_attn = WhisperCrossAttention(
            embed_dim=config.d_model,
            num_heads=config.decoder_attention_heads,
            cache_config=cache_config,
            quant_config=quant_config,
            prefix=f"{prefix}.encoder_attn",
        )
        self.encoder_attn_layer_norm = nn.LayerNorm(config.d_model)
        self.mlp = WhisperMLP(
            embed_dim=config.d_model,
            ffn_dim=config.decoder_ffn_dim,
            act_fn=config.activation_function,
            quant_config=quant_config,
            prefix=f"{prefix}.mlp",
        )
        self.final_layer_norm = nn.LayerNorm(config.d_model)

    def forward(
        self,
        hidden_states: torch.Tensor,
        encoder_hidden_states: Optional[torch.Tensor],
    ):
        residual = hidden_states
        hidden_states = self.self_attn_layer_norm(hidden_states)
        hidden_states = self.self_attn(hidden_states=hidden_states)
        hidden_states = residual + hidden_states

        residual = hidden_states
        hidden_states = self.encoder_attn_layer_norm(hidden_states)
        hidden_states = self.encoder_attn(
            hidden_states=hidden_states,
            encoder_hidden_states=encoder_hidden_states,
        )
        hidden_states = residual + hidden_states

        residual = hidden_states
        hidden_states = self.final_layer_norm(hidden_states)
        hidden_states = self.mlp(hidden_states)
        hidden_states = residual + hidden_states

        return hidden_states

encoder_attn instance-attribute 

encoder_attn = WhisperCrossAttention(
    embed_dim=d_model,
    num_heads=decoder_attention_heads,
    cache_config=cache_config,
    quant_config=quant_config,
    prefix=f"{prefix}.encoder_attn",
)

encoder_attn_layer_norm instance-attribute 

encoder_attn_layer_norm = LayerNorm(d_model)

final_layer_norm instance-attribute 

final_layer_norm = LayerNorm(d_model)

mlp instance-attribute 

mlp = WhisperMLP(
    embed_dim=d_model,
    ffn_dim=decoder_ffn_dim,
    act_fn=activation_function,
    quant_config=quant_config,
    prefix=f"{prefix}.mlp",
)

self_attn instance-attribute 

self_attn = WhisperAttention(
    embed_dim=d_model,
    num_heads=decoder_attention_heads,
    attn_type=DECODER,
    cache_config=cache_config,
    quant_config=quant_config,
    prefix=f"{prefix}.self_attn",
)

self_attn_layer_norm instance-attribute 

self_attn_layer_norm = LayerNorm(d_model)

__init__ 

__init__(*, vllm_config: VllmConfig, prefix: str = '')
Source code in vllm/model_executor/models/whisper.py 
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
def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
    super().__init__()
    config = vllm_config.model_config.hf_config
    cache_config = vllm_config.cache_config
    quant_config = vllm_config.quant_config

    self.self_attn = WhisperAttention(
        embed_dim=config.d_model,
        num_heads=config.decoder_attention_heads,
        attn_type=AttentionType.DECODER,
        cache_config=cache_config,
        quant_config=quant_config,
        prefix=f"{prefix}.self_attn",
    )
    self.self_attn_layer_norm = nn.LayerNorm(config.d_model)
    self.encoder_attn = WhisperCrossAttention(
        embed_dim=config.d_model,
        num_heads=config.decoder_attention_heads,
        cache_config=cache_config,
        quant_config=quant_config,
        prefix=f"{prefix}.encoder_attn",
    )
    self.encoder_attn_layer_norm = nn.LayerNorm(config.d_model)
    self.mlp = WhisperMLP(
        embed_dim=config.d_model,
        ffn_dim=config.decoder_ffn_dim,
        act_fn=config.activation_function,
        quant_config=quant_config,
        prefix=f"{prefix}.mlp",
    )
    self.final_layer_norm = nn.LayerNorm(config.d_model)

forward 

forward(
    hidden_states: Tensor,
    encoder_hidden_states: Optional[Tensor],
)
Source code in vllm/model_executor/models/whisper.py 
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
def forward(
    self,
    hidden_states: torch.Tensor,
    encoder_hidden_states: Optional[torch.Tensor],
):
    residual = hidden_states
    hidden_states = self.self_attn_layer_norm(hidden_states)
    hidden_states = self.self_attn(hidden_states=hidden_states)
    hidden_states = residual + hidden_states

    residual = hidden_states
    hidden_states = self.encoder_attn_layer_norm(hidden_states)
    hidden_states = self.encoder_attn(
        hidden_states=hidden_states,
        encoder_hidden_states=encoder_hidden_states,
    )
    hidden_states = residual + hidden_states

    residual = hidden_states
    hidden_states = self.final_layer_norm(hidden_states)
    hidden_states = self.mlp(hidden_states)
    hidden_states = residual + hidden_states

    return hidden_states

WhisperDummyInputsBuilder

Bases: BaseDummyInputsBuilder[WhisperProcessingInfo]

Source code in vllm/model_executor/models/whisper.py 
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
class WhisperDummyInputsBuilder(BaseDummyInputsBuilder[WhisperProcessingInfo]):

    def get_dummy_text(self, mm_counts: Mapping[str, int]) -> str:
        num_audios = mm_counts.get("audio", 0)

        return "<|startoftranscript|>" * num_audios

    def get_dummy_mm_data(
        self,
        seq_len: int,
        mm_counts: Mapping[str, int],
    ) -> MultiModalDataDict:
        feature_extractor = self.info.get_feature_extractor()

        sampling_rate = feature_extractor.sampling_rate
        audio_len = feature_extractor.chunk_length * sampling_rate
        num_audios = mm_counts.get("audio", 0)

        return {
            "audio":
            self._get_dummy_audios(length=audio_len, num_audios=num_audios)
        }

get_dummy_mm_data 

get_dummy_mm_data(
    seq_len: int, mm_counts: Mapping[str, int]
) -> MultiModalDataDict
Source code in vllm/model_executor/models/whisper.py 
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
def get_dummy_mm_data(
    self,
    seq_len: int,
    mm_counts: Mapping[str, int],
) -> MultiModalDataDict:
    feature_extractor = self.info.get_feature_extractor()

    sampling_rate = feature_extractor.sampling_rate
    audio_len = feature_extractor.chunk_length * sampling_rate
    num_audios = mm_counts.get("audio", 0)

    return {
        "audio":
        self._get_dummy_audios(length=audio_len, num_audios=num_audios)
    }

get_dummy_text 

get_dummy_text(mm_counts: Mapping[str, int]) -> str
Source code in vllm/model_executor/models/whisper.py 
546
547
548
549
def get_dummy_text(self, mm_counts: Mapping[str, int]) -> str:
    num_audios = mm_counts.get("audio", 0)

    return "<|startoftranscript|>" * num_audios

WhisperEncoder

Bases: Module

Source code in vllm/model_executor/models/whisper.py 
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
class WhisperEncoder(nn.Module):

    def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
        super().__init__()
        config = vllm_config.model_config.hf_config
        embed_dim = config.d_model
        self.num_mel_bins = config.num_mel_bins
        self.max_source_positions = config.max_source_positions
        self.embed_scale = (math.sqrt(embed_dim)
                            if config.scale_embedding else 1.0)

        self.conv1 = nn.Conv1d(self.num_mel_bins,
                               embed_dim,
                               kernel_size=3,
                               padding=1)
        self.conv2 = nn.Conv1d(embed_dim,
                               embed_dim,
                               kernel_size=3,
                               stride=2,
                               padding=1)
        self.embed_positions = nn.Embedding(self.max_source_positions,
                                            embed_dim)
        self.start_layer, self.end_layer, self.layers = make_layers(
            config.encoder_layers,
            lambda prefix: WhisperEncoderLayer(vllm_config=vllm_config,
                                               prefix=f"{prefix}.layers"),
            prefix=f"{prefix}.layers",
        )
        self.layer_norm = nn.LayerNorm(config.d_model)

        with torch.no_grad():
            self.embed_positions.weight.copy_(
                sinusoids(*self.embed_positions.weight.shape))

    def forward(self, input_features: Union[torch.Tensor, list[torch.Tensor]]):
        hidden_states = []
        for features in input_features:
            embeds = nn.functional.gelu(self.conv1(features))
            embeds = nn.functional.gelu(self.conv2(embeds))
            embeds = embeds.permute(1, 0)
            embeds = embeds + self.embed_positions.weight[:embeds.size(0), :]
            hidden_states.append(embeds)
        hidden_states = torch.cat(hidden_states)

        for encoder_layer in self.layers:
            hidden_states = encoder_layer(hidden_states)

        hidden_states = self.layer_norm(hidden_states)
        return hidden_states

conv1 instance-attribute 

conv1 = Conv1d(
    num_mel_bins, embed_dim, kernel_size=3, padding=1
)

conv2 instance-attribute 

conv2 = Conv1d(
    embed_dim, embed_dim, kernel_size=3, stride=2, padding=1
)

embed_positions instance-attribute 

embed_positions = Embedding(max_source_positions, embed_dim)

embed_scale instance-attribute 

embed_scale = sqrt(embed_dim) if scale_embedding else 1.0

layer_norm instance-attribute 

layer_norm = LayerNorm(d_model)

max_source_positions instance-attribute 

max_source_positions = max_source_positions

num_mel_bins instance-attribute 

num_mel_bins = num_mel_bins

__init__ 

__init__(*, vllm_config: VllmConfig, prefix: str = '')
Source code in vllm/model_executor/models/whisper.py 
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
def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
    super().__init__()
    config = vllm_config.model_config.hf_config
    embed_dim = config.d_model
    self.num_mel_bins = config.num_mel_bins
    self.max_source_positions = config.max_source_positions
    self.embed_scale = (math.sqrt(embed_dim)
                        if config.scale_embedding else 1.0)

    self.conv1 = nn.Conv1d(self.num_mel_bins,
                           embed_dim,
                           kernel_size=3,
                           padding=1)
    self.conv2 = nn.Conv1d(embed_dim,
                           embed_dim,
                           kernel_size=3,
                           stride=2,
                           padding=1)
    self.embed_positions = nn.Embedding(self.max_source_positions,
                                        embed_dim)
    self.start_layer, self.end_layer, self.layers = make_layers(
        config.encoder_layers,
        lambda prefix: WhisperEncoderLayer(vllm_config=vllm_config,
                                           prefix=f"{prefix}.layers"),
        prefix=f"{prefix}.layers",
    )
    self.layer_norm = nn.LayerNorm(config.d_model)

    with torch.no_grad():
        self.embed_positions.weight.copy_(
            sinusoids(*self.embed_positions.weight.shape))

forward 

forward(input_features: Union[Tensor, list[Tensor]])
Source code in vllm/model_executor/models/whisper.py 
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
def forward(self, input_features: Union[torch.Tensor, list[torch.Tensor]]):
    hidden_states = []
    for features in input_features:
        embeds = nn.functional.gelu(self.conv1(features))
        embeds = nn.functional.gelu(self.conv2(embeds))
        embeds = embeds.permute(1, 0)
        embeds = embeds + self.embed_positions.weight[:embeds.size(0), :]
        hidden_states.append(embeds)
    hidden_states = torch.cat(hidden_states)

    for encoder_layer in self.layers:
        hidden_states = encoder_layer(hidden_states)

    hidden_states = self.layer_norm(hidden_states)
    return hidden_states

WhisperEncoderLayer

Bases: Module

Source code in vllm/model_executor/models/whisper.py 
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
class WhisperEncoderLayer(nn.Module):

    def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
        super().__init__()
        config = vllm_config.model_config.hf_config
        cache_config = vllm_config.cache_config
        quant_config = vllm_config.quant_config

        self.embed_dim = config.d_model
        self.self_attn = WhisperAttention(
            embed_dim=self.embed_dim,
            num_heads=config.encoder_attention_heads,
            attn_type=AttentionType.ENCODER,
            cache_config=cache_config,
            quant_config=quant_config,
            prefix=f"{prefix}.self_attn",
        )
        self.self_attn_layer_norm = nn.LayerNorm(self.embed_dim)
        self.mlp = WhisperMLP(
            embed_dim=config.d_model,
            ffn_dim=config.encoder_ffn_dim,
            act_fn=config.activation_function,
            quant_config=quant_config,
            prefix=f"{prefix}.mlp",
        )
        self.final_layer_norm = nn.LayerNorm(self.embed_dim)

    def forward(
        self,
        hidden_states: torch.Tensor,
    ):
        residual = hidden_states
        hidden_states = self.self_attn_layer_norm(hidden_states)
        hidden_states = self.self_attn(hidden_states=hidden_states)
        hidden_states = residual + hidden_states
        residual = hidden_states
        hidden_states = self.final_layer_norm(hidden_states)
        hidden_states = self.mlp(hidden_states)
        hidden_states = residual + hidden_states

        hidden_states = cast_overflow_tensors(hidden_states)

        return hidden_states

embed_dim instance-attribute 

embed_dim = d_model

final_layer_norm instance-attribute 

final_layer_norm = LayerNorm(embed_dim)

mlp instance-attribute 

mlp = WhisperMLP(
    embed_dim=d_model,
    ffn_dim=encoder_ffn_dim,
    act_fn=activation_function,
    quant_config=quant_config,
    prefix=f"{prefix}.mlp",
)

self_attn instance-attribute 

self_attn = WhisperAttention(
    embed_dim=embed_dim,
    num_heads=encoder_attention_heads,
    attn_type=ENCODER,
    cache_config=cache_config,
    quant_config=quant_config,
    prefix=f"{prefix}.self_attn",
)

self_attn_layer_norm instance-attribute 

self_attn_layer_norm = LayerNorm(embed_dim)

__init__ 

__init__(*, vllm_config: VllmConfig, prefix: str = '')
Source code in vllm/model_executor/models/whisper.py 
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
    super().__init__()
    config = vllm_config.model_config.hf_config
    cache_config = vllm_config.cache_config
    quant_config = vllm_config.quant_config

    self.embed_dim = config.d_model
    self.self_attn = WhisperAttention(
        embed_dim=self.embed_dim,
        num_heads=config.encoder_attention_heads,
        attn_type=AttentionType.ENCODER,
        cache_config=cache_config,
        quant_config=quant_config,
        prefix=f"{prefix}.self_attn",
    )
    self.self_attn_layer_norm = nn.LayerNorm(self.embed_dim)
    self.mlp = WhisperMLP(
        embed_dim=config.d_model,
        ffn_dim=config.encoder_ffn_dim,
        act_fn=config.activation_function,
        quant_config=quant_config,
        prefix=f"{prefix}.mlp",
    )
    self.final_layer_norm = nn.LayerNorm(self.embed_dim)

forward 

forward(hidden_states: Tensor)
Source code in vllm/model_executor/models/whisper.py 
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
def forward(
    self,
    hidden_states: torch.Tensor,
):
    residual = hidden_states
    hidden_states = self.self_attn_layer_norm(hidden_states)
    hidden_states = self.self_attn(hidden_states=hidden_states)
    hidden_states = residual + hidden_states
    residual = hidden_states
    hidden_states = self.final_layer_norm(hidden_states)
    hidden_states = self.mlp(hidden_states)
    hidden_states = residual + hidden_states

    hidden_states = cast_overflow_tensors(hidden_states)

    return hidden_states

WhisperForConditionalGeneration

Bases: Module, SupportsTranscription, SupportsMultiModal, SupportsV0Only

Source code in vllm/model_executor/models/whisper.py 
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
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
@MULTIMODAL_REGISTRY.register_processor(WhisperMultiModalProcessor,
                                        info=WhisperProcessingInfo,
                                        dummy_inputs=WhisperDummyInputsBuilder)
class WhisperForConditionalGeneration(nn.Module, SupportsTranscription,
                                      SupportsMultiModal, SupportsV0Only):
    packed_modules_mapping = {
        "self_attn.qkv_proj": [
            "self_attn.q_proj",
            "self_attn.k_proj",
            "self_attn.v_proj",
        ],
        "encoder_attn.kv_proj": ["encoder_attn.k_proj", "encoder_attn.v_proj"],
    }

    hf_to_vllm_mapper = WeightsMapper(orig_to_new_substr={
        ".fc1.": ".mlp.fc1.",
        ".fc2.": ".mlp.fc2."
    })

    def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
        super().__init__()
        config = vllm_config.model_config.hf_config
        quant_config = vllm_config.quant_config
        self.config = config
        self.dtype = vllm_config.model_config.dtype

        self.model = WhisperModel(vllm_config=vllm_config, prefix=prefix)
        self.unpadded_vocab_size = config.vocab_size
        self.proj_out = ParallelLMHead(config.vocab_size,
                                       config.d_model,
                                       quant_config=quant_config)
        self.proj_out = self.proj_out.tie_weights(
            self.model.decoder.embed_tokens)
        logit_scale = getattr(config, "logit_scale", 1.0)
        self.logits_processor = LogitsProcessor(self.unpadded_vocab_size,
                                                config.vocab_size, logit_scale)

    def forward(
        self,
        input_ids: torch.Tensor,
        positions: torch.Tensor,
        **kwargs,
    ) -> torch.Tensor:
        audio_input = self._parse_and_validate_audio_input(**kwargs)
        decoder_outputs = self.model(
            input_features=audio_input["input_features"],
            input_ids=input_ids,
            positions=positions,
        )
        return decoder_outputs

    def get_language_model(self) -> torch.nn.Module:
        return self.model.decoder

    def get_multimodal_embeddings(
            self, **kwargs: object) -> Optional[MultiModalEmbeddings]:
        # TODO: This method does not obey the interface for SupportsMultiModal.
        # Refactor this once encoder/decoder support is implemented in V1.
        audio_input = self._parse_and_validate_audio_input(**kwargs)
        return self.model.get_encoder_outputs(audio_input["input_features"])

    def get_input_embeddings(
        self,
        input_ids: torch.Tensor,
        multimodal_embeddings: Optional[NestedTensors] = None,
    ) -> torch.Tensor:
        # TODO: This method just returns the decoder sequence embeddings since
        # Whisper does not have encoder text tokens. Refactor this once
        # encoder/decoder support is implemented in V1.
        return self.model.decoder.get_input_embeddings(input_ids)

    def _parse_and_validate_audio_input(
            self, **kwargs: object) -> WhisperAudioInputs:
        input_features = kwargs.pop("input_features", None)

        if input_features is not None:
            if not isinstance(input_features, (torch.Tensor, list)):
                raise ValueError("Incorrect type of audio features. "
                                 f"Got type: {type(input_features)}")
            input_features = torch.cat(
                [feat.to(self.dtype) for feat in input_features])

        return WhisperAudioInputs(input_features=input_features)

    def compute_logits(self, hidden_states: torch.Tensor,
                       sampling_metadata: SamplingMetadata) -> torch.Tensor:
        logits = self.logits_processor(self.proj_out, hidden_states,
                                       sampling_metadata)
        return logits

    def load_weights(self, weights: Iterable[tuple[str,
                                                   torch.Tensor]]) -> set[str]:
        loader = AutoWeightsLoader(self, skip_prefixes=["proj_out."])

        # add fake zeros bias for k_proj to state_dict
        weights = _create_fake_bias_for_k_proj(weights)
        return loader.load_weights(weights, mapper=self.hf_to_vllm_mapper)

config instance-attribute 

config = config

dtype instance-attribute 

dtype = dtype

hf_to_vllm_mapper class-attribute instance-attribute 

hf_to_vllm_mapper = WeightsMapper(
    orig_to_new_substr={
        ".fc1.": ".mlp.fc1.",
        ".fc2.": ".mlp.fc2.",
    }
)

logits_processor instance-attribute 

logits_processor = LogitsProcessor(
    unpadded_vocab_size, vocab_size, logit_scale
)

model instance-attribute 

model = WhisperModel(vllm_config=vllm_config, prefix=prefix)

packed_modules_mapping class-attribute instance-attribute 

packed_modules_mapping = {
    "self_attn.qkv_proj": [
        "self_attn.q_proj",
        "self_attn.k_proj",
        "self_attn.v_proj",
    ],
    "encoder_attn.kv_proj": [
        "encoder_attn.k_proj",
        "encoder_attn.v_proj",
    ],
}

proj_out instance-attribute 

proj_out = tie_weights(embed_tokens)

unpadded_vocab_size instance-attribute 

unpadded_vocab_size = vocab_size

__init__ 

__init__(*, vllm_config: VllmConfig, prefix: str = '')
Source code in vllm/model_executor/models/whisper.py 
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
    super().__init__()
    config = vllm_config.model_config.hf_config
    quant_config = vllm_config.quant_config
    self.config = config
    self.dtype = vllm_config.model_config.dtype

    self.model = WhisperModel(vllm_config=vllm_config, prefix=prefix)
    self.unpadded_vocab_size = config.vocab_size
    self.proj_out = ParallelLMHead(config.vocab_size,
                                   config.d_model,
                                   quant_config=quant_config)
    self.proj_out = self.proj_out.tie_weights(
        self.model.decoder.embed_tokens)
    logit_scale = getattr(config, "logit_scale", 1.0)
    self.logits_processor = LogitsProcessor(self.unpadded_vocab_size,
                                            config.vocab_size, logit_scale)

_parse_and_validate_audio_input 

_parse_and_validate_audio_input(
    **kwargs: object,
) -> WhisperAudioInputs
Source code in vllm/model_executor/models/whisper.py 
706
707
708
709
710
711
712
713
714
715
716
717
def _parse_and_validate_audio_input(
        self, **kwargs: object) -> WhisperAudioInputs:
    input_features = kwargs.pop("input_features", None)

    if input_features is not None:
        if not isinstance(input_features, (torch.Tensor, list)):
            raise ValueError("Incorrect type of audio features. "
                             f"Got type: {type(input_features)}")
        input_features = torch.cat(
            [feat.to(self.dtype) for feat in input_features])

    return WhisperAudioInputs(input_features=input_features)

compute_logits 

compute_logits(
    hidden_states: Tensor,
    sampling_metadata: SamplingMetadata,
) -> Tensor
Source code in vllm/model_executor/models/whisper.py 
719
720
721
722
723
def compute_logits(self, hidden_states: torch.Tensor,
                   sampling_metadata: SamplingMetadata) -> torch.Tensor:
    logits = self.logits_processor(self.proj_out, hidden_states,
                                   sampling_metadata)
    return logits

forward 

forward(
    input_ids: Tensor, positions: Tensor, **kwargs
) -> Tensor
Source code in vllm/model_executor/models/whisper.py 
672
673
674
675
676
677
678
679
680
681
682
683
684
def forward(
    self,
    input_ids: torch.Tensor,
    positions: torch.Tensor,
    **kwargs,
) -> torch.Tensor:
    audio_input = self._parse_and_validate_audio_input(**kwargs)
    decoder_outputs = self.model(
        input_features=audio_input["input_features"],
        input_ids=input_ids,
        positions=positions,
    )
    return decoder_outputs

get_input_embeddings 

get_input_embeddings(
    input_ids: Tensor,
    multimodal_embeddings: Optional[NestedTensors] = None,
) -> Tensor
Source code in vllm/model_executor/models/whisper.py 
696
697
698
699
700
701
702
703
704
def get_input_embeddings(
    self,
    input_ids: torch.Tensor,
    multimodal_embeddings: Optional[NestedTensors] = None,
) -> torch.Tensor:
    # TODO: This method just returns the decoder sequence embeddings since
    # Whisper does not have encoder text tokens. Refactor this once
    # encoder/decoder support is implemented in V1.
    return self.model.decoder.get_input_embeddings(input_ids)

get_language_model 

get_language_model() -> Module
Source code in vllm/model_executor/models/whisper.py 
686
687
def get_language_model(self) -> torch.nn.Module:
    return self.model.decoder

get_multimodal_embeddings 

get_multimodal_embeddings(
    **kwargs: object,
) -> Optional[MultiModalEmbeddings]
Source code in vllm/model_executor/models/whisper.py 
689
690
691
692
693
694
def get_multimodal_embeddings(
        self, **kwargs: object) -> Optional[MultiModalEmbeddings]:
    # TODO: This method does not obey the interface for SupportsMultiModal.
    # Refactor this once encoder/decoder support is implemented in V1.
    audio_input = self._parse_and_validate_audio_input(**kwargs)
    return self.model.get_encoder_outputs(audio_input["input_features"])

load_weights 

load_weights(
    weights: Iterable[tuple[str, Tensor]],
) -> set[str]
Source code in vllm/model_executor/models/whisper.py 
725
726
727
728
729
730
731
def load_weights(self, weights: Iterable[tuple[str,
                                               torch.Tensor]]) -> set[str]:
    loader = AutoWeightsLoader(self, skip_prefixes=["proj_out."])

    # add fake zeros bias for k_proj to state_dict
    weights = _create_fake_bias_for_k_proj(weights)
    return loader.load_weights(weights, mapper=self.hf_to_vllm_mapper)

WhisperMLP

Bases: Module

Source code in vllm/model_executor/models/whisper.py 
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
class WhisperMLP(nn.Module):

    def __init__(
        self,
        embed_dim: int,
        ffn_dim: int,
        act_fn: str,
        quant_config: Optional[QuantizationConfig] = None,
        prefix: str = "",
    ):
        super().__init__()

        self.activation_fn = get_act_fn(act_fn)
        self.fc1 = ColumnParallelLinear(
            input_size=embed_dim,
            output_size=ffn_dim,
            quant_config=quant_config,
            prefix=f"{prefix}.fc1",
        )
        self.fc2 = RowParallelLinear(
            input_size=ffn_dim,
            output_size=embed_dim,
            quant_config=quant_config,
            prefix=f"{prefix}.fc2",
        )

    def forward(self, hidden_states: torch.Tensor):
        hidden_states, _ = self.fc1(hidden_states)
        hidden_states = self.activation_fn(hidden_states)
        hidden_states, _ = self.fc2(hidden_states)
        return hidden_states

activation_fn instance-attribute 

activation_fn = get_act_fn(act_fn)

fc1 instance-attribute 

fc1 = ColumnParallelLinear(
    input_size=embed_dim,
    output_size=ffn_dim,
    quant_config=quant_config,
    prefix=f"{prefix}.fc1",
)

fc2 instance-attribute 

fc2 = RowParallelLinear(
    input_size=ffn_dim,
    output_size=embed_dim,
    quant_config=quant_config,
    prefix=f"{prefix}.fc2",
)

__init__ 

__init__(
    embed_dim: int,
    ffn_dim: int,
    act_fn: str,
    quant_config: Optional[QuantizationConfig] = None,
    prefix: str = "",
)
Source code in vllm/model_executor/models/whisper.py 
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
def __init__(
    self,
    embed_dim: int,
    ffn_dim: int,
    act_fn: str,
    quant_config: Optional[QuantizationConfig] = None,
    prefix: str = "",
):
    super().__init__()

    self.activation_fn = get_act_fn(act_fn)
    self.fc1 = ColumnParallelLinear(
        input_size=embed_dim,
        output_size=ffn_dim,
        quant_config=quant_config,
        prefix=f"{prefix}.fc1",
    )
    self.fc2 = RowParallelLinear(
        input_size=ffn_dim,
        output_size=embed_dim,
        quant_config=quant_config,
        prefix=f"{prefix}.fc2",
    )

forward 

forward(hidden_states: Tensor)
Source code in vllm/model_executor/models/whisper.py 
239
240
241
242
243
def forward(self, hidden_states: torch.Tensor):
    hidden_states, _ = self.fc1(hidden_states)
    hidden_states = self.activation_fn(hidden_states)
    hidden_states, _ = self.fc2(hidden_states)
    return hidden_states

WhisperModel

Bases: Module

Source code in vllm/model_executor/models/whisper.py 
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
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
class WhisperModel(nn.Module):

    def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
        super().__init__()
        self.encoder = WhisperEncoder(vllm_config=vllm_config,
                                      prefix=f"{prefix}.encoder")
        self.decoder = WhisperDecoder(vllm_config=vllm_config,
                                      prefix=f"{prefix}.decoder")

    def forward(
        self,
        input_features: Optional[Union[torch.Tensor, list[torch.Tensor]]],
        input_ids: Optional[torch.Tensor],
        positions: torch.Tensor,
    ) -> torch.Tensor:
        encoder_outputs = self.get_encoder_outputs(input_features)
        decoder_outputs = self.decoder(
            input_ids=input_ids,
            positions=positions,
            encoder_hidden_states=encoder_outputs,
        )
        return decoder_outputs

    def get_encoder_outputs(
        self,
        input_features: Optional[Union[torch.Tensor, list[torch.Tensor]]],
    ) -> Optional[torch.Tensor]:
        if input_features is None:
            return None
        return self.encoder(input_features)

    def load_weights(self, weights: Iterable[tuple[str,
                                                   torch.Tensor]]) -> set[str]:
        stacked_params_mapping = [
            # (param_name, shard_name, shard_id)
            (".self_attn.qkv_proj", ".self_attn.q_proj", "q"),
            (".self_attn.qkv_proj", ".self_attn.k_proj", "k"),
            (".self_attn.qkv_proj", ".self_attn.v_proj", "v"),
            (".encoder_attn.kv_proj", ".encoder_attn.k_proj", "k"),
            (".encoder_attn.kv_proj", ".encoder_attn.v_proj", "v"),
        ]
        params_dict = dict(self.named_parameters())
        loaded_params: set[str] = set()
        for name, loaded_weight in weights:
            for param_name, weight_name, shard_id in stacked_params_mapping:
                if weight_name not in name:
                    continue
                name = name.replace(weight_name, param_name)
                # Skip loading extra bias for GPTQ models.
                if name.endswith(".bias") and name not in params_dict:
                    continue

                param = params_dict[name]
                weight_loader = param.weight_loader
                weight_loader(param, loaded_weight, shard_id)
                break
            else:
                # Skip loading extra bias for GPTQ models.
                if name.endswith(".bias") and name not in params_dict:
                    continue

                param = params_dict[name]
                weight_loader = getattr(param, "weight_loader",
                                        default_weight_loader)
                weight_loader(param, loaded_weight)
            loaded_params.add(name)
        return loaded_params

decoder instance-attribute 

decoder = WhisperDecoder(
    vllm_config=vllm_config, prefix=f"{prefix}.decoder"
)

encoder instance-attribute 

encoder = WhisperEncoder(
    vllm_config=vllm_config, prefix=f"{prefix}.encoder"
)

__init__ 

__init__(*, vllm_config: VllmConfig, prefix: str = '')
Source code in vllm/model_executor/models/whisper.py 
454
455
456
457
458
459
def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
    super().__init__()
    self.encoder = WhisperEncoder(vllm_config=vllm_config,
                                  prefix=f"{prefix}.encoder")
    self.decoder = WhisperDecoder(vllm_config=vllm_config,
                                  prefix=f"{prefix}.decoder")

forward 

forward(
    input_features: Optional[Union[Tensor, list[Tensor]]],
    input_ids: Optional[Tensor],
    positions: Tensor,
) -> Tensor
Source code in vllm/model_executor/models/whisper.py 
461
462
463
464
465
466
467
468
469
470
471
472
473
def forward(
    self,
    input_features: Optional[Union[torch.Tensor, list[torch.Tensor]]],
    input_ids: Optional[torch.Tensor],
    positions: torch.Tensor,
) -> torch.Tensor:
    encoder_outputs = self.get_encoder_outputs(input_features)
    decoder_outputs = self.decoder(
        input_ids=input_ids,
        positions=positions,
        encoder_hidden_states=encoder_outputs,
    )
    return decoder_outputs

get_encoder_outputs 

get_encoder_outputs(
    input_features: Optional[Union[Tensor, list[Tensor]]],
) -> Optional[Tensor]
Source code in vllm/model_executor/models/whisper.py 
475
476
477
478
479
480
481
def get_encoder_outputs(
    self,
    input_features: Optional[Union[torch.Tensor, list[torch.Tensor]]],
) -> Optional[torch.Tensor]:
    if input_features is None:
        return None
    return self.encoder(input_features)

load_weights 

load_weights(
    weights: Iterable[tuple[str, Tensor]],
) -> set[str]
Source code in vllm/model_executor/models/whisper.py 
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
def load_weights(self, weights: Iterable[tuple[str,
                                               torch.Tensor]]) -> set[str]:
    stacked_params_mapping = [
        # (param_name, shard_name, shard_id)
        (".self_attn.qkv_proj", ".self_attn.q_proj", "q"),
        (".self_attn.qkv_proj", ".self_attn.k_proj", "k"),
        (".self_attn.qkv_proj", ".self_attn.v_proj", "v"),
        (".encoder_attn.kv_proj", ".encoder_attn.k_proj", "k"),
        (".encoder_attn.kv_proj", ".encoder_attn.v_proj", "v"),
    ]
    params_dict = dict(self.named_parameters())
    loaded_params: set[str] = set()
    for name, loaded_weight in weights:
        for param_name, weight_name, shard_id in stacked_params_mapping:
            if weight_name not in name:
                continue
            name = name.replace(weight_name, param_name)
            # Skip loading extra bias for GPTQ models.
            if name.endswith(".bias") and name not in params_dict:
                continue

            param = params_dict[name]
            weight_loader = param.weight_loader
            weight_loader(param, loaded_weight, shard_id)
            break
        else:
            # Skip loading extra bias for GPTQ models.
            if name.endswith(".bias") and name not in params_dict:
                continue

            param = params_dict[name]
            weight_loader = getattr(param, "weight_loader",
                                    default_weight_loader)
            weight_loader(param, loaded_weight)
        loaded_params.add(name)
    return loaded_params

WhisperMultiModalProcessor

Bases: EncDecMultiModalProcessor[WhisperProcessingInfo]

Source code in vllm/model_executor/models/whisper.py 
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
625
626
627
628
629
630
631
632
class WhisperMultiModalProcessor(
        EncDecMultiModalProcessor[WhisperProcessingInfo]):

    def _get_data_parser(self) -> MultiModalDataParser:
        feature_extractor = self.info.get_feature_extractor()
        return MultiModalDataParser(target_sr=feature_extractor.sampling_rate)

    @property
    def pad_dummy_encoder_prompt(self) -> bool:
        return True

    def create_encoder_prompt(
        self,
        prompt: Union[str, list[int]],
        mm_data: MultiModalDataDict,
    ) -> Union[str, list[int]]:
        # Strictly speaking, whisper encoder only accept audio features.
        # We create a dummy encoder prompt here which will be padded to
        # num_audio_tokens. So that we can create dummy data from this
        # for encoder profiling.
        return [0]

    def _call_hf_processor(
        self,
        prompt: str,
        mm_data: Mapping[str, object],
        mm_kwargs: Mapping[str, object],
    ) -> BatchFeature:
        if mm_data:
            feature_extractor = self.info.get_feature_extractor(**mm_kwargs)
            mm_data = dict(audio=mm_data.pop("audios"))
            mm_kwargs = dict(
                **mm_kwargs,
                sampling_rate=feature_extractor.sampling_rate,
            )
        processed_outputs = super()._call_hf_processor(
            prompt=prompt,
            mm_data=mm_data,
            mm_kwargs=mm_kwargs,
        )
        if "labels" in processed_outputs:
            processed_outputs["input_ids"] = processed_outputs.pop("labels")
        return processed_outputs

    def _get_mm_fields_config(
        self,
        hf_inputs: BatchFeature,
        hf_processor_mm_kwargs: Mapping[str, object],
    ) -> Mapping[str, MultiModalFieldConfig]:
        return dict(input_features=MultiModalFieldConfig.batched("audio"))

    def _get_prompt_updates(
        self,
        mm_items: MultiModalDataItems,
        hf_processor_mm_kwargs: Mapping[str, object],
        out_mm_kwargs: MultiModalKwargs,
    ) -> Sequence[PromptUpdate]:
        num_tokens = self.info.get_num_audio_tokens()
        return [
            PromptReplacement(
                modality="audio",
                target=[0],
                replacement=[0] * num_tokens,
            )
        ]

pad_dummy_encoder_prompt property 

pad_dummy_encoder_prompt: bool

_call_hf_processor 

_call_hf_processor(
    prompt: str,
    mm_data: Mapping[str, object],
    mm_kwargs: Mapping[str, object],
) -> BatchFeature
Source code in vllm/model_executor/models/whisper.py 
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
def _call_hf_processor(
    self,
    prompt: str,
    mm_data: Mapping[str, object],
    mm_kwargs: Mapping[str, object],
) -> BatchFeature:
    if mm_data:
        feature_extractor = self.info.get_feature_extractor(**mm_kwargs)
        mm_data = dict(audio=mm_data.pop("audios"))
        mm_kwargs = dict(
            **mm_kwargs,
            sampling_rate=feature_extractor.sampling_rate,
        )
    processed_outputs = super()._call_hf_processor(
        prompt=prompt,
        mm_data=mm_data,
        mm_kwargs=mm_kwargs,
    )
    if "labels" in processed_outputs:
        processed_outputs["input_ids"] = processed_outputs.pop("labels")
    return processed_outputs

_get_data_parser 

_get_data_parser() -> MultiModalDataParser
Source code in vllm/model_executor/models/whisper.py 
571
572
573
def _get_data_parser(self) -> MultiModalDataParser:
    feature_extractor = self.info.get_feature_extractor()
    return MultiModalDataParser(target_sr=feature_extractor.sampling_rate)

_get_mm_fields_config 

_get_mm_fields_config(
    hf_inputs: BatchFeature,
    hf_processor_mm_kwargs: Mapping[str, object],
) -> Mapping[str, MultiModalFieldConfig]
Source code in vllm/model_executor/models/whisper.py 
612
613
614
615
616
617
def _get_mm_fields_config(
    self,
    hf_inputs: BatchFeature,
    hf_processor_mm_kwargs: Mapping[str, object],
) -> Mapping[str, MultiModalFieldConfig]:
    return dict(input_features=MultiModalFieldConfig.batched("audio"))

_get_prompt_updates 

_get_prompt_updates(
    mm_items: MultiModalDataItems,
    hf_processor_mm_kwargs: Mapping[str, object],
    out_mm_kwargs: MultiModalKwargs,
) -> Sequence[PromptUpdate]
Source code in vllm/model_executor/models/whisper.py 
619
620
621
622
623
624
625
626
627
628
629
630
631
632
def _get_prompt_updates(
    self,
    mm_items: MultiModalDataItems,
    hf_processor_mm_kwargs: Mapping[str, object],
    out_mm_kwargs: MultiModalKwargs,
) -> Sequence[PromptUpdate]:
    num_tokens = self.info.get_num_audio_tokens()
    return [
        PromptReplacement(
            modality="audio",
            target=[0],
            replacement=[0] * num_tokens,
        )
    ]

create_encoder_prompt 

create_encoder_prompt(
    prompt: Union[str, list[int]],
    mm_data: MultiModalDataDict,
) -> Union[str, list[int]]
Source code in vllm/model_executor/models/whisper.py 
579
580
581
582
583
584
585
586
587
588
def create_encoder_prompt(
    self,
    prompt: Union[str, list[int]],
    mm_data: MultiModalDataDict,
) -> Union[str, list[int]]:
    # Strictly speaking, whisper encoder only accept audio features.
    # We create a dummy encoder prompt here which will be padded to
    # num_audio_tokens. So that we can create dummy data from this
    # for encoder profiling.
    return [0]

WhisperPositionalEmbedding

Bases: Embedding

Source code in vllm/model_executor/models/whisper.py 
49
50
51
52
53
54
55
class WhisperPositionalEmbedding(nn.Embedding):

    def __init__(self, num_positions: int, embedding_dim: int):
        super().__init__(num_positions, embedding_dim)

    def forward(self, position_ids):
        return self.weight[position_ids]

__init__ 

__init__(num_positions: int, embedding_dim: int)
Source code in vllm/model_executor/models/whisper.py 
51
52
def __init__(self, num_positions: int, embedding_dim: int):
    super().__init__(num_positions, embedding_dim)

forward 

forward(position_ids)
Source code in vllm/model_executor/models/whisper.py 
54
55
def forward(self, position_ids):
    return self.weight[position_ids]

WhisperProcessingInfo

Bases: BaseProcessingInfo

Source code in vllm/model_executor/models/whisper.py 
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
class WhisperProcessingInfo(BaseProcessingInfo):

    def get_hf_config(self) -> WhisperConfig:
        return self.ctx.get_hf_config(WhisperConfig)

    def get_hf_processor(self,
                         sampling_rate: Optional[int] = None
                         ) -> WhisperProcessor:
        return self.ctx.get_hf_processor(WhisperProcessor)

    def get_supported_mm_limits(self) -> Mapping[str, Optional[int]]:
        return {"audio": 1}

    def get_feature_extractor(self) -> WhisperFeatureExtractor:
        hf_processor = self.get_hf_processor()
        feature_extractor = hf_processor.feature_extractor  # type: ignore
        assert isinstance(feature_extractor, WhisperFeatureExtractor)
        return feature_extractor

    def get_num_audio_tokens(self) -> int:
        return self.get_hf_config().max_source_positions

get_feature_extractor 

get_feature_extractor() -> WhisperFeatureExtractor
Source code in vllm/model_executor/models/whisper.py 
534
535
536
537
538
def get_feature_extractor(self) -> WhisperFeatureExtractor:
    hf_processor = self.get_hf_processor()
    feature_extractor = hf_processor.feature_extractor  # type: ignore
    assert isinstance(feature_extractor, WhisperFeatureExtractor)
    return feature_extractor

get_hf_config 

get_hf_config() -> WhisperConfig
Source code in vllm/model_executor/models/whisper.py 
523
524
def get_hf_config(self) -> WhisperConfig:
    return self.ctx.get_hf_config(WhisperConfig)

get_hf_processor 

get_hf_processor(
    sampling_rate: Optional[int] = None,
) -> WhisperProcessor
Source code in vllm/model_executor/models/whisper.py 
526
527
528
529
def get_hf_processor(self,
                     sampling_rate: Optional[int] = None
                     ) -> WhisperProcessor:
    return self.ctx.get_hf_processor(WhisperProcessor)

get_num_audio_tokens 

get_num_audio_tokens() -> int
Source code in vllm/model_executor/models/whisper.py 
540
541
def get_num_audio_tokens(self) -> int:
    return self.get_hf_config().max_source_positions

get_supported_mm_limits 

get_supported_mm_limits() -> Mapping[str, Optional[int]]
Source code in vllm/model_executor/models/whisper.py 
531
532
def get_supported_mm_limits(self) -> Mapping[str, Optional[int]]:
    return {"audio": 1}

_create_fake_bias_for_k_proj 

_create_fake_bias_for_k_proj(
    weights: Iterable[tuple[str, Tensor]],
) -> Iterable[tuple[str, Tensor]]

Create full zeros bias for k_proj weight in self-attn and x-attn layers. So that the bias for k_proj in qkv_proj can be initialized with zeros.

Source code in vllm/model_executor/models/whisper.py 
734
735
736
737
738
739
740
741
742
743
744
745
746
def _create_fake_bias_for_k_proj(
    weights: Iterable[tuple[str, torch.Tensor]]
) -> Iterable[tuple[str, torch.Tensor]]:
    """
    Create full zeros bias for k_proj weight in self-attn and x-attn layers.
    So that the bias for k_proj in qkv_proj can be initialized with zeros.
    """
    for name, weight in weights:
        if name.endswith(".k_proj.weight"):
            bias = torch.zeros(weight.size(0))
            bias_name = name.replace("weight", "bias")
            yield from [(name, weight), (bias_name, bias)]
        yield name, weight