Skip to content

vllm.model_executor.models.bamba

Inference-only Bamba model.

ALL_DECODER_LAYER_TYPES module-attribute 

ALL_DECODER_LAYER_TYPES = {
    "attention": BambaAttentionDecoderLayer,
    "mamba": BambaMixerDecoderLayer,
}

BambaAttentionDecoderLayer

Bases: Module

Source code in vllm/model_executor/models/bamba.py 
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
class BambaAttentionDecoderLayer(nn.Module):

    def __init__(
        self,
        config: BambaConfig,
        layer_idx: int,
        cache_config: Optional[CacheConfig] = None,
        quant_config: Optional[QuantizationConfig] = None,
        prefix: str = "",
    ) -> None:
        super().__init__()
        rope_theta = getattr(config, "rope_theta", 10000)
        rope_scaling = getattr(config, "rope_scaling", None)
        max_position_embeddings = getattr(config, "max_position_embeddings",
                                          8192)
        self.hidden_size = config.hidden_size
        tp_size = get_tensor_model_parallel_world_size()
        self.total_num_heads = config.num_attention_heads
        assert self.total_num_heads % tp_size == 0
        self.num_heads = self.total_num_heads // tp_size
        self.total_num_kv_heads = config.num_key_value_heads
        if self.total_num_kv_heads >= tp_size:
            # Number of KV heads is greater than TP size, so we partition
            # the KV heads across multiple tensor parallel GPUs.
            assert self.total_num_kv_heads % tp_size == 0
        else:
            # Number of KV heads is less than TP size, so we replicate
            # the KV heads across multiple tensor parallel GPUs.
            assert tp_size % self.total_num_kv_heads == 0
        self.num_kv_heads = max(1, self.total_num_kv_heads // tp_size)
        self.head_dim = config.hidden_size // self.total_num_heads
        self.q_size = self.num_heads * self.head_dim
        self.kv_size = self.num_kv_heads * self.head_dim
        self.scaling = self.head_dim**-0.5
        self.rope_theta = rope_theta
        self.max_position_embeddings = max_position_embeddings

        if hasattr(config, "partial_rotary_factor"):
            rotary_dim = self.head_dim * config.partial_rotary_factor
        elif hasattr(config, "attn_rotary_emb"):
            rotary_dim = config.attn_rotary_emb  # for backward compatibility
        else:
            rotary_dim = self.head_dim  # default

        self.rotary_emb = get_rope(
            head_size=self.head_dim,
            rotary_dim=rotary_dim,
            max_position=max_position_embeddings,
            rope_scaling=rope_scaling,
            base=rope_theta,
            is_neox_style=True,
            dtype=torch.get_default_dtype(),  # see impl of get_rope
        )

        self.qkv_proj = QKVParallelLinear(
            config.hidden_size,
            self.head_dim,
            self.total_num_heads,
            self.total_num_kv_heads,
            bias=False,
            quant_config=quant_config,
        )
        self.o_proj = RowParallelLinear(self.total_num_heads * self.head_dim,
                                        config.hidden_size,
                                        bias=False,
                                        quant_config=quant_config)

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

        self.feed_forward = BambaMLP(config, quant_config=quant_config)
        self.input_layernorm = RMSNorm(config.hidden_size,
                                       eps=config.rms_norm_eps)
        self.pre_ff_layernorm = RMSNorm(config.hidden_size,
                                        eps=config.rms_norm_eps)

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

        q, k = self.rotary_emb(positions, q, k)
        attn_output = self.attn(q, k, v)
        output, _ = self.o_proj(attn_output)
        return output

    def forward(
        self,
        positions: torch.Tensor,
        hidden_states: torch.Tensor,
        residual: Optional[torch.Tensor],
        **kwargs,
    ):
        if residual is None:
            residual = hidden_states
            hidden_states = self.input_layernorm(hidden_states)
        else:
            hidden_states, residual = self.input_layernorm(
                hidden_states, residual)

        hidden_states = self.self_attention(
            positions=positions,
            hidden_states=hidden_states,
        )
        # Fully Connected
        hidden_states, residual = self.pre_ff_layernorm(
            hidden_states, residual)
        hidden_states = self.feed_forward(hidden_states)
        return hidden_states, residual

attn instance-attribute 

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

feed_forward instance-attribute 

feed_forward = BambaMLP(config, quant_config=quant_config)

head_dim instance-attribute 

head_dim = hidden_size // total_num_heads

hidden_size instance-attribute 

hidden_size = hidden_size

input_layernorm instance-attribute 

input_layernorm = RMSNorm(hidden_size, eps=rms_norm_eps)

kv_size instance-attribute 

kv_size = num_kv_heads * head_dim

max_position_embeddings instance-attribute 

max_position_embeddings = max_position_embeddings

num_heads instance-attribute 

num_heads = total_num_heads // tp_size

num_kv_heads instance-attribute 

num_kv_heads = max(1, total_num_kv_heads // tp_size)

o_proj instance-attribute 

o_proj = RowParallelLinear(
    total_num_heads * head_dim,
    hidden_size,
    bias=False,
    quant_config=quant_config,
)

pre_ff_layernorm instance-attribute 

pre_ff_layernorm = RMSNorm(hidden_size, eps=rms_norm_eps)

q_size instance-attribute 

q_size = num_heads * head_dim

qkv_proj instance-attribute 

qkv_proj = QKVParallelLinear(
    hidden_size,
    head_dim,
    total_num_heads,
    total_num_kv_heads,
    bias=False,
    quant_config=quant_config,
)

rope_theta instance-attribute 

rope_theta = rope_theta

rotary_emb instance-attribute 

rotary_emb = get_rope(
    head_size=head_dim,
    rotary_dim=rotary_dim,
    max_position=max_position_embeddings,
    rope_scaling=rope_scaling,
    base=rope_theta,
    is_neox_style=True,
    dtype=get_default_dtype(),
)

scaling instance-attribute 

scaling = head_dim ** -0.5

total_num_heads instance-attribute 

total_num_heads = num_attention_heads

total_num_kv_heads instance-attribute 

total_num_kv_heads = num_key_value_heads

__init__ 

__init__(
    config: BambaConfig,
    layer_idx: int,
    cache_config: Optional[CacheConfig] = None,
    quant_config: Optional[QuantizationConfig] = None,
    prefix: str = "",
) -> None
Source code in vllm/model_executor/models/bamba.py 
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
def __init__(
    self,
    config: BambaConfig,
    layer_idx: int,
    cache_config: Optional[CacheConfig] = None,
    quant_config: Optional[QuantizationConfig] = None,
    prefix: str = "",
) -> None:
    super().__init__()
    rope_theta = getattr(config, "rope_theta", 10000)
    rope_scaling = getattr(config, "rope_scaling", None)
    max_position_embeddings = getattr(config, "max_position_embeddings",
                                      8192)
    self.hidden_size = config.hidden_size
    tp_size = get_tensor_model_parallel_world_size()
    self.total_num_heads = config.num_attention_heads
    assert self.total_num_heads % tp_size == 0
    self.num_heads = self.total_num_heads // tp_size
    self.total_num_kv_heads = config.num_key_value_heads
    if self.total_num_kv_heads >= tp_size:
        # Number of KV heads is greater than TP size, so we partition
        # the KV heads across multiple tensor parallel GPUs.
        assert self.total_num_kv_heads % tp_size == 0
    else:
        # Number of KV heads is less than TP size, so we replicate
        # the KV heads across multiple tensor parallel GPUs.
        assert tp_size % self.total_num_kv_heads == 0
    self.num_kv_heads = max(1, self.total_num_kv_heads // tp_size)
    self.head_dim = config.hidden_size // self.total_num_heads
    self.q_size = self.num_heads * self.head_dim
    self.kv_size = self.num_kv_heads * self.head_dim
    self.scaling = self.head_dim**-0.5
    self.rope_theta = rope_theta
    self.max_position_embeddings = max_position_embeddings

    if hasattr(config, "partial_rotary_factor"):
        rotary_dim = self.head_dim * config.partial_rotary_factor
    elif hasattr(config, "attn_rotary_emb"):
        rotary_dim = config.attn_rotary_emb  # for backward compatibility
    else:
        rotary_dim = self.head_dim  # default

    self.rotary_emb = get_rope(
        head_size=self.head_dim,
        rotary_dim=rotary_dim,
        max_position=max_position_embeddings,
        rope_scaling=rope_scaling,
        base=rope_theta,
        is_neox_style=True,
        dtype=torch.get_default_dtype(),  # see impl of get_rope
    )

    self.qkv_proj = QKVParallelLinear(
        config.hidden_size,
        self.head_dim,
        self.total_num_heads,
        self.total_num_kv_heads,
        bias=False,
        quant_config=quant_config,
    )
    self.o_proj = RowParallelLinear(self.total_num_heads * self.head_dim,
                                    config.hidden_size,
                                    bias=False,
                                    quant_config=quant_config)

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

    self.feed_forward = BambaMLP(config, quant_config=quant_config)
    self.input_layernorm = RMSNorm(config.hidden_size,
                                   eps=config.rms_norm_eps)
    self.pre_ff_layernorm = RMSNorm(config.hidden_size,
                                    eps=config.rms_norm_eps)

forward 

forward(
    positions: Tensor,
    hidden_states: Tensor,
    residual: Optional[Tensor],
    **kwargs,
)
Source code in vllm/model_executor/models/bamba.py 
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
def forward(
    self,
    positions: torch.Tensor,
    hidden_states: torch.Tensor,
    residual: Optional[torch.Tensor],
    **kwargs,
):
    if residual is None:
        residual = hidden_states
        hidden_states = self.input_layernorm(hidden_states)
    else:
        hidden_states, residual = self.input_layernorm(
            hidden_states, residual)

    hidden_states = self.self_attention(
        positions=positions,
        hidden_states=hidden_states,
    )
    # Fully Connected
    hidden_states, residual = self.pre_ff_layernorm(
        hidden_states, residual)
    hidden_states = self.feed_forward(hidden_states)
    return hidden_states, residual

self_attention 

self_attention(
    positions: Tensor, hidden_states: Tensor, **kwargs
) -> Tensor
Source code in vllm/model_executor/models/bamba.py 
217
218
219
220
221
222
223
224
225
226
227
228
229
def self_attention(
    self,
    positions: torch.Tensor,
    hidden_states: torch.Tensor,
    **kwargs,
) -> torch.Tensor:
    qkv, _ = self.qkv_proj(hidden_states)
    q, k, v = qkv.split([self.q_size, self.kv_size, self.kv_size], dim=-1)

    q, k = self.rotary_emb(positions, q, k)
    attn_output = self.attn(q, k, v)
    output, _ = self.o_proj(attn_output)
    return output

BambaForCausalLM

Bases: Module, HasInnerState, SupportsLoRA, SupportsPP, IsHybrid, SupportsQuant

Source code in vllm/model_executor/models/bamba.py 
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
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
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
class BambaForCausalLM(nn.Module, HasInnerState, SupportsLoRA, SupportsPP,
                       IsHybrid, SupportsQuant):
    packed_modules_mapping = {
        "qkv_proj": [
            "q_proj",
            "k_proj",
            "v_proj",
        ],
        "gate_up_proj": ["up_proj", "down_proj"]
    }

    # LoRA specific attributes
    embedding_modules = {
        "embed_tokens": "input_embeddings",
        "lm_head": "output_embeddings",
    }
    embedding_padding_modules = ["lm_head"]

    def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
        config = vllm_config.model_config.hf_config
        self.vllm_config = vllm_config
        self.model_config = vllm_config.model_config
        cache_config = vllm_config.cache_config
        lora_config = vllm_config.lora_config
        scheduler_config = vllm_config.scheduler_config
        assert not cache_config.enable_prefix_caching, \
            "Bamba currently does not support prefix caching"

        self.quant_config = vllm_config.quant_config

        super().__init__()
        self.config = config
        self.scheduler_config = scheduler_config
        self.model = BambaModel(vllm_config=vllm_config,
                                prefix=maybe_prefix(prefix, "model"))
        self.unpadded_vocab_size = config.vocab_size
        if lora_config:
            self.unpadded_vocab_size += lora_config.lora_extra_vocab_size
        self.lm_head = ParallelLMHead(
            self.unpadded_vocab_size,
            config.hidden_size,
            org_num_embeddings=config.vocab_size,
            padding_size=DEFAULT_VOCAB_PADDING_SIZE
            # We need bigger padding if using lora for kernel
            # compatibility
            if not lora_config else lora_config.lora_vocab_padding_size,
        )
        # Used to track and store by the Mamba cache between steps.
        self.mamba_cache: Optional[MambaCacheManager] = None

        self.logits_processor = LogitsProcessor(self.unpadded_vocab_size,
                                                config.vocab_size)

        self.make_empty_intermediate_tensors = (
            self.model.make_empty_intermediate_tensors)

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

    def forward(self,
                input_ids: torch.Tensor,
                positions: torch.Tensor,
                intermediate_tensors: Optional[IntermediateTensors] = None,
                inputs_embeds: Optional[torch.Tensor] = None,
                **kwargs):

        mamba_cache_params = None
        if not envs.VLLM_USE_V1:
            if self.mamba_cache is None:
                num_mamba_layers = \
                    self.model_config.get_num_layers_by_block_type(
                        self.vllm_config.parallel_config,
                        LayerBlockType.mamba
                    )

                self.mamba_cache = MambaCacheManager(
                    self.vllm_config, self.lm_head.weight.dtype,
                    num_mamba_layers, *self._get_mamba_cache_shape())

            mamba_cache_params = self.mamba_cache.current_run_tensors(**kwargs)

        hidden_states = self.model(input_ids, positions, mamba_cache_params,
                                   intermediate_tensors, inputs_embeds)

        return hidden_states

    def copy_inputs_before_cuda_graphs(self, input_buffers, **kwargs):
        return self.mamba_cache.copy_inputs_before_cuda_graphs(
            input_buffers, **kwargs)

    def get_seqlen_agnostic_capture_inputs(self, batch_size: int):
        return self.mamba_cache.get_seqlen_agnostic_capture_inputs(batch_size)

    def _get_mamba_cache_shape(
            self) -> tuple[tuple[int, int], tuple[int, int]]:
        world_size = get_tensor_model_parallel_world_size()
        hidden_size = self.config.hidden_size

        conv_state_shape, temporal_state_shape = None, None

        intermediate_size = self.config.mamba_expand * hidden_size

        # if n_groups is not divisible by world_size, need to extend the shards
        # to ensure all groups needed by a head is sharded along with it
        n_groups = (self.config.mamba_n_groups + extra_groups_for_head_shards(
            self.config.mamba_n_groups, world_size))

        # - heads and n_groups are TP-ed
        conv_dim = (intermediate_size +
                    2 * n_groups * self.config.mamba_d_state)
        conv_state_shape = (
            divide(conv_dim, world_size),
            self.config.mamba_d_conv - 1,
        )

        # These are not TP-ed as they depend on A, dt_bias, D
        # - they are typically small
        #   e.g., (h_heads, d_head, d_state) = (128, 64, 128)
        temporal_state_shape = (
            divide(self.config.mamba_n_heads, world_size),
            self.config.mamba_d_head,
            self.config.mamba_d_state,
        )
        return conv_state_shape, temporal_state_shape

    def compute_logits(
        self,
        hidden_states: torch.Tensor,
        sampling_metadata: SamplingMetadata,
    ) -> Optional[torch.Tensor]:
        logits = self.logits_processor(self.lm_head, hidden_states,
                                       sampling_metadata)
        return logits

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

config instance-attribute 

config = config

embedding_modules class-attribute instance-attribute 

embedding_modules = {
    "embed_tokens": "input_embeddings",
    "lm_head": "output_embeddings",
}

embedding_padding_modules class-attribute instance-attribute 

embedding_padding_modules = ['lm_head']

lm_head instance-attribute 

lm_head = ParallelLMHead(
    unpadded_vocab_size,
    hidden_size,
    org_num_embeddings=vocab_size,
    padding_size=DEFAULT_VOCAB_PADDING_SIZE
    if not lora_config
    else lora_vocab_padding_size,
)

logits_processor instance-attribute 

logits_processor = LogitsProcessor(
    unpadded_vocab_size, vocab_size
)

make_empty_intermediate_tensors instance-attribute 

make_empty_intermediate_tensors = (
    make_empty_intermediate_tensors
)

mamba_cache instance-attribute 

mamba_cache: Optional[MambaCacheManager] = None

model instance-attribute 

model = BambaModel(
    vllm_config=vllm_config,
    prefix=maybe_prefix(prefix, "model"),
)

model_config instance-attribute 

model_config = model_config

packed_modules_mapping class-attribute instance-attribute 

packed_modules_mapping = {
    "qkv_proj": ["q_proj", "k_proj", "v_proj"],
    "gate_up_proj": ["up_proj", "down_proj"],
}

quant_config instance-attribute 

quant_config = quant_config

scheduler_config instance-attribute 

scheduler_config = scheduler_config

unpadded_vocab_size instance-attribute 

unpadded_vocab_size = vocab_size

vllm_config instance-attribute 

vllm_config = vllm_config

__init__ 

__init__(*, vllm_config: VllmConfig, prefix: str = '')
Source code in vllm/model_executor/models/bamba.py 
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
471
472
473
474
475
def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
    config = vllm_config.model_config.hf_config
    self.vllm_config = vllm_config
    self.model_config = vllm_config.model_config
    cache_config = vllm_config.cache_config
    lora_config = vllm_config.lora_config
    scheduler_config = vllm_config.scheduler_config
    assert not cache_config.enable_prefix_caching, \
        "Bamba currently does not support prefix caching"

    self.quant_config = vllm_config.quant_config

    super().__init__()
    self.config = config
    self.scheduler_config = scheduler_config
    self.model = BambaModel(vllm_config=vllm_config,
                            prefix=maybe_prefix(prefix, "model"))
    self.unpadded_vocab_size = config.vocab_size
    if lora_config:
        self.unpadded_vocab_size += lora_config.lora_extra_vocab_size
    self.lm_head = ParallelLMHead(
        self.unpadded_vocab_size,
        config.hidden_size,
        org_num_embeddings=config.vocab_size,
        padding_size=DEFAULT_VOCAB_PADDING_SIZE
        # We need bigger padding if using lora for kernel
        # compatibility
        if not lora_config else lora_config.lora_vocab_padding_size,
    )
    # Used to track and store by the Mamba cache between steps.
    self.mamba_cache: Optional[MambaCacheManager] = None

    self.logits_processor = LogitsProcessor(self.unpadded_vocab_size,
                                            config.vocab_size)

    self.make_empty_intermediate_tensors = (
        self.model.make_empty_intermediate_tensors)

_get_mamba_cache_shape 

_get_mamba_cache_shape() -> tuple[
    tuple[int, int], tuple[int, int]
]
Source code in vllm/model_executor/models/bamba.py 
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
def _get_mamba_cache_shape(
        self) -> tuple[tuple[int, int], tuple[int, int]]:
    world_size = get_tensor_model_parallel_world_size()
    hidden_size = self.config.hidden_size

    conv_state_shape, temporal_state_shape = None, None

    intermediate_size = self.config.mamba_expand * hidden_size

    # if n_groups is not divisible by world_size, need to extend the shards
    # to ensure all groups needed by a head is sharded along with it
    n_groups = (self.config.mamba_n_groups + extra_groups_for_head_shards(
        self.config.mamba_n_groups, world_size))

    # - heads and n_groups are TP-ed
    conv_dim = (intermediate_size +
                2 * n_groups * self.config.mamba_d_state)
    conv_state_shape = (
        divide(conv_dim, world_size),
        self.config.mamba_d_conv - 1,
    )

    # These are not TP-ed as they depend on A, dt_bias, D
    # - they are typically small
    #   e.g., (h_heads, d_head, d_state) = (128, 64, 128)
    temporal_state_shape = (
        divide(self.config.mamba_n_heads, world_size),
        self.config.mamba_d_head,
        self.config.mamba_d_state,
    )
    return conv_state_shape, temporal_state_shape

compute_logits 

compute_logits(
    hidden_states: Tensor,
    sampling_metadata: SamplingMetadata,
) -> Optional[Tensor]
Source code in vllm/model_executor/models/bamba.py 
546
547
548
549
550
551
552
553
def compute_logits(
    self,
    hidden_states: torch.Tensor,
    sampling_metadata: SamplingMetadata,
) -> Optional[torch.Tensor]:
    logits = self.logits_processor(self.lm_head, hidden_states,
                                   sampling_metadata)
    return logits

copy_inputs_before_cuda_graphs 

copy_inputs_before_cuda_graphs(input_buffers, **kwargs)
Source code in vllm/model_executor/models/bamba.py 
507
508
509
def copy_inputs_before_cuda_graphs(self, input_buffers, **kwargs):
    return self.mamba_cache.copy_inputs_before_cuda_graphs(
        input_buffers, **kwargs)

forward 

forward(
    input_ids: Tensor,
    positions: Tensor,
    intermediate_tensors: Optional[
        IntermediateTensors
    ] = None,
    inputs_embeds: Optional[Tensor] = None,
    **kwargs,
)
Source code in vllm/model_executor/models/bamba.py 
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
def forward(self,
            input_ids: torch.Tensor,
            positions: torch.Tensor,
            intermediate_tensors: Optional[IntermediateTensors] = None,
            inputs_embeds: Optional[torch.Tensor] = None,
            **kwargs):

    mamba_cache_params = None
    if not envs.VLLM_USE_V1:
        if self.mamba_cache is None:
            num_mamba_layers = \
                self.model_config.get_num_layers_by_block_type(
                    self.vllm_config.parallel_config,
                    LayerBlockType.mamba
                )

            self.mamba_cache = MambaCacheManager(
                self.vllm_config, self.lm_head.weight.dtype,
                num_mamba_layers, *self._get_mamba_cache_shape())

        mamba_cache_params = self.mamba_cache.current_run_tensors(**kwargs)

    hidden_states = self.model(input_ids, positions, mamba_cache_params,
                               intermediate_tensors, inputs_embeds)

    return hidden_states

get_input_embeddings 

get_input_embeddings(input_ids: Tensor) -> Tensor
Source code in vllm/model_executor/models/bamba.py 
477
478
def get_input_embeddings(self, input_ids: torch.Tensor) -> torch.Tensor:
    return self.model.get_input_embeddings(input_ids)

get_seqlen_agnostic_capture_inputs 

get_seqlen_agnostic_capture_inputs(batch_size: int)
Source code in vllm/model_executor/models/bamba.py 
511
512
def get_seqlen_agnostic_capture_inputs(self, batch_size: int):
    return self.mamba_cache.get_seqlen_agnostic_capture_inputs(batch_size)

load_weights 

load_weights(
    weights: Iterable[tuple[str, Tensor]],
) -> set[str]
Source code in vllm/model_executor/models/bamba.py 
555
556
557
558
def load_weights(self, weights: Iterable[tuple[str,
                                               torch.Tensor]]) -> set[str]:
    loader = AutoWeightsLoader(self)
    return loader.load_weights(weights)

BambaMLP

Bases: Module

Source code in vllm/model_executor/models/bamba.py 
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
class BambaMLP(nn.Module):

    def __init__(
        self,
        config: BambaConfig,
        quant_config: Optional[QuantizationConfig] = None,
        bias: bool = False,
    ) -> None:
        super().__init__()
        self.gate_up_proj = MergedColumnParallelLinear(
            input_size=config.hidden_size,
            output_sizes=[config.intermediate_size] * 2,
            bias=bias,
            quant_config=quant_config,
        )
        self.down_proj = RowParallelLinear(
            input_size=config.intermediate_size,
            output_size=config.hidden_size,
            bias=bias,
            quant_config=quant_config,
        )
        if config.hidden_act != "silu":
            raise ValueError(f"Unsupported activation: {config.hidden_act}. "
                             "Only silu is supported for now.")
        self.act_fn = SiluAndMul()

    def forward(self, x):
        x, _ = self.gate_up_proj(x)
        x = self.act_fn(x)
        x, _ = self.down_proj(x)
        return x

act_fn instance-attribute 

act_fn = SiluAndMul()

down_proj instance-attribute 

down_proj = RowParallelLinear(
    input_size=intermediate_size,
    output_size=hidden_size,
    bias=bias,
    quant_config=quant_config,
)

gate_up_proj instance-attribute 

gate_up_proj = MergedColumnParallelLinear(
    input_size=hidden_size,
    output_sizes=[intermediate_size] * 2,
    bias=bias,
    quant_config=quant_config,
)

__init__ 

__init__(
    config: BambaConfig,
    quant_config: Optional[QuantizationConfig] = None,
    bias: bool = False,
) -> None
Source code in vllm/model_executor/models/bamba.py 
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
def __init__(
    self,
    config: BambaConfig,
    quant_config: Optional[QuantizationConfig] = None,
    bias: bool = False,
) -> None:
    super().__init__()
    self.gate_up_proj = MergedColumnParallelLinear(
        input_size=config.hidden_size,
        output_sizes=[config.intermediate_size] * 2,
        bias=bias,
        quant_config=quant_config,
    )
    self.down_proj = RowParallelLinear(
        input_size=config.intermediate_size,
        output_size=config.hidden_size,
        bias=bias,
        quant_config=quant_config,
    )
    if config.hidden_act != "silu":
        raise ValueError(f"Unsupported activation: {config.hidden_act}. "
                         "Only silu is supported for now.")
    self.act_fn = SiluAndMul()

forward 

forward(x)
Source code in vllm/model_executor/models/bamba.py 
72
73
74
75
76
def forward(self, x):
    x, _ = self.gate_up_proj(x)
    x = self.act_fn(x)
    x, _ = self.down_proj(x)
    return x

BambaMixerDecoderLayer

Bases: Module

Source code in vllm/model_executor/models/bamba.py 
 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
class BambaMixerDecoderLayer(nn.Module):

    def __init__(self,
                 config: BambaConfig,
                 layer_idx: int,
                 cache_config: Optional[CacheConfig] = None,
                 quant_config: Optional[QuantizationConfig] = None,
                 prefix: str = "") -> None:
        super().__init__()
        self.config = config
        self.mamba = MambaMixer2(hidden_size= config.hidden_size,
                                ssm_state_size = config.mamba_d_state,
                                conv_kernel_size = config.mamba_d_conv,
                                intermediate_size = config.mamba_expand *\
                                                    config.hidden_size,
                                use_conv_bias = config.mamba_conv_bias,
                                use_bias = config.mamba_proj_bias,
                                n_groups=config.mamba_n_groups,
                                num_heads=config.mamba_n_heads,
                                head_dim=config.mamba_d_head,
                                rms_norm_eps=config.rms_norm_eps,
                                activation=config.hidden_act,
                                quant_config=quant_config,
                                prefix=f"{prefix}.mixer",
                                chunk_size=config.mamba_chunk_size)

        self.feed_forward = BambaMLP(config, quant_config=quant_config)
        self.input_layernorm = RMSNorm(config.hidden_size,
                                       eps=config.rms_norm_eps)
        self.pre_ff_layernorm = RMSNorm(config.hidden_size,
                                        eps=config.rms_norm_eps)

    def forward(
        self,
        hidden_states: torch.Tensor,
        residual: Optional[torch.Tensor],
        mamba_cache_params: MambaCacheParams,
        mamba2_metadata: Mamba2Metadata,
        **kwargs,
    ):
        if residual is None:
            residual = hidden_states
            hidden_states = self.input_layernorm(hidden_states)
        else:
            hidden_states, residual = self.input_layernorm(
                hidden_states, residual)

        hidden_states = self.mamba(hidden_states, mamba_cache_params,
                                   mamba2_metadata)
        # Fully Connected
        hidden_states, residual = self.pre_ff_layernorm(
            hidden_states, residual)
        hidden_states = self.feed_forward(hidden_states)
        return hidden_states, residual

config instance-attribute 

config = config

feed_forward instance-attribute 

feed_forward = BambaMLP(config, quant_config=quant_config)

input_layernorm instance-attribute 

input_layernorm = RMSNorm(hidden_size, eps=rms_norm_eps)

mamba instance-attribute 

mamba = MambaMixer2(
    hidden_size=hidden_size,
    ssm_state_size=mamba_d_state,
    conv_kernel_size=mamba_d_conv,
    intermediate_size=mamba_expand * hidden_size,
    use_conv_bias=mamba_conv_bias,
    use_bias=mamba_proj_bias,
    n_groups=mamba_n_groups,
    num_heads=mamba_n_heads,
    head_dim=mamba_d_head,
    rms_norm_eps=rms_norm_eps,
    activation=hidden_act,
    quant_config=quant_config,
    prefix=f"{prefix}.mixer",
    chunk_size=mamba_chunk_size,
)

pre_ff_layernorm instance-attribute 

pre_ff_layernorm = RMSNorm(hidden_size, eps=rms_norm_eps)

__init__ 

__init__(
    config: BambaConfig,
    layer_idx: int,
    cache_config: Optional[CacheConfig] = None,
    quant_config: Optional[QuantizationConfig] = None,
    prefix: str = "",
) -> None
Source code in vllm/model_executor/models/bamba.py 
 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
def __init__(self,
             config: BambaConfig,
             layer_idx: int,
             cache_config: Optional[CacheConfig] = None,
             quant_config: Optional[QuantizationConfig] = None,
             prefix: str = "") -> None:
    super().__init__()
    self.config = config
    self.mamba = MambaMixer2(hidden_size= config.hidden_size,
                            ssm_state_size = config.mamba_d_state,
                            conv_kernel_size = config.mamba_d_conv,
                            intermediate_size = config.mamba_expand *\
                                                config.hidden_size,
                            use_conv_bias = config.mamba_conv_bias,
                            use_bias = config.mamba_proj_bias,
                            n_groups=config.mamba_n_groups,
                            num_heads=config.mamba_n_heads,
                            head_dim=config.mamba_d_head,
                            rms_norm_eps=config.rms_norm_eps,
                            activation=config.hidden_act,
                            quant_config=quant_config,
                            prefix=f"{prefix}.mixer",
                            chunk_size=config.mamba_chunk_size)

    self.feed_forward = BambaMLP(config, quant_config=quant_config)
    self.input_layernorm = RMSNorm(config.hidden_size,
                                   eps=config.rms_norm_eps)
    self.pre_ff_layernorm = RMSNorm(config.hidden_size,
                                    eps=config.rms_norm_eps)

forward 

forward(
    hidden_states: Tensor,
    residual: Optional[Tensor],
    mamba_cache_params: MambaCacheParams,
    mamba2_metadata: Mamba2Metadata,
    **kwargs,
)
Source code in vllm/model_executor/models/bamba.py 
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
def forward(
    self,
    hidden_states: torch.Tensor,
    residual: Optional[torch.Tensor],
    mamba_cache_params: MambaCacheParams,
    mamba2_metadata: Mamba2Metadata,
    **kwargs,
):
    if residual is None:
        residual = hidden_states
        hidden_states = self.input_layernorm(hidden_states)
    else:
        hidden_states, residual = self.input_layernorm(
            hidden_states, residual)

    hidden_states = self.mamba(hidden_states, mamba_cache_params,
                               mamba2_metadata)
    # Fully Connected
    hidden_states, residual = self.pre_ff_layernorm(
        hidden_states, residual)
    hidden_states = self.feed_forward(hidden_states)
    return hidden_states, residual

BambaModel

Bases: Module

Source code in vllm/model_executor/models/bamba.py 
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
class BambaModel(nn.Module):

    def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
        super().__init__()

        config: BambaConfig = vllm_config.model_config.hf_config
        cache_config = vllm_config.cache_config
        quant_config = vllm_config.quant_config
        lora_config = vllm_config.lora_config

        self.config = config
        lora_vocab = ((lora_config.lora_extra_vocab_size *
                       (lora_config.max_loras or 1)) if lora_config else 0)
        self.vocab_size = config.vocab_size + lora_vocab
        self.org_vocab_size = config.vocab_size

        self.embed_tokens = VocabParallelEmbedding(
            self.vocab_size,
            config.hidden_size,
            org_num_embeddings=config.vocab_size,
        )

        def get_layer(prefix: str):
            layer_idx = int(prefix.rsplit(".", 1)[1])
            layer_class = ALL_DECODER_LAYER_TYPES[
                config.layers_block_type[layer_idx]]
            return layer_class(
                config,
                layer_idx,
                cache_config,
                quant_config=quant_config,
                prefix=prefix,
            )

        self.start_layer, self.end_layer, self.layers = make_layers(
            config.num_hidden_layers, get_layer, prefix=f"{prefix}.layers")
        self.make_empty_intermediate_tensors = (
            make_empty_intermediate_tensors_factory(
                ["hidden_states", "residual"], config.hidden_size))

        self.final_layernorm = RMSNorm(config.hidden_size,
                                       eps=config.rms_norm_eps)

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

    def forward(
        self,
        input_ids: torch.Tensor,
        positions: torch.Tensor,
        mamba_cache_params: MambaCacheParams,
        intermediate_tensors: Optional[IntermediateTensors] = None,
        inputs_embeds: Optional[torch.Tensor] = None,
    ) -> torch.Tensor:

        attn_metadata = get_forward_context().attn_metadata

        if not envs.VLLM_USE_V1:
            mamba2_metadata = prepare_mamba2_metadata(
                chunk_size=self.config.mamba_chunk_size,
                attn_metadata=attn_metadata,
            )
        else:
            # v1 get mamba2_metadata from forward_context
            mamba2_metadata = None

        if get_pp_group().is_first_rank:
            if inputs_embeds is not None:
                hidden_states = inputs_embeds
            else:
                hidden_states = self.get_input_embeddings(input_ids)
            residual = None
        else:
            assert intermediate_tensors is not None
            hidden_states = intermediate_tensors["hidden_states"]
            residual = intermediate_tensors["residual"]

        residual = None
        num_attn = 0
        for i in range(len(self.layers)):
            layer = self.layers[i]
            if isinstance(layer, BambaAttentionDecoderLayer):
                num_attn += 1

            layer_mamba_cache_params = None
            if isinstance(layer,
                          BambaMixerDecoderLayer) and mamba_cache_params:
                layer_mamba_cache_params = mamba_cache_params.at_layer_idx(
                    i - num_attn)

            hidden_states, residual = layer(
                positions=positions,
                hidden_states=hidden_states,
                residual=residual,
                mamba_cache_params=layer_mamba_cache_params,
                mamba2_metadata=mamba2_metadata,
            )

        if not get_pp_group().is_last_rank:
            return IntermediateTensors({
                "hidden_states": hidden_states,
                "residual": residual
            })
        hidden_states, _ = self.final_layernorm(hidden_states, residual)
        return hidden_states

    def load_weights(self, weights: Iterable[tuple[str,
                                                   torch.Tensor]]) -> set[str]:
        stacked_params_mapping = [
            # (param_name, shard_name, shard_id)
            ("qkv_proj", "q_proj", "q"),
            ("qkv_proj", "k_proj", "k"),
            ("qkv_proj", "v_proj", "v"),
            ("gate_up_proj", "gate_proj", 0),
            ("gate_up_proj", "up_proj", 1),
        ]

        params_dict = dict(self.named_parameters())
        loaded_params: set[str] = set()
        for name, loaded_weight in weights:
            if "rotary_emb.inv_freq" in name:
                continue

            if "A_log" in name:
                name = name.replace("A_log", "A")

            if ".self_attn." in name:
                name = name.replace(".self_attn", "")

            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
                # Skip layers on other devices.
                if is_pp_missing_parameter(name, self):
                    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
                if is_pp_missing_parameter(name, self):
                    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

config instance-attribute 

config = config

embed_tokens instance-attribute 

embed_tokens = VocabParallelEmbedding(
    vocab_size, hidden_size, org_num_embeddings=vocab_size
)

final_layernorm instance-attribute 

final_layernorm = RMSNorm(hidden_size, eps=rms_norm_eps)

make_empty_intermediate_tensors instance-attribute 

make_empty_intermediate_tensors = (
    make_empty_intermediate_tensors_factory(
        ["hidden_states", "residual"], hidden_size
    )
)

org_vocab_size instance-attribute 

org_vocab_size = vocab_size

vocab_size instance-attribute 

vocab_size = vocab_size + lora_vocab

__init__ 

__init__(*, vllm_config: VllmConfig, prefix: str = '')
Source code in vllm/model_executor/models/bamba.py 
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
def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
    super().__init__()

    config: BambaConfig = vllm_config.model_config.hf_config
    cache_config = vllm_config.cache_config
    quant_config = vllm_config.quant_config
    lora_config = vllm_config.lora_config

    self.config = config
    lora_vocab = ((lora_config.lora_extra_vocab_size *
                   (lora_config.max_loras or 1)) if lora_config else 0)
    self.vocab_size = config.vocab_size + lora_vocab
    self.org_vocab_size = config.vocab_size

    self.embed_tokens = VocabParallelEmbedding(
        self.vocab_size,
        config.hidden_size,
        org_num_embeddings=config.vocab_size,
    )

    def get_layer(prefix: str):
        layer_idx = int(prefix.rsplit(".", 1)[1])
        layer_class = ALL_DECODER_LAYER_TYPES[
            config.layers_block_type[layer_idx]]
        return layer_class(
            config,
            layer_idx,
            cache_config,
            quant_config=quant_config,
            prefix=prefix,
        )

    self.start_layer, self.end_layer, self.layers = make_layers(
        config.num_hidden_layers, get_layer, prefix=f"{prefix}.layers")
    self.make_empty_intermediate_tensors = (
        make_empty_intermediate_tensors_factory(
            ["hidden_states", "residual"], config.hidden_size))

    self.final_layernorm = RMSNorm(config.hidden_size,
                                   eps=config.rms_norm_eps)

forward 

forward(
    input_ids: Tensor,
    positions: Tensor,
    mamba_cache_params: MambaCacheParams,
    intermediate_tensors: Optional[
        IntermediateTensors
    ] = None,
    inputs_embeds: Optional[Tensor] = None,
) -> Tensor
Source code in vllm/model_executor/models/bamba.py 
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
def forward(
    self,
    input_ids: torch.Tensor,
    positions: torch.Tensor,
    mamba_cache_params: MambaCacheParams,
    intermediate_tensors: Optional[IntermediateTensors] = None,
    inputs_embeds: Optional[torch.Tensor] = None,
) -> torch.Tensor:

    attn_metadata = get_forward_context().attn_metadata

    if not envs.VLLM_USE_V1:
        mamba2_metadata = prepare_mamba2_metadata(
            chunk_size=self.config.mamba_chunk_size,
            attn_metadata=attn_metadata,
        )
    else:
        # v1 get mamba2_metadata from forward_context
        mamba2_metadata = None

    if get_pp_group().is_first_rank:
        if inputs_embeds is not None:
            hidden_states = inputs_embeds
        else:
            hidden_states = self.get_input_embeddings(input_ids)
        residual = None
    else:
        assert intermediate_tensors is not None
        hidden_states = intermediate_tensors["hidden_states"]
        residual = intermediate_tensors["residual"]

    residual = None
    num_attn = 0
    for i in range(len(self.layers)):
        layer = self.layers[i]
        if isinstance(layer, BambaAttentionDecoderLayer):
            num_attn += 1

        layer_mamba_cache_params = None
        if isinstance(layer,
                      BambaMixerDecoderLayer) and mamba_cache_params:
            layer_mamba_cache_params = mamba_cache_params.at_layer_idx(
                i - num_attn)

        hidden_states, residual = layer(
            positions=positions,
            hidden_states=hidden_states,
            residual=residual,
            mamba_cache_params=layer_mamba_cache_params,
            mamba2_metadata=mamba2_metadata,
        )

    if not get_pp_group().is_last_rank:
        return IntermediateTensors({
            "hidden_states": hidden_states,
            "residual": residual
        })
    hidden_states, _ = self.final_layernorm(hidden_states, residual)
    return hidden_states

get_input_embeddings 

get_input_embeddings(input_ids: Tensor) -> Tensor
Source code in vllm/model_executor/models/bamba.py 
305
306
def get_input_embeddings(self, input_ids: torch.Tensor) -> torch.Tensor:
    return self.embed_tokens(input_ids)

load_weights 

load_weights(
    weights: Iterable[tuple[str, Tensor]],
) -> set[str]
Source code in vllm/model_executor/models/bamba.py 
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
def load_weights(self, weights: Iterable[tuple[str,
                                               torch.Tensor]]) -> set[str]:
    stacked_params_mapping = [
        # (param_name, shard_name, shard_id)
        ("qkv_proj", "q_proj", "q"),
        ("qkv_proj", "k_proj", "k"),
        ("qkv_proj", "v_proj", "v"),
        ("gate_up_proj", "gate_proj", 0),
        ("gate_up_proj", "up_proj", 1),
    ]

    params_dict = dict(self.named_parameters())
    loaded_params: set[str] = set()
    for name, loaded_weight in weights:
        if "rotary_emb.inv_freq" in name:
            continue

        if "A_log" in name:
            name = name.replace("A_log", "A")

        if ".self_attn." in name:
            name = name.replace(".self_attn", "")

        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
            # Skip layers on other devices.
            if is_pp_missing_parameter(name, self):
                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
            if is_pp_missing_parameter(name, self):
                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