Skip to content

vllm.model_executor.models.mamba2

PyTorch MAMBA2 model.

KVCache module-attribute 

KVCache = tuple[Tensor, Tensor]

Mamba2DecoderLayer

Bases: Module

Source code in vllm/model_executor/models/mamba2.py 
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
class Mamba2DecoderLayer(nn.Module):

    def __init__(self,
                 config: MambaConfig,
                 quant_config: Optional[QuantizationConfig] = None,
                 prefix: str = "") -> None:
        super().__init__()
        self.config = config
        self.mixer = MambaMixer2(hidden_size=config.hidden_size,
                                 ssm_state_size=config.state_size,
                                 conv_kernel_size=config.conv_kernel,
                                 intermediate_size=getattr(
                                     config, "intermediate_size",
                                     config.expand * config.hidden_size),
                                 use_conv_bias=config.use_conv_bias,
                                 use_bias=config.use_bias,
                                 n_groups=config.n_groups,
                                 num_heads=config.num_heads,
                                 head_dim=config.head_dim,
                                 rms_norm_eps=config.layer_norm_epsilon,
                                 activation=config.hidden_act,
                                 quant_config=quant_config,
                                 prefix=f"{prefix}.mixer",
                                 chunk_size=config.chunk_size)

        self.norm = RMSNorm(config.hidden_size, eps=config.layer_norm_epsilon)

    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.norm(hidden_states)
        else:
            hidden_states, residual = self.norm(hidden_states, residual)

        hidden_states = self.mixer(hidden_states, mamba_cache_params,
                                   mamba2_metadata)
        return hidden_states, residual

config instance-attribute 

config = config

mixer instance-attribute 

mixer = MambaMixer2(
    hidden_size=hidden_size,
    ssm_state_size=state_size,
    conv_kernel_size=conv_kernel,
    intermediate_size=getattr(
        config, "intermediate_size", expand * hidden_size
    ),
    use_conv_bias=use_conv_bias,
    use_bias=use_bias,
    n_groups=n_groups,
    num_heads=num_heads,
    head_dim=head_dim,
    rms_norm_eps=layer_norm_epsilon,
    activation=hidden_act,
    quant_config=quant_config,
    prefix=f"{prefix}.mixer",
    chunk_size=chunk_size,
)

norm instance-attribute 

norm = RMSNorm(hidden_size, eps=layer_norm_epsilon)

__init__ 

__init__(
    config: MambaConfig,
    quant_config: Optional[QuantizationConfig] = None,
    prefix: str = "",
) -> None
Source code in vllm/model_executor/models/mamba2.py 
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
def __init__(self,
             config: MambaConfig,
             quant_config: Optional[QuantizationConfig] = None,
             prefix: str = "") -> None:
    super().__init__()
    self.config = config
    self.mixer = MambaMixer2(hidden_size=config.hidden_size,
                             ssm_state_size=config.state_size,
                             conv_kernel_size=config.conv_kernel,
                             intermediate_size=getattr(
                                 config, "intermediate_size",
                                 config.expand * config.hidden_size),
                             use_conv_bias=config.use_conv_bias,
                             use_bias=config.use_bias,
                             n_groups=config.n_groups,
                             num_heads=config.num_heads,
                             head_dim=config.head_dim,
                             rms_norm_eps=config.layer_norm_epsilon,
                             activation=config.hidden_act,
                             quant_config=quant_config,
                             prefix=f"{prefix}.mixer",
                             chunk_size=config.chunk_size)

    self.norm = RMSNorm(config.hidden_size, eps=config.layer_norm_epsilon)

forward 

forward(
    hidden_states: Tensor,
    residual: Optional[Tensor],
    mamba_cache_params: MambaCacheParams,
    mamba2_metadata: Mamba2Metadata,
    **kwargs,
)
Source code in vllm/model_executor/models/mamba2.py 
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
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.norm(hidden_states)
    else:
        hidden_states, residual = self.norm(hidden_states, residual)

    hidden_states = self.mixer(hidden_states, mamba_cache_params,
                               mamba2_metadata)
    return hidden_states, residual

Mamba2ForCausalLM

Bases: Module, HasInnerState, IsAttentionFree

Source code in vllm/model_executor/models/mamba2.py 
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
class Mamba2ForCausalLM(nn.Module, HasInnerState, IsAttentionFree):

    def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
        config = vllm_config.model_config.hf_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, \
            "Mamba does not support prefix caching"

        super().__init__()
        self.config = config
        self.vllm_config = vllm_config
        self.scheduler_config = scheduler_config
        self.model_config = vllm_config.model_config
        self.backbone = Mamba2Model(vllm_config=vllm_config,
                                    prefix=maybe_prefix(prefix, "backbone"))
        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,
        )
        if config.tie_word_embeddings:
            self.lm_head = self.lm_head.tie_weights(self.backbone.embeddings)

        # 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.backbone.make_empty_intermediate_tensors)

    def get_input_embeddings(self, input_ids: torch.Tensor) -> torch.Tensor:
        return self.backbone.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):
        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)
        else:
            # NOTE: mamba_cache_params is not needed for v1
            mamba_cache_params = None

        hidden_states = self.backbone(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()

        conv_state_shape, temporal_state_shape = None, None

        intermediate_size = getattr(
            self.config, "intermediate_size",
            self.config.expand * self.config.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.n_groups +
            extra_groups_for_head_shards(self.config.n_groups, world_size))

        # - heads and n_groups are TP-ed
        conv_dim = (intermediate_size + 2 * n_groups * self.config.state_size)
        conv_state_shape = (
            divide(conv_dim, world_size),
            self.config.conv_kernel - 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.num_heads, world_size),
            self.config.head_dim,
            self.config.state_size,
        )
        return conv_state_shape, temporal_state_shape

    def compute_logits(self, hidden_states: torch.Tensor,
                       sampling_metadata: SamplingMetadata) -> 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)

backbone instance-attribute 

backbone = Mamba2Model(
    vllm_config=vllm_config,
    prefix=maybe_prefix(prefix, "backbone"),
)

config instance-attribute 

config = config

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_config instance-attribute 

model_config = model_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/mamba2.py 
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
def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
    config = vllm_config.model_config.hf_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, \
        "Mamba does not support prefix caching"

    super().__init__()
    self.config = config
    self.vllm_config = vllm_config
    self.scheduler_config = scheduler_config
    self.model_config = vllm_config.model_config
    self.backbone = Mamba2Model(vllm_config=vllm_config,
                                prefix=maybe_prefix(prefix, "backbone"))
    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,
    )
    if config.tie_word_embeddings:
        self.lm_head = self.lm_head.tie_weights(self.backbone.embeddings)

    # 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.backbone.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/mamba2.py 
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
def _get_mamba_cache_shape(
        self) -> tuple[tuple[int, int], tuple[int, int]]:
    world_size = get_tensor_model_parallel_world_size()

    conv_state_shape, temporal_state_shape = None, None

    intermediate_size = getattr(
        self.config, "intermediate_size",
        self.config.expand * self.config.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.n_groups +
        extra_groups_for_head_shards(self.config.n_groups, world_size))

    # - heads and n_groups are TP-ed
    conv_dim = (intermediate_size + 2 * n_groups * self.config.state_size)
    conv_state_shape = (
        divide(conv_dim, world_size),
        self.config.conv_kernel - 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.num_heads, world_size),
        self.config.head_dim,
        self.config.state_size,
    )
    return conv_state_shape, temporal_state_shape

compute_logits 

compute_logits(
    hidden_states: Tensor,
    sampling_metadata: SamplingMetadata,
) -> Tensor
Source code in vllm/model_executor/models/mamba2.py 
311
312
313
314
315
def compute_logits(self, hidden_states: torch.Tensor,
                   sampling_metadata: SamplingMetadata) -> 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/mamba2.py 
271
272
273
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/mamba2.py 
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
def forward(self,
            input_ids: torch.Tensor,
            positions: torch.Tensor,
            intermediate_tensors: Optional[IntermediateTensors] = None,
            inputs_embeds: Optional[torch.Tensor] = None,
            **kwargs):
    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)
    else:
        # NOTE: mamba_cache_params is not needed for v1
        mamba_cache_params = None

    hidden_states = self.backbone(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/mamba2.py 
242
243
def get_input_embeddings(self, input_ids: torch.Tensor) -> torch.Tensor:
    return self.backbone.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/mamba2.py 
275
276
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/mamba2.py 
317
318
319
320
def load_weights(self, weights: Iterable[tuple[str,
                                               torch.Tensor]]) -> set[str]:
    loader = AutoWeightsLoader(self)
    return loader.load_weights(weights)

Mamba2Model

Bases: Module

Source code in vllm/model_executor/models/mamba2.py 
 89
 90
 91
 92
 93
 94
 95
 96
 97
 98
 99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
class Mamba2Model(nn.Module):

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

        config = vllm_config.model_config.hf_config
        quant_config = vllm_config.quant_config
        lora_config = vllm_config.lora_config
        is_lora_enabled = bool(lora_config)
        assert not is_lora_enabled

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

        self.start_layer, self.end_layer, self.layers = make_layers(
            config.num_hidden_layers,
            lambda prefix: Mamba2DecoderLayer(
                config, quant_config=quant_config, prefix=prefix),
            prefix=f"{prefix}.layers")

        self.norm_f = RMSNorm(config.hidden_size,
                              eps=config.layer_norm_epsilon)
        self.make_empty_intermediate_tensors = (
            make_empty_intermediate_tensors_factory(
                ["hidden_states", "residual"], config.hidden_size))

    def get_input_embeddings(self, input_ids: torch.Tensor) -> torch.Tensor:
        return self.embeddings(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:
        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"]

        attn_metadata: AttentionMetadata = get_forward_context().attn_metadata

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

        for i in range(len(self.layers)):
            layer = self.layers[i]

            hidden_states, residual = layer(
                positions=positions,
                hidden_states=hidden_states,
                residual=residual,
                mamba_cache_params=mamba_cache_params.at_layer_idx(
                    i - self.start_layer) if mamba_cache_params else None,
                mamba2_metadata=mamba2_metadata)

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

        hidden_states, _ = self.norm_f(hidden_states, residual)

        return hidden_states

    def load_weights(self, weights: Iterable[tuple[str,
                                                   torch.Tensor]]) -> set[str]:
        params_dict = dict(self.named_parameters())
        loaded_params: set[str] = set()
        for name, loaded_weight in weights:
            if "A_log" in name:
                name = name.replace("A_log", "A")

            # 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

embeddings instance-attribute 

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

make_empty_intermediate_tensors instance-attribute 

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

norm_f instance-attribute 

norm_f = RMSNorm(hidden_size, eps=layer_norm_epsilon)

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/mamba2.py 
 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
def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
    super().__init__()

    config = vllm_config.model_config.hf_config
    quant_config = vllm_config.quant_config
    lora_config = vllm_config.lora_config
    is_lora_enabled = bool(lora_config)
    assert not is_lora_enabled

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

    self.start_layer, self.end_layer, self.layers = make_layers(
        config.num_hidden_layers,
        lambda prefix: Mamba2DecoderLayer(
            config, quant_config=quant_config, prefix=prefix),
        prefix=f"{prefix}.layers")

    self.norm_f = RMSNorm(config.hidden_size,
                          eps=config.layer_norm_epsilon)
    self.make_empty_intermediate_tensors = (
        make_empty_intermediate_tensors_factory(
            ["hidden_states", "residual"], config.hidden_size))

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/mamba2.py 
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
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:
    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"]

    attn_metadata: AttentionMetadata = get_forward_context().attn_metadata

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

    for i in range(len(self.layers)):
        layer = self.layers[i]

        hidden_states, residual = layer(
            positions=positions,
            hidden_states=hidden_states,
            residual=residual,
            mamba_cache_params=mamba_cache_params.at_layer_idx(
                i - self.start_layer) if mamba_cache_params else None,
            mamba2_metadata=mamba2_metadata)

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

    hidden_states, _ = self.norm_f(hidden_states, residual)

    return hidden_states

get_input_embeddings 

get_input_embeddings(input_ids: Tensor) -> Tensor
Source code in vllm/model_executor/models/mamba2.py 
124
125
def get_input_embeddings(self, input_ids: torch.Tensor) -> torch.Tensor:
    return self.embeddings(input_ids)

load_weights 

load_weights(
    weights: Iterable[tuple[str, Tensor]],
) -> set[str]
Source code in vllm/model_executor/models/mamba2.py 
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
def load_weights(self, weights: Iterable[tuple[str,
                                               torch.Tensor]]) -> set[str]:
    params_dict = dict(self.named_parameters())
    loaded_params: set[str] = set()
    for name, loaded_weight in weights:
        if "A_log" in name:
            name = name.replace("A_log", "A")

        # 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