Skip to content

vllm.model_executor.models.granitemoe

Inference-only GraniteMoe model.

GraniteMoeAttention

Bases: Module

Source code in vllm/model_executor/models/granitemoe.py 
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
class GraniteMoeAttention(nn.Module):

    def __init__(
        self,
        hidden_size: int,
        num_heads: int,
        num_kv_heads: int,
        max_position: int = 4096 * 32,
        rope_theta: float = 10000,
        cache_config: Optional[CacheConfig] = None,
        quant_config: Optional[QuantizationConfig] = None,
        attention_multiplier: Optional[float] = None,
        prefix: str = "",
    ) -> None:
        super().__init__()
        self.hidden_size = hidden_size
        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
        self.total_num_kv_heads = num_kv_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 = 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 = (attention_multiplier if attention_multiplier
                        is not None else self.head_dim**-1)
        self.rope_theta = rope_theta

        self.qkv_proj = QKVParallelLinear(
            hidden_size,
            self.head_dim,
            self.total_num_heads,
            self.total_num_kv_heads,
            bias=False,
            quant_config=quant_config,
            prefix=f"{prefix}.qkv_proj",
        )
        self.o_proj = RowParallelLinear(
            self.total_num_heads * self.head_dim,
            hidden_size,
            bias=False,
            quant_config=quant_config,
            prefix=f"{prefix}.o_proj",
        )
        self.rotary_emb = get_rope(
            self.head_dim,
            rotary_dim=self.head_dim,
            max_position=max_position,
            base=int(self.rope_theta),
            is_neox_style=True,
        )
        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")

    def forward(
        self,
        positions: torch.Tensor,
        hidden_states: torch.Tensor,
    ) -> 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

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",
)

head_dim instance-attribute 

head_dim = hidden_size // total_num_heads

hidden_size instance-attribute 

hidden_size = hidden_size

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_kv_heads // tp_size)

o_proj instance-attribute 

o_proj = RowParallelLinear(
    total_num_heads * head_dim,
    hidden_size,
    bias=False,
    quant_config=quant_config,
    prefix=f"{prefix}.o_proj",
)

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,
    prefix=f"{prefix}.qkv_proj",
)

rope_theta instance-attribute 

rope_theta = rope_theta

rotary_emb instance-attribute 

rotary_emb = get_rope(
    head_dim,
    rotary_dim=head_dim,
    max_position=max_position,
    base=int(rope_theta),
    is_neox_style=True,
)

scaling instance-attribute 

scaling = (
    attention_multiplier
    if attention_multiplier is not None
    else head_dim**-1
)

total_num_heads instance-attribute 

total_num_heads = num_heads

total_num_kv_heads instance-attribute 

total_num_kv_heads = num_kv_heads

__init__ 

__init__(
    hidden_size: int,
    num_heads: int,
    num_kv_heads: int,
    max_position: int = 4096 * 32,
    rope_theta: float = 10000,
    cache_config: Optional[CacheConfig] = None,
    quant_config: Optional[QuantizationConfig] = None,
    attention_multiplier: Optional[float] = None,
    prefix: str = "",
) -> None
Source code in vllm/model_executor/models/granitemoe.py 
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
def __init__(
    self,
    hidden_size: int,
    num_heads: int,
    num_kv_heads: int,
    max_position: int = 4096 * 32,
    rope_theta: float = 10000,
    cache_config: Optional[CacheConfig] = None,
    quant_config: Optional[QuantizationConfig] = None,
    attention_multiplier: Optional[float] = None,
    prefix: str = "",
) -> None:
    super().__init__()
    self.hidden_size = hidden_size
    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
    self.total_num_kv_heads = num_kv_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 = 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 = (attention_multiplier if attention_multiplier
                    is not None else self.head_dim**-1)
    self.rope_theta = rope_theta

    self.qkv_proj = QKVParallelLinear(
        hidden_size,
        self.head_dim,
        self.total_num_heads,
        self.total_num_kv_heads,
        bias=False,
        quant_config=quant_config,
        prefix=f"{prefix}.qkv_proj",
    )
    self.o_proj = RowParallelLinear(
        self.total_num_heads * self.head_dim,
        hidden_size,
        bias=False,
        quant_config=quant_config,
        prefix=f"{prefix}.o_proj",
    )
    self.rotary_emb = get_rope(
        self.head_dim,
        rotary_dim=self.head_dim,
        max_position=max_position,
        base=int(self.rope_theta),
        is_neox_style=True,
    )
    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")

forward 

forward(positions: Tensor, hidden_states: Tensor) -> Tensor
Source code in vllm/model_executor/models/granitemoe.py 
173
174
175
176
177
178
179
180
181
182
183
def forward(
    self,
    positions: torch.Tensor,
    hidden_states: torch.Tensor,
) -> 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

GraniteMoeDecoderLayer

Bases: Module

Source code in vllm/model_executor/models/granitemoe.py 
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
class GraniteMoeDecoderLayer(nn.Module):

    def __init__(
        self,
        config: GraniteMoeConfig,
        cache_config: Optional[CacheConfig] = None,
        quant_config: Optional[QuantizationConfig] = None,
        prefix: str = "",
    ) -> None:
        super().__init__()
        self.hidden_size = config.hidden_size
        # Requires transformers > 4.32.0
        rope_theta = getattr(config, "rope_theta", 10000)
        self.self_attn = GraniteMoeAttention(
            hidden_size=self.hidden_size,
            num_heads=config.num_attention_heads,
            max_position=config.max_position_embeddings,
            num_kv_heads=config.num_key_value_heads,
            rope_theta=rope_theta,
            cache_config=cache_config,
            quant_config=quant_config,
            prefix=f"{prefix}.self_attn",
            attention_multiplier=config.attention_multiplier)
        self.block_sparse_moe = GraniteMoeMoE(
            num_experts=config.num_local_experts,
            top_k=config.num_experts_per_tok,
            hidden_size=config.hidden_size,
            intermediate_size=config.intermediate_size,
            quant_config=quant_config,
            prefix=f"{prefix}.block_sparse_moe")

        self.input_layernorm = RMSNorm(config.hidden_size,
                                       eps=config.rms_norm_eps)
        self.post_attention_layernorm = RMSNorm(config.hidden_size,
                                                eps=config.rms_norm_eps)

        self.residual_multiplier = config.residual_multiplier

    def forward(
        self,
        positions: torch.Tensor,
        hidden_states: torch.Tensor,
    ) -> torch.Tensor:
        # Self Attention
        residual = hidden_states
        hidden_states = self.input_layernorm(hidden_states)
        hidden_states = self.self_attn(
            positions=positions,
            hidden_states=hidden_states,
        )
        hidden_states = residual + hidden_states * self.residual_multiplier
        residual = hidden_states
        hidden_states = self.post_attention_layernorm(hidden_states)
        hidden_states = self.block_sparse_moe(hidden_states)
        hidden_states = residual + hidden_states * self.residual_multiplier

        return hidden_states

block_sparse_moe instance-attribute 

block_sparse_moe = GraniteMoeMoE(
    num_experts=num_local_experts,
    top_k=num_experts_per_tok,
    hidden_size=hidden_size,
    intermediate_size=intermediate_size,
    quant_config=quant_config,
    prefix=f"{prefix}.block_sparse_moe",
)

hidden_size instance-attribute 

hidden_size = hidden_size

input_layernorm instance-attribute 

input_layernorm = RMSNorm(hidden_size, eps=rms_norm_eps)

post_attention_layernorm instance-attribute 

post_attention_layernorm = RMSNorm(
    hidden_size, eps=rms_norm_eps
)

residual_multiplier instance-attribute 

residual_multiplier = residual_multiplier

self_attn instance-attribute 

self_attn = GraniteMoeAttention(
    hidden_size=hidden_size,
    num_heads=num_attention_heads,
    max_position=max_position_embeddings,
    num_kv_heads=num_key_value_heads,
    rope_theta=rope_theta,
    cache_config=cache_config,
    quant_config=quant_config,
    prefix=f"{prefix}.self_attn",
    attention_multiplier=attention_multiplier,
)

__init__ 

__init__(
    config: GraniteMoeConfig,
    cache_config: Optional[CacheConfig] = None,
    quant_config: Optional[QuantizationConfig] = None,
    prefix: str = "",
) -> None
Source code in vllm/model_executor/models/granitemoe.py 
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
def __init__(
    self,
    config: GraniteMoeConfig,
    cache_config: Optional[CacheConfig] = None,
    quant_config: Optional[QuantizationConfig] = None,
    prefix: str = "",
) -> None:
    super().__init__()
    self.hidden_size = config.hidden_size
    # Requires transformers > 4.32.0
    rope_theta = getattr(config, "rope_theta", 10000)
    self.self_attn = GraniteMoeAttention(
        hidden_size=self.hidden_size,
        num_heads=config.num_attention_heads,
        max_position=config.max_position_embeddings,
        num_kv_heads=config.num_key_value_heads,
        rope_theta=rope_theta,
        cache_config=cache_config,
        quant_config=quant_config,
        prefix=f"{prefix}.self_attn",
        attention_multiplier=config.attention_multiplier)
    self.block_sparse_moe = GraniteMoeMoE(
        num_experts=config.num_local_experts,
        top_k=config.num_experts_per_tok,
        hidden_size=config.hidden_size,
        intermediate_size=config.intermediate_size,
        quant_config=quant_config,
        prefix=f"{prefix}.block_sparse_moe")

    self.input_layernorm = RMSNorm(config.hidden_size,
                                   eps=config.rms_norm_eps)
    self.post_attention_layernorm = RMSNorm(config.hidden_size,
                                            eps=config.rms_norm_eps)

    self.residual_multiplier = config.residual_multiplier

forward 

forward(positions: Tensor, hidden_states: Tensor) -> Tensor
Source code in vllm/model_executor/models/granitemoe.py 
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
def forward(
    self,
    positions: torch.Tensor,
    hidden_states: torch.Tensor,
) -> torch.Tensor:
    # Self Attention
    residual = hidden_states
    hidden_states = self.input_layernorm(hidden_states)
    hidden_states = self.self_attn(
        positions=positions,
        hidden_states=hidden_states,
    )
    hidden_states = residual + hidden_states * self.residual_multiplier
    residual = hidden_states
    hidden_states = self.post_attention_layernorm(hidden_states)
    hidden_states = self.block_sparse_moe(hidden_states)
    hidden_states = residual + hidden_states * self.residual_multiplier

    return hidden_states

GraniteMoeForCausalLM

Bases: Module, SupportsLoRA, SupportsPP

Source code in vllm/model_executor/models/granitemoe.py 
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
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
class GraniteMoeForCausalLM(nn.Module, SupportsLoRA, SupportsPP):
    fall_back_to_pt_during_load = False

    packed_modules_mapping = {
        "qkv_proj": [
            "q_proj",
            "k_proj",
            "v_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 = ""):
        super().__init__()
        config = vllm_config.model_config.hf_config
        quant_config = vllm_config.quant_config
        lora_config = vllm_config.lora_config

        self.config = config
        self.lora_config = lora_config

        self.model = GraniteMoeModel(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,
            quant_config=quant_config,
        )
        if config.tie_word_embeddings:
            self.lm_head.weight = self.model.embed_tokens.weight

        self.logits_processor = LogitsProcessor(self.unpadded_vocab_size,
                                                config.vocab_size,
                                                scale=1 /
                                                self.config.logits_scaling)

    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,
    ) -> torch.Tensor:
        hidden_states = self.model(input_ids, positions, intermediate_tensors,
                                   inputs_embeds)
        return hidden_states

    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 make_empty_intermediate_tensors(
            self, batch_size: int, dtype: torch.dtype,
            device: torch.device) -> IntermediateTensors:
        return IntermediateTensors({
            "hidden_states":
            torch.zeros((batch_size, self.config.hidden_size),
                        dtype=dtype,
                        device=device),
            "residual":
            torch.zeros((batch_size, self.config.hidden_size),
                        dtype=dtype,
                        device=device),
        })

    def load_weights(self, weights: Iterable[tuple[str,
                                                   torch.Tensor]]) -> set[str]:
        loader = AutoWeightsLoader(
            self,
            skip_prefixes=(["lm_head."]
                           if self.config.tie_word_embeddings else None),
        )
        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']

fall_back_to_pt_during_load class-attribute instance-attribute 

fall_back_to_pt_during_load = False

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,
    quant_config=quant_config,
)

logits_processor instance-attribute 

logits_processor = LogitsProcessor(
    unpadded_vocab_size,
    vocab_size,
    scale=1 / logits_scaling,
)

lora_config instance-attribute 

lora_config = lora_config

model instance-attribute 

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

packed_modules_mapping class-attribute instance-attribute 

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

unpadded_vocab_size instance-attribute 

unpadded_vocab_size = vocab_size

__init__ 

__init__(*, vllm_config: VllmConfig, prefix: str = '')
Source code in vllm/model_executor/models/granitemoe.py 
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
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

    self.config = config
    self.lora_config = lora_config

    self.model = GraniteMoeModel(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,
        quant_config=quant_config,
    )
    if config.tie_word_embeddings:
        self.lm_head.weight = self.model.embed_tokens.weight

    self.logits_processor = LogitsProcessor(self.unpadded_vocab_size,
                                            config.vocab_size,
                                            scale=1 /
                                            self.config.logits_scaling)

compute_logits 

compute_logits(
    hidden_states: Tensor,
    sampling_metadata: SamplingMetadata,
) -> Optional[Tensor]
Source code in vllm/model_executor/models/granitemoe.py 
409
410
411
412
413
414
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

forward 

forward(
    input_ids: Tensor,
    positions: Tensor,
    intermediate_tensors: Optional[
        IntermediateTensors
    ] = None,
    inputs_embeds: Optional[Tensor] = None,
) -> Tensor
Source code in vllm/model_executor/models/granitemoe.py 
398
399
400
401
402
403
404
405
406
407
def forward(
    self,
    input_ids: torch.Tensor,
    positions: torch.Tensor,
    intermediate_tensors: Optional[IntermediateTensors] = None,
    inputs_embeds: Optional[torch.Tensor] = None,
) -> torch.Tensor:
    hidden_states = self.model(input_ids, positions, intermediate_tensors,
                               inputs_embeds)
    return hidden_states

get_input_embeddings 

get_input_embeddings(input_ids: Tensor) -> Tensor
Source code in vllm/model_executor/models/granitemoe.py 
395
396
def get_input_embeddings(self, input_ids: torch.Tensor) -> torch.Tensor:
    return self.model.get_input_embeddings(input_ids)

load_weights 

load_weights(
    weights: Iterable[tuple[str, Tensor]],
) -> set[str]
Source code in vllm/model_executor/models/granitemoe.py 
430
431
432
433
434
435
436
437
def load_weights(self, weights: Iterable[tuple[str,
                                               torch.Tensor]]) -> set[str]:
    loader = AutoWeightsLoader(
        self,
        skip_prefixes=(["lm_head."]
                       if self.config.tie_word_embeddings else None),
    )
    return loader.load_weights(weights)

make_empty_intermediate_tensors 

make_empty_intermediate_tensors(
    batch_size: int, dtype: dtype, device: device
) -> IntermediateTensors
Source code in vllm/model_executor/models/granitemoe.py 
416
417
418
419
420
421
422
423
424
425
426
427
428
def make_empty_intermediate_tensors(
        self, batch_size: int, dtype: torch.dtype,
        device: torch.device) -> IntermediateTensors:
    return IntermediateTensors({
        "hidden_states":
        torch.zeros((batch_size, self.config.hidden_size),
                    dtype=dtype,
                    device=device),
        "residual":
        torch.zeros((batch_size, self.config.hidden_size),
                    dtype=dtype,
                    device=device),
    })

GraniteMoeMoE

Bases: Module

A tensor-parallel MoE implementation for GraniteMoe that shards each expert across all ranks. Each expert's weights are sharded across all ranks and a fused MoE kernel is used for the forward pass, and finally we reduce the outputs across ranks.

Source code in vllm/model_executor/models/granitemoe.py 
 56
 57
 58
 59
 60
 61
 62
 63
 64
 65
 66
 67
 68
 69
 70
 71
 72
 73
 74
 75
 76
 77
 78
 79
 80
 81
 82
 83
 84
 85
 86
 87
 88
 89
 90
 91
 92
 93
 94
 95
 96
 97
 98
 99
100
101
102
class GraniteMoeMoE(nn.Module):
    """A tensor-parallel MoE implementation for GraniteMoe that shards each
    expert across all ranks.
    Each expert's weights are sharded across all ranks and a fused MoE
    kernel is used for the forward pass, and finally we reduce the outputs
    across ranks.
    """

    def __init__(self,
                 num_experts: int,
                 top_k: int,
                 hidden_size: int,
                 intermediate_size: int,
                 params_dtype: Optional[torch.dtype] = None,
                 quant_config: Optional[QuantizationConfig] = None,
                 tp_size: Optional[int] = None,
                 prefix: str = ""):
        super().__init__()
        self.hidden_size = hidden_size

        # Gate always runs at half / full precision for now.
        self.gate = ReplicatedLinear(hidden_size,
                                     num_experts,
                                     bias=False,
                                     params_dtype=params_dtype,
                                     quant_config=None,
                                     prefix=f"{prefix}.gate")

        self.experts = FusedMoE(num_experts=num_experts,
                                top_k=top_k,
                                hidden_size=hidden_size,
                                intermediate_size=intermediate_size,
                                params_dtype=params_dtype,
                                reduce_results=True,
                                renormalize=True,
                                quant_config=quant_config,
                                tp_size=tp_size,
                                prefix=f"{prefix}.experts")

    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
        # NOTE: hidden_states can have either 1D or 2D shape.
        orig_shape = hidden_states.shape
        hidden_states = hidden_states.view(-1, self.hidden_size)
        # router_logits: (num_tokens, n_experts)
        router_logits, _ = self.gate(hidden_states)
        final_hidden_states = self.experts(hidden_states, router_logits)
        return final_hidden_states.view(orig_shape)

experts instance-attribute 

experts = FusedMoE(
    num_experts=num_experts,
    top_k=top_k,
    hidden_size=hidden_size,
    intermediate_size=intermediate_size,
    params_dtype=params_dtype,
    reduce_results=True,
    renormalize=True,
    quant_config=quant_config,
    tp_size=tp_size,
    prefix=f"{prefix}.experts",
)

gate instance-attribute 

gate = ReplicatedLinear(
    hidden_size,
    num_experts,
    bias=False,
    params_dtype=params_dtype,
    quant_config=None,
    prefix=f"{prefix}.gate",
)

hidden_size instance-attribute 

hidden_size = hidden_size

__init__ 

__init__(
    num_experts: int,
    top_k: int,
    hidden_size: int,
    intermediate_size: int,
    params_dtype: Optional[dtype] = None,
    quant_config: Optional[QuantizationConfig] = None,
    tp_size: Optional[int] = None,
    prefix: str = "",
)
Source code in vllm/model_executor/models/granitemoe.py 
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
def __init__(self,
             num_experts: int,
             top_k: int,
             hidden_size: int,
             intermediate_size: int,
             params_dtype: Optional[torch.dtype] = None,
             quant_config: Optional[QuantizationConfig] = None,
             tp_size: Optional[int] = None,
             prefix: str = ""):
    super().__init__()
    self.hidden_size = hidden_size

    # Gate always runs at half / full precision for now.
    self.gate = ReplicatedLinear(hidden_size,
                                 num_experts,
                                 bias=False,
                                 params_dtype=params_dtype,
                                 quant_config=None,
                                 prefix=f"{prefix}.gate")

    self.experts = FusedMoE(num_experts=num_experts,
                            top_k=top_k,
                            hidden_size=hidden_size,
                            intermediate_size=intermediate_size,
                            params_dtype=params_dtype,
                            reduce_results=True,
                            renormalize=True,
                            quant_config=quant_config,
                            tp_size=tp_size,
                            prefix=f"{prefix}.experts")

forward 

forward(hidden_states: Tensor) -> Tensor
Source code in vllm/model_executor/models/granitemoe.py 
 95
 96
 97
 98
 99
100
101
102
def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
    # NOTE: hidden_states can have either 1D or 2D shape.
    orig_shape = hidden_states.shape
    hidden_states = hidden_states.view(-1, self.hidden_size)
    # router_logits: (num_tokens, n_experts)
    router_logits, _ = self.gate(hidden_states)
    final_hidden_states = self.experts(hidden_states, router_logits)
    return final_hidden_states.view(orig_shape)

GraniteMoeModel

Bases: Module

Source code in vllm/model_executor/models/granitemoe.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
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
@support_torch_compile
class GraniteMoeModel(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
        lora_config = vllm_config.lora_config

        self.config = config
        self.quant_config = quant_config  # Required by MixtralModel
        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,
        )
        self.embedding_multiplier = config.embedding_multiplier

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

        self.norm = 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,
        intermediate_tensors: Optional[IntermediateTensors],
        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)
            hidden_states *= self.embedding_multiplier
            residual = None
        else:
            assert intermediate_tensors is not None
            hidden_states = intermediate_tensors["hidden_states"]
            residual = intermediate_tensors["residual"]
        for layer in self.layers[self.start_layer:self.end_layer]:
            hidden_states = layer(positions, hidden_states)
        if not get_pp_group().is_last_rank:
            return IntermediateTensors({
                "hidden_states": hidden_states,
                "residual": residual
            })
        hidden_states = self.norm(hidden_states)
        return hidden_states

    def load_weights(self, weights: Iterable[tuple[str,
                                                   torch.Tensor]]) -> set[str]:
        new_weights = {}
        for n, p in weights:
            if n.endswith('.block_sparse_moe.input_linear.weight'):
                for e in range(p.size(0)):
                    w1_name = n.replace(
                        '.block_sparse_moe.input_linear.weight',
                        f".block_sparse_moe.experts.{e}.w1.weight")
                    w3_name = n.replace(
                        '.block_sparse_moe.input_linear.weight',
                        f".block_sparse_moe.experts.{e}.w3.weight")
                    w1_param, w3_param = p[e].chunk(2, dim=0)
                    assert w1_name not in new_weights
                    assert w3_name not in new_weights
                    new_weights[w1_name] = w1_param
                    new_weights[w3_name] = w3_param
            elif n.endswith('.block_sparse_moe.output_linear.weight'):
                for e in range(p.size(0)):
                    w2_name = n.replace(
                        '.block_sparse_moe.output_linear.weight',
                        f".block_sparse_moe.experts.{e}.w2.weight")
                    w2_param = p[e]
                    assert w2_name not in new_weights
                    new_weights[w2_name] = w2_param
            elif n.endswith('.block_sparse_moe.router.layer.weight'):
                gate_name = n.replace('.block_sparse_moe.router.layer.weight',
                                      ".block_sparse_moe.gate.weight")
                assert gate_name not in new_weights
                new_weights[gate_name] = p
            else:
                new_weights[n] = p
        return mixtral.MixtralModel.load_weights(self, new_weights.items())

config instance-attribute 

config = config

embed_tokens instance-attribute 

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

embedding_multiplier instance-attribute 

embedding_multiplier = embedding_multiplier

norm instance-attribute 

norm = RMSNorm(hidden_size, eps=rms_norm_eps)

org_vocab_size instance-attribute 

org_vocab_size = vocab_size

quant_config instance-attribute 

quant_config = quant_config

vocab_size instance-attribute 

vocab_size = vocab_size + lora_vocab

__init__ 

__init__(*, vllm_config: VllmConfig, prefix: str = '')
Source code in vllm/model_executor/models/granitemoe.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
272
273
274
275
276
277
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
    lora_config = vllm_config.lora_config

    self.config = config
    self.quant_config = quant_config  # Required by MixtralModel
    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,
    )
    self.embedding_multiplier = config.embedding_multiplier

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

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

forward 

forward(
    input_ids: Tensor,
    positions: Tensor,
    intermediate_tensors: Optional[IntermediateTensors],
    inputs_embeds: Optional[Tensor] = None,
) -> Tensor
Source code in vllm/model_executor/models/granitemoe.py 
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
def forward(
    self,
    input_ids: torch.Tensor,
    positions: torch.Tensor,
    intermediate_tensors: Optional[IntermediateTensors],
    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)
        hidden_states *= self.embedding_multiplier
        residual = None
    else:
        assert intermediate_tensors is not None
        hidden_states = intermediate_tensors["hidden_states"]
        residual = intermediate_tensors["residual"]
    for layer in self.layers[self.start_layer:self.end_layer]:
        hidden_states = layer(positions, hidden_states)
    if not get_pp_group().is_last_rank:
        return IntermediateTensors({
            "hidden_states": hidden_states,
            "residual": residual
        })
    hidden_states = self.norm(hidden_states)
    return hidden_states

get_input_embeddings 

get_input_embeddings(input_ids: Tensor) -> Tensor
Source code in vllm/model_executor/models/granitemoe.py 
279
280
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/granitemoe.py 
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
def load_weights(self, weights: Iterable[tuple[str,
                                               torch.Tensor]]) -> set[str]:
    new_weights = {}
    for n, p in weights:
        if n.endswith('.block_sparse_moe.input_linear.weight'):
            for e in range(p.size(0)):
                w1_name = n.replace(
                    '.block_sparse_moe.input_linear.weight',
                    f".block_sparse_moe.experts.{e}.w1.weight")
                w3_name = n.replace(
                    '.block_sparse_moe.input_linear.weight',
                    f".block_sparse_moe.experts.{e}.w3.weight")
                w1_param, w3_param = p[e].chunk(2, dim=0)
                assert w1_name not in new_weights
                assert w3_name not in new_weights
                new_weights[w1_name] = w1_param
                new_weights[w3_name] = w3_param
        elif n.endswith('.block_sparse_moe.output_linear.weight'):
            for e in range(p.size(0)):
                w2_name = n.replace(
                    '.block_sparse_moe.output_linear.weight',
                    f".block_sparse_moe.experts.{e}.w2.weight")
                w2_param = p[e]
                assert w2_name not in new_weights
                new_weights[w2_name] = w2_param
        elif n.endswith('.block_sparse_moe.router.layer.weight'):
            gate_name = n.replace('.block_sparse_moe.router.layer.weight',
                                  ".block_sparse_moe.gate.weight")
            assert gate_name not in new_weights
            new_weights[gate_name] = p
        else:
            new_weights[n] = p
    return mixtral.MixtralModel.load_weights(self, new_weights.items())