vllm.model_executor.models.phimoe

class PhiMoE(nn.Module):
    """A tensor-parallel MoE implementation for PhiMoE 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,
        )

        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=False,
            quant_config=quant_config,
            tp_size=tp_size,
            custom_routing_function=phimoe_routing_function,
            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)

class PhiMoEAttention(nn.Module):

    def __init__(
        self,
        hidden_size: int,
        num_heads: int,
        num_kv_heads: int,
        head_dim: Optional[int] = None,
        max_position: int = 4096 * 32,
        rope_theta: float = 10000,
        cache_config: Optional[CacheConfig] = None,
        quant_config: Optional[QuantizationConfig] = None,
        rope_scaling: Optional[dict] = 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)
        if head_dim is None:
            head_dim = hidden_size // num_heads
        self.head_dim = head_dim
        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.rope_scaling = rope_scaling

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

class PhiMoEConfig(PretrainedConfig):

    model_type = "phimoe"
    keys_to_ignore_at_inference = ["past_key_values"]

    def __init__(
        self,
        vocab_size=32000,
        hidden_size=4096,
        intermediate_size=14336,
        num_hidden_layers=32,
        num_attention_heads=32,
        num_key_value_heads=8,
        head_dim=None,
        hidden_act="silu",
        max_position_embeddings=4096 * 32,
        initializer_range=0.02,
        rms_norm_eps=1e-5,
        use_cache=True,
        pad_token_id=None,
        bos_token_id=1,
        eos_token_id=2,
        tie_word_embeddings=False,
        rope_theta=1e6,
        sliding_window=None,
        attention_dropout=0.0,
        num_experts_per_tok=2,
        num_local_experts=16,
        output_router_logits=False,
        router_aux_loss_coef=0.001,
        router_jitter_noise=0.0,
        attention_bias=False,
        lm_head_bias=False,
        **kwargs,
    ):
        self.vocab_size = vocab_size
        self.max_position_embeddings = max_position_embeddings
        self.hidden_size = hidden_size
        self.intermediate_size = intermediate_size
        self.num_hidden_layers = num_hidden_layers
        self.num_attention_heads = num_attention_heads
        self.sliding_window = sliding_window
        self.attention_bias = attention_bias
        self.lm_head_bias = lm_head_bias
        # for backward compatibility
        if num_key_value_heads is None:
            num_key_value_heads = num_attention_heads
        if head_dim is None:
            head_dim = hidden_size // num_attention_heads

        self.num_key_value_heads = num_key_value_heads
        self.head_dim = head_dim
        self.hidden_act = hidden_act
        self.initializer_range = initializer_range
        self.rms_norm_eps = rms_norm_eps
        self.use_cache = use_cache
        self.rope_theta = rope_theta
        self.attention_dropout = attention_dropout

        self.num_experts_per_tok = num_experts_per_tok
        self.num_local_experts = num_local_experts
        self.output_router_logits = output_router_logits
        self.router_aux_loss_coef = router_aux_loss_coef
        self.router_jitter_noise = router_jitter_noise
        super().__init__(
            pad_token_id=pad_token_id,
            bos_token_id=bos_token_id,
            eos_token_id=eos_token_id,
            tie_word_embeddings=tie_word_embeddings,
            **kwargs,
        )

class PhiMoEDecoderLayer(nn.Module):

    def __init__(
        self,
        config: PhiMoEConfig,
        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 = PhiMoEAttention(
            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,
            head_dim=getattr(config, "head_dim",
                             self.hidden_size // config.num_attention_heads),
            rope_theta=rope_theta,
            cache_config=cache_config,
            quant_config=quant_config,
            rope_scaling=config.rope_scaling,
            prefix=f"{prefix}.self_attn",
        )
        self.block_sparse_moe = PhiMoE(
            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 = nn.LayerNorm(config.hidden_size,
                                            eps=config.rms_norm_eps,
                                            elementwise_affine=True)
        self.post_attention_layernorm = nn.LayerNorm(config.hidden_size,
                                                     eps=config.rms_norm_eps,
                                                     elementwise_affine=True)

    def forward(
        self,
        positions: torch.Tensor,
        hidden_states: torch.Tensor,
        residual: Optional[torch.Tensor],
    ) -> torch.Tensor:
        residual = hidden_states

        # Self Attention
        hidden_states = self.input_layernorm(hidden_states)

        hidden_states = self.self_attn(
            positions=positions,
            hidden_states=hidden_states,
        )
        hidden_states = hidden_states + residual

        # Fully Connected
        residual = hidden_states
        hidden_states = self.post_attention_layernorm(hidden_states)
        hidden_states = self.block_sparse_moe(hidden_states)

        hidden_states = hidden_states + residual
        return hidden_states, residual

class PhiMoEForCausalLM(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
        lora_config = vllm_config.lora_config
        self.config = config
        self.lora_config = lora_config
        self.quant_config = vllm_config.quant_config

        self.model = PhiMoEModel(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=None,
            bias=True,
        )
        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,
    ) -> Union[torch.Tensor, IntermediateTensors]:
        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) -> 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)

    def get_expert_mapping(self) -> list[tuple[str, str, int, str]]:
        return self.model.get_expert_mapping()