srt/models/phimoe.py

from typing import Iterable, Optional, Tuple, Union

import torch
from torch import nn
from transformers.configuration_utils import PretrainedConfig

from sglang.srt.distributed import get_tensor_model_parallel_world_size
from sglang.srt.layers.dp_attention import get_attention_tp_rank, get_attention_tp_size
from sglang.srt.layers.linear import (
    QKVParallelLinear,
    ReplicatedLinear,
    RowParallelLinear,
)
from sglang.srt.layers.logits_processor import LogitsProcessor, LogitsProcessorOutput
from sglang.srt.layers.moe.fused_moe_triton import FusedMoE
from sglang.srt.layers.moe.topk import TopK
from sglang.srt.layers.pooler import Pooler, PoolingType
from sglang.srt.layers.quantization.base_config import QuantizationConfig
from sglang.srt.layers.radix_attention import RadixAttention
from sglang.srt.layers.rotary_embedding import get_rope
from sglang.srt.layers.vocab_parallel_embedding import (
    DEFAULT_VOCAB_PADDING_SIZE,
    ParallelLMHead,
    VocabParallelEmbedding,
)
from sglang.srt.model_executor.forward_batch_info import ForwardBatch
from sglang.srt.model_loader.weight_utils import (
    default_weight_loader,
    maybe_remap_kv_scale_name,
)
from sglang.srt.utils import add_prefix, make_layers


class PhiMoEConfig(PretrainedConfig):

    model_type = "phimoe"

    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,
        )


def sparsemixer(scores, jitter_eps=0.01):
    ################ Select first expert (topk=2) ################

    # compute mask for sparsity
    mask_logits_threshold, max_ind = scores.max(dim=-1, keepdim=True)
    factor = scores.abs().clamp(min=mask_logits_threshold)
    mask_logits_threshold = ((mask_logits_threshold - scores) / factor) > (
        2 * jitter_eps
    )

    # apply mask
    masked_gates = scores.masked_fill(mask_logits_threshold, float("-inf"))
    selected_experts = max_ind

    # compute scores for gradients
    masked_gates = torch.softmax(masked_gates, dim=-1)
    multiplier_o = masked_gates.gather(dim=-1, index=selected_experts)

    multiplier = multiplier_o

    # masked out first expert
    masked_scores = torch.scatter(
        scores,
        -1,
        selected_experts,
        float("-inf"),
    )

    ################ Select second expert (topk=2) ################
    # compute mask for sparsity
    mask_logits_threshold, max_ind = masked_scores.max(dim=-1, keepdim=True)
    factor = scores.abs().clamp(min=mask_logits_threshold)
    mask_logits_threshold = ((mask_logits_threshold - scores) / factor) > (
        2 * jitter_eps
    )

    # apply mask
    masked_gates_top2 = masked_scores.masked_fill(mask_logits_threshold, float("-inf"))
    selected_experts_top2 = max_ind
    # compute scores for gradients
    masked_gates_top2 = torch.softmax(masked_gates_top2, dim=-1)
    multiplier_top2 = masked_gates_top2.gather(dim=-1, index=selected_experts_top2)

    multiplier = torch.concat((multiplier, multiplier_top2), dim=-1)
    selected_experts = torch.concat((selected_experts, selected_experts_top2), dim=-1)

    return (
        multiplier,
        selected_experts,
    )


def phimoe_routing_function(
    hidden_states: torch.Tensor,
    gating_output: torch.Tensor,
    topk: int,
    renormalize: bool,
):
    assert hidden_states.shape[0] == gating_output.shape[0], "Number of tokens mismatch"
    assert topk == 2, "Only top-2 routing is supported"
    assert renormalize is False, "Renormalization is not supported"

    topk_weights, topk_ids = sparsemixer(gating_output)
    return topk_weights, topk_ids


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,
        layer_id: int,
        quant_config: Optional[QuantizationConfig] = None,
        prefix: str = "",
    ):
        super().__init__()
        self.hidden_size = hidden_size
        self.tp_size = get_tensor_model_parallel_world_size()

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

        self.topk = TopK(
            top_k=top_k,
            renormalize=False,
            custom_routing_function=phimoe_routing_function,
        )

        self.experts = FusedMoE(
            num_experts=num_experts,
            top_k=top_k,
            layer_id=layer_id,
            hidden_size=hidden_size,
            intermediate_size=intermediate_size,
            reduce_results=True,
            quant_config=quant_config,
            prefix=add_prefix("experts", prefix),
        )

    def forward(
        self, hidden_states: torch.Tensor, forward_batch: Optional[ForwardBatch] = None
    ) -> 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, _ = self.gate(hidden_states)
        topk_output = self.topk(hidden_states, router_logits)
        final_hidden_states = self.experts(hidden_states, topk_output)
        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,
        layer_id: int = 0,
        attention_bias: bool = False,
        quant_config: Optional[QuantizationConfig] = None,
        rope_scaling: Optional[dict] = None,
        prefix: str = "",
    ) -> None:
        super().__init__()
        self.hidden_size = hidden_size

        attn_tp_rank = get_attention_tp_rank()
        attn_tp_size = get_attention_tp_size()

        self.total_num_heads = num_heads
        assert self.total_num_heads % attn_tp_size == 0
        self.num_heads = self.total_num_heads // attn_tp_size
        self.total_num_kv_heads = num_kv_heads
        if self.total_num_kv_heads >= attn_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 % attn_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 attn_tp_size % self.total_num_kv_heads == 0
        self.num_kv_heads = max(1, self.total_num_kv_heads // attn_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=attention_bias,
            quant_config=quant_config,
            tp_rank=attn_tp_rank,
            tp_size=attn_tp_size,
            prefix=add_prefix("qkv_proj", prefix),
        )
        self.o_proj = RowParallelLinear(
            self.total_num_heads * self.head_dim,
            hidden_size,
            bias=attention_bias,
            quant_config=quant_config,
            tp_rank=attn_tp_rank,
            tp_size=attn_tp_size,
            prefix=add_prefix("o_proj", prefix),
        )
        self.rotary_emb = get_rope(
            self.head_dim,
            rotary_dim=self.head_dim,
            max_position=max_position,
            base=int(self.rope_theta),
            rope_scaling=self.rope_scaling,
        )
        self.attn = RadixAttention(
            self.num_heads,
            self.head_dim,
            self.scaling,
            num_kv_heads=self.num_kv_heads,
            layer_id=layer_id,
            quant_config=quant_config,
            prefix=add_prefix("attn", prefix),
        )

    def forward(
        self,
        positions: torch.Tensor,
        hidden_states: torch.Tensor,
        forward_batch: ForwardBatch,
    ) -> 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, forward_batch)
        output, _ = self.o_proj(attn_output)
        return output


class PhiMoEDecoderLayer(nn.Module):

    def __init__(
        self,
        config: PhiMoEConfig,
        layer_id: int,
        quant_config: Optional[QuantizationConfig] = None,
        prefix: str = "",
    ) -> None:
        super().__init__()
        self.hidden_size = config.hidden_size
        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,
            layer_id=layer_id,
            attention_bias=config.attention_bias,
            quant_config=quant_config,
            rope_scaling=config.rope_scaling,
            prefix=add_prefix("self_attn", prefix),
        )
        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,
            layer_id=layer_id,
            quant_config=quant_config,
            prefix=add_prefix("block_sparse_moe", prefix),
        )
        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],
        forward_batch: ForwardBatch,
    ) -> Tuple[torch.Tensor, torch.Tensor]:
        residual = hidden_states

        hidden_states = self.input_layernorm(hidden_states)

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

        residual = hidden_states
        hidden_states = self.post_attention_layernorm(hidden_states)
        hidden_states = self.block_sparse_moe(
            hidden_states, forward_batch=forward_batch
        )

        hidden_states = hidden_states + residual
        return hidden_states, residual


class PhiMoEModel(nn.Module):

    def __init__(
        self,
        config: PhiMoEConfig,
        quant_config: Optional[QuantizationConfig] = None,
        prefix: str = "",
    ):
        super().__init__()

        self.config = config
        self.quant_config = quant_config
        self.vocab_size = config.vocab_size
        self.embed_tokens = VocabParallelEmbedding(
            config.vocab_size,
            config.hidden_size,
            quant_config=quant_config,
            prefix=add_prefix("embed_tokens", prefix),
        )

        self.layers = make_layers(
            config.num_hidden_layers,
            lambda idx, prefix: PhiMoEDecoderLayer(
                config, int(prefix.split(".")[-1]), quant_config, prefix=prefix
            ),
            prefix=add_prefix("layers", prefix),
        )
        self.norm = nn.LayerNorm(
            config.hidden_size, eps=config.rms_norm_eps, elementwise_affine=True
        )

    def forward(
        self,
        input_ids: torch.Tensor,
        positions: torch.Tensor,
        forward_batch: ForwardBatch,
        input_embeds: Optional[torch.Tensor] = None,
    ) -> Union[torch.Tensor]:
        if input_embeds is None:
            hidden_states = self.embed_tokens(input_ids)
        else:
            hidden_states = input_embeds
        residual = None

        for layer in self.layers:
            hidden_states, residual = layer(
                positions, hidden_states, residual, forward_batch=forward_batch
            )

        hidden_states = self.norm(hidden_states)
        return hidden_states


class PhiMoEForCausalLM(nn.Module):

    def __init__(
        self,
        config: PhiMoEConfig,
        quant_config: Optional[QuantizationConfig] = None,
        prefix: str = "",
    ):

        super().__init__()
        self.config = config
        self.quant_config = quant_config

        self.model = PhiMoEModel(
            config=config, quant_config=quant_config, prefix=add_prefix("model", prefix)
        )
        self.lm_head = ParallelLMHead(
            config.vocab_size,
            config.hidden_size,
            org_num_embeddings=config.vocab_size,
            padding_size=DEFAULT_VOCAB_PADDING_SIZE,
            quant_config=quant_config,
            bias=True,
            prefix=add_prefix("lm_head", prefix),
        )
        if self.config.tie_word_embeddings:
            self.lm_head.weight = self.model.embed_tokens.weight
        self.logits_processor = LogitsProcessor(config)
        self.pooler = Pooler(pooling_type=PoolingType.LAST, normalize=True)

    @torch.no_grad()
    def forward(
        self,
        input_ids: torch.Tensor,
        positions: torch.Tensor,
        forward_batch: ForwardBatch,
        inputs_embeds: Optional[torch.Tensor] = None,
        get_embedding: bool = False,
    ) -> LogitsProcessorOutput:
        hidden_states = self.model(input_ids, positions, forward_batch, inputs_embeds)

        if not get_embedding:
            return self.logits_processor(
                input_ids, hidden_states, self.lm_head, forward_batch
            )

        else:
            return self.pooler(hidden_states, forward_batch)

    def load_weights(self, weights: Iterable[Tuple[str, torch.Tensor]]):
        stacked_params_mapping = [
            # (param_name, shard_name, shard_id)
            ("qkv_proj", "q_proj", "q"),
            ("qkv_proj", "k_proj", "k"),
            ("qkv_proj", "v_proj", "v"),
        ]

        expert_params_mapping = FusedMoE.make_expert_params_mapping(
            ckpt_gate_proj_name="w1",
            ckpt_down_proj_name="w2",
            ckpt_up_proj_name="w3",
            num_experts=self.config.num_local_experts,
        )

        params_dict = dict(self.named_parameters())
        for name, loaded_weight in weights:
            for param_name, weight_name, shard_id in stacked_params_mapping:
                if weight_name not in name:
                    continue
                name = name.replace(weight_name, param_name)
                if name.endswith(".bias") and name not in params_dict:
                    continue
                param = params_dict[name]
                weight_loader = param.weight_loader
                weight_loader(param, loaded_weight, shard_id)
                break
            else:
                for mapping in expert_params_mapping:
                    param_name, weight_name, expert_id, shard_id = mapping
                    if weight_name not in name:
                        continue
                    name = name.replace(weight_name, param_name)
                    param = params_dict[name]
                    weight_loader = param.weight_loader
                    weight_loader(
                        param,
                        loaded_weight,
                        name,
                        shard_id=shard_id,
                        expert_id=expert_id,
                    )
                    break
                else:
                    if name.endswith(".bias") and name not in params_dict:
                        continue
                    # Remapping the name of FP8 kv-scale.
                    name = maybe_remap_kv_scale_name(name, params_dict)
                    if name is None:
                        continue

                    param = params_dict[name]
                    weight_loader = getattr(
                        param, "weight_loader", default_weight_loader
                    )
                    weight_loader(param, loaded_weight)


EntryClass = PhiMoEForCausalLM