server/text_generation_server/models/custom_modeling/flash_phi_modeling.py

import torch
import torch.distributed

from torch import nn
from transformers.activations import ACT2FN
from transformers.configuration_utils import PretrainedConfig
from typing import Optional, List, Tuple

from text_generation_server.layers.attention import (
    paged_attention,
    attention,
    Seqlen,
)
from text_generation_server.layers import (
    TensorParallelRowLinear,
    TensorParallelColumnLinear,
    TensorParallelEmbedding,
    SpeculativeHead,
    get_linear,
)
from text_generation_server.layers.attention.kv_cache import get_kv_scales
from text_generation_server.layers.layernorm import (
    FastLayerNorm,
)
from text_generation_server.layers.rotary import (
    PositionRotaryEmbedding,
)


class PhiConfig(PretrainedConfig):
    def __init__(
        self,
        vocab_size=51200,
        hidden_size=2560,
        num_hidden_layers=32,
        num_attention_heads=32,
        num_key_value_heads=32,
        hidden_act="gelu_fast",  # llama uses silu
        layer_norm_eps=1e-05,  # rms in llama,
        pad_token_id=0,
        bos_token_id=1,
        eos_token_id=2,
        tie_word_embeddings=False,
        rope_theta=10000.0,
        resid_pdrop=0.1,  # llama doesn't have this
        partial_rotary_factor=0.5,  # important difference between llama and phi
        **kwargs,
    ):
        self.vocab_size = vocab_size
        self.hidden_size = hidden_size
        self.num_hidden_layers = num_hidden_layers
        self.num_attention_heads = num_attention_heads
        self.num_key_value_heads = num_key_value_heads
        self.hidden_act = hidden_act
        self.layer_norm_eps = layer_norm_eps
        self.rope_theta = rope_theta
        self.resid_pdrop = resid_pdrop
        self.partial_rotary_factor = partial_rotary_factor

        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,
        )


# this is the same as llama except for Phi uses bias=True
def load_attention(config, prefix, weights):
    if config.num_attention_heads != config.num_key_value_heads:
        return _load_gqa(config, prefix, weights)
    else:
        return TensorParallelColumnLinear.load_multi(
            config,
            prefixes=[f"{prefix}.q_proj", f"{prefix}.k_proj", f"{prefix}.v_proj"],
            dim=0,
            weights=weights,
            bias=True,
        )


def _load_gqa(config, prefix: str, weights):
    assert config.hidden_size % config.num_attention_heads == 0
    assert config.num_attention_heads % weights.process_group.size() == 0

    weight = weights.get_multi_weights_col(
        prefixes=[f"{prefix}.q_proj", f"{prefix}.k_proj", f"{prefix}.v_proj"],
        dim=0,
    )

    if config.quantize not in ["gptq", "awq", "marlin"]:
        weight = weight.to(dtype=weights.dtype).to(device=weights.device)

        head_size = config.hidden_size // config.num_attention_heads
        num_heads = config.num_attention_heads // weights.process_group.size()
        num_key_value_heads = config.num_key_value_heads // weights.process_group.size()
        assert list(weight.shape) == [
            (num_heads + 2 * num_key_value_heads) * head_size,
            config.hidden_size,
        ], f"{list(weight.shape)} != {[(num_heads + 2 * config.num_key_value_heads) * head_size, config.hidden_size]}"

    # this is the same as llama except for Phi uses bias=True
    return TensorParallelColumnLinear(get_linear(weight, bias=True))


class FlashPhiAttention(torch.nn.Module):
    def __init__(
        self,
        prefix: str,
        config,
        weights,
    ):
        super().__init__()
        self.num_heads = config.num_attention_heads
        self.hidden_size = config.hidden_size
        self.head_size = self.hidden_size // self.num_heads

        self.softmax_scale = self.head_size**-0.5
        self.rotary_dim = int(config.partial_rotary_factor * self.head_size)

        self.rotary_emb = PositionRotaryEmbedding.static(
            config=config,
            dim=self.rotary_dim,
            base=config.rope_theta,
            device=weights.device,
        )

        if self.num_heads % weights.process_group.size() != 0:
            raise ValueError(
                f"`num_heads` must be divisible by `num_shards` (got `num_heads`: {self.num_heads} "
                f"and `num_shards`: {weights.process_group.size()}"
            )

        self.num_heads = self.num_heads // weights.process_group.size()
        self.num_key_value_heads = (
            config.num_key_value_heads // weights.process_group.size()
        )

        self.query_key_value = load_attention(config, prefix, weights)
        self.kv_scales = get_kv_scales(weights, f"{prefix}")

        # in llama the dense layer is called "o_proj" and has bias=False
        self.dense = TensorParallelRowLinear.load(
            config,
            prefix=f"{prefix}.dense",
            weights=weights,
            bias=True,
        )
        self.num_groups = self.num_heads // self.num_key_value_heads
        self.kv_head_mapping = torch.arange(
            0, self.num_key_value_heads, dtype=torch.int32, device=weights.device
        ).repeat_interleave(self.num_groups)

    def forward(
        self,
        hidden_states,
        cos,
        sin,
        cu_seqlen_prefill,
        kv_cache,
        block_tables,
        slots,
        seqlen,
        max_s,
    ):
        # Compute query, key, value and split
        qkv = self.query_key_value(hidden_states)
        query, kv = qkv.split(
            [
                self.head_size * self.num_heads,
                2 * self.head_size * self.num_key_value_heads,
            ],
            dim=1,
        )

        # Reshape query and key for rotary embeddings
        query = query.view(-1, self.num_heads, self.head_size)
        kv = kv.view(-1, 2, self.num_key_value_heads, self.head_size)

        # NOTE: this is the main difference between Llama and Phi
        # in llama the rotary embeddings are applied to the whole query and key.
        # Phi uses PARTIAL rotary embeddings, which are applied to the first 32 dimensions
        #
        # Apply partial positional embeddings in place
        self.rotary_emb(
            query[:, :, : self.rotary_dim], kv[:, 0, :, : self.rotary_dim], cos, sin
        )

        # Reshape key and value and cache
        kv_cache.store(
            key=kv[:, 0],
            value=kv[:, 1],
            slots=slots,
            kv_scales=self.kv_scales,
        )

        # Prefill
        if cu_seqlen_prefill is not None:
            attn_output = attention(
                query=query,
                key=kv[:, 0],
                value=kv[:, 1],
                kv_scales=self.kv_scales,
                kv_cache=kv_cache,
                seqlen=seqlen,
                block_tables=block_tables,
                softmax_scale=self.softmax_scale,
            )
        # Decode
        else:
            attn_output = paged_attention(
                query,
                kv_cache,
                self.kv_head_mapping,
                self.softmax_scale,
                block_tables,
                seqlen,
                max_s,
                kv_scales=self.kv_scales,
            )

        return self.dense(attn_output.view(-1, self.num_heads * self.head_size))


class PhiMLP(nn.Module):
    def __init__(self, prefix, config, weights):
        super().__init__()
        act = config.hidden_act
        self.act = (
            ACT2FN[act]
            if "gelu" not in act
            else lambda x: torch.nn.functional.gelu(
                x,
                approximate=(
                    "tanh" if act in ["gelu_fast", "gelu_pytorch_tanh"] else "none"
                ),
            )
        )

        # llama weights are up_proj and down_proj and bias=False
        self.up_proj = TensorParallelColumnLinear.load(
            config,
            prefix=f"{prefix}.fc1",
            weights=weights,
            bias=True,
        )
        self.down_proj = TensorParallelRowLinear.load(
            config,
            prefix=f"{prefix}.fc2",
            weights=weights,
            bias=True,
        )

    def forward(self, hidden_states):
        # NOTE: Llama requires the gate up states to an intermediate size
        # Phi does not and we can avoid the `view` operation
        return self.down_proj(self.act(self.up_proj(hidden_states)))


class FlashPhiLayer(nn.Module):
    def __init__(self, prefix: str, layer_id, config, weights):
        super().__init__()
        prefix = f"{prefix}.layers.{layer_id}"
        self.self_attn = FlashPhiAttention(
            prefix=f"{prefix}.self_attn", config=config, weights=weights
        )
        self.mlp = PhiMLP(prefix=f"{prefix}.mlp", config=config, weights=weights)
        self.input_layernorm = FastLayerNorm.load(
            prefix=f"{prefix}.input_layernorm",
            weights=weights,
            eps=config.layer_norm_eps,
        )
        self.resid_dropout = torch.nn.Dropout(config.resid_pdrop)

    def forward(
        self,
        hidden_states,
        residual,
        cos,
        sin,
        cu_seqlen_prefill,
        kv_cache,
        block_tables,
        slots,
        seqlen,
        max_s,
    ):
        hidden_states, res = self.input_layernorm(hidden_states, residual)
        # Self Attention
        attn_output = self.self_attn(
            hidden_states,
            cos,
            sin,
            cu_seqlen_prefill,
            kv_cache,
            block_tables,
            slots,
            seqlen,
            max_s,
        )

        hidden_states = self.resid_dropout(attn_output).add(
            self.resid_dropout(self.mlp(hidden_states))
        )

        return hidden_states, res


class FlashPhiModel(torch.nn.Module):
    def __init__(self, prefix: str, config, weights):
        super().__init__()

        process_group = weights.process_group
        self.tp_rank = process_group.rank()
        self.tp_world_size = process_group.size()
        self.embed_tokens = TensorParallelEmbedding(
            prefix=f"{prefix}.embed_tokens", weights=weights
        )
        self.layers = nn.ModuleList(
            [
                FlashPhiLayer(
                    prefix,
                    layer_id,
                    config,
                    weights,
                )
                for layer_id in range(config.num_hidden_layers)
            ]
        )
        self.gradient_checkpointing = False

        self.head_size = self.layers[0].self_attn.head_size
        self.num_heads = self.layers[0].self_attn.num_heads
        self.num_key_value_heads = self.layers[0].self_attn.num_key_value_heads

        self.norm = FastLayerNorm.load(
            prefix="model.final_layernorm",
            weights=weights,
            eps=config.layer_norm_eps,
        )

    def forward(
        self,
        input_ids: torch.Tensor,
        position_ids: torch.Tensor,
        cu_seqlen_prefill: Optional[torch.Tensor],
        kv_cache: List[Tuple[torch.Tensor, torch.Tensor]],
        block_tables: torch.Tensor,
        slots: torch.Tensor,
        seqlen: Seqlen,
        max_s: int,
    ) -> torch.Tensor:
        hidden_states = self.embed_tokens(input_ids)

        # Get rotary cos and sin for this forward
        # Avoid to index in each layer
        cos, sin = self.layers[0].self_attn.rotary_emb.get_cos_sin(
            position_ids, max_s, hidden_states.dtype
        )

        residual = None
        for i, layer in enumerate(self.layers):
            hidden_states, residual = layer(
                hidden_states,
                residual,
                cos,
                sin,
                cu_seqlen_prefill,
                kv_cache[i],
                block_tables,
                slots,
                seqlen,
                max_s,
            )

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

        return hidden_states


class FlashPhiForCausalLM(torch.nn.Module):
    def __init__(self, prefix: str, config, weights):
        super().__init__()

        if not prefix:
            prefix = "model"
        else:
            prefix = f"{prefix}.model"

        self.model = FlashPhiModel(prefix, config, weights)
        self.lm_head = SpeculativeHead.load(
            config,
            prefix="lm_head",
            weights=weights,
        )

    def forward(
        self,
        input_ids: torch.Tensor,
        position_ids: torch.Tensor,
        cu_seqlen_prefill: Optional[torch.Tensor],
        kv_cache: List[Tuple[torch.Tensor, torch.Tensor]],
        block_tables: torch.Tensor,
        slots: torch.Tensor,
        seqlen: Seqlen,
        max_s: int,
        prefill_cache_indices: Optional[torch.Tensor],
        lm_head_indices: Optional[torch.Tensor] = None,
        adapter_data: Optional[torch.Tensor] = None,
    ) -> torch.Tensor:
        hidden_states = self.model(
            input_ids,
            position_ids,
            cu_seqlen_prefill,
            kv_cache,
            block_tables,
            slots,
            seqlen,
            max_s,
        )
        if lm_head_indices is not None:
            hidden_states = hidden_states[lm_head_indices]

        return self.lm_head(hidden_states)