modeling_mistral4.py

from collections.abc import Callable

import torch
import torch.nn.functional as F
from torch import nn

from transformers.cache_utils import Cache, DynamicCache
from transformers.generation import GenerationMixin
from transformers.modeling_outputs import BaseModelOutputWithPast, CausalLMOutputWithPast
from transformers.activations import ACT2FN
from transformers.modeling_utils import ALL_ATTENTION_FUNCTIONS, PreTrainedModel
from transformers.utils import logging
from transformers.models.deepseek_v3.modeling_deepseek_v3 import (
    DeepseekV3Attention,
    DeepseekV3DecoderLayer,
    apply_rotary_pos_emb_interleave,
)
from transformers.models.llama.modeling_llama import (
    LlamaRMSNorm,
    LlamaRotaryEmbedding,
    apply_rotary_pos_emb,
    eager_attention_forward,
)
from transformers.masking_utils import create_causal_mask
from transformers.models.gemma.modeling_gemma import GemmaMLP
from .configuration_mistral4 import Mistral4Config


logger = logging.get_logger(__name__)


def get_llama_4_attn_scale(positions_ids: torch.Tensor, beta: float, max_position_embeddings: int) -> torch.Tensor:
    scaling = 1 + beta * torch.log(1 + torch.floor(positions_ids / max_position_embeddings))
    return scaling[:, None, :, None]

class Mistral4RMSNorm(LlamaRMSNorm):
    pass


class Mistral4RotaryEmbedding(LlamaRotaryEmbedding):
    pass


class Mistral4MLP(GemmaMLP):
    def __init__(self, config, intermediate_size=None):
        super().__init__(config)
        self.config = config
        self.hidden_size = config.hidden_size
        self.intermediate_size = config.intermediate_size if intermediate_size is None else intermediate_size
        self.gate_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=False)
        self.up_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=False)
        self.down_proj = nn.Linear(self.intermediate_size, self.hidden_size, bias=False)
        self.act_fn = ACT2FN[config.hidden_act]


class Mistral4TopkRouter(nn.Module):
    def __init__(self, config):
        super().__init__()
        self.config = config
        self.n_routed_experts = config.n_routed_experts

        self.weight = nn.Parameter(torch.empty((self.n_routed_experts, config.hidden_size)))

    def forward(self, hidden_states):
        hidden_states = hidden_states.view(-1, self.config.hidden_size)
        router_logits = F.linear(hidden_states, self.weight)
        return router_logits


class Mistral4NaiveMoe(nn.Module):
    """Collection of expert weights stored as 3D tensors."""

    def __init__(self, config):
        super().__init__()
        self.num_experts = config.num_local_experts
        self.hidden_dim = config.hidden_size
        self.intermediate_dim = config.moe_intermediate_size
        self.gate_up_proj = nn.Parameter(torch.empty(self.num_experts, 2 * self.intermediate_dim, self.hidden_dim))
        self.down_proj = nn.Parameter(torch.empty(self.num_experts, self.hidden_dim, self.intermediate_dim))
        self.act_fn = ACT2FN[config.hidden_act]

    def forward(
        self,
        hidden_states: torch.Tensor,
        top_k_index: torch.Tensor,
        top_k_weights: torch.Tensor,
    ) -> torch.Tensor:
        final_hidden_states = torch.zeros_like(hidden_states)
        with torch.no_grad():
            expert_mask = torch.nn.functional.one_hot(top_k_index, num_classes=self.num_experts)
            expert_mask = expert_mask.permute(2, 1, 0)
            expert_hit = torch.greater(expert_mask.sum(dim=(-1, -2)), 0).nonzero()

        for expert_idx in expert_hit:
            expert_idx = expert_idx[0]
            if expert_idx == self.num_experts:
                continue
            top_k_pos, token_idx = torch.where(expert_mask[expert_idx])
            current_state = hidden_states[token_idx]
            gate, up = nn.functional.linear(current_state, self.gate_up_proj[expert_idx]).chunk(2, dim=-1)
            current_hidden_states = self.act_fn(gate) * up
            current_hidden_states = nn.functional.linear(current_hidden_states, self.down_proj[expert_idx])
            current_hidden_states = current_hidden_states * top_k_weights[token_idx, top_k_pos, None]
            final_hidden_states.index_add_(0, token_idx, current_hidden_states.to(final_hidden_states.dtype))

        return final_hidden_states


class Mistral4MoE(nn.Module):
    """
    A mixed expert module containing shared experts.
    """

    def __init__(self, config):
        super().__init__()
        self.config = config
        self.experts = Mistral4NaiveMoe(config)
        self.gate = Mistral4TopkRouter(config)
        if config.n_shared_experts > 0:
            self.shared_experts = Mistral4MLP(
                config=config, intermediate_size=config.moe_intermediate_size * config.n_shared_experts
            )
        else:
            self.shared_experts = None
        self.n_routed_experts = config.n_routed_experts
        self.n_group = config.n_group
        self.topk_group = config.topk_group
        self.norm_topk_prob = config.norm_topk_prob
        self.routed_scaling_factor = config.routed_scaling_factor
        self.top_k = config.num_experts_per_tok

    def route_tokens_to_experts(self, router_logits: torch.Tensor) -> tuple[torch.Tensor, torch.Tensor]:
        router_logits = router_logits.softmax(-1)
        group_scores = (
            router_logits.view(-1, self.n_group, self.n_routed_experts // self.n_group).topk(2, dim=-1)[0].sum(dim=-1)
        )
        group_idx = torch.topk(group_scores, k=self.topk_group, dim=-1, sorted=False)[1]
        group_mask = torch.zeros_like(group_scores)
        group_mask.scatter_(1, group_idx, 1)
        score_mask = (
            group_mask.unsqueeze(-1)
            .expand(-1, self.n_group, self.n_routed_experts // self.n_group)
            .reshape(-1, self.n_routed_experts)
        )
        scores_for_choice = router_logits.masked_fill(~score_mask.bool(), 0.0)
        topk_indices = torch.topk(scores_for_choice, k=self.top_k, dim=-1, sorted=False)[1]
        topk_weights = router_logits.gather(1, topk_indices)
        if self.norm_topk_prob:
            denominator = topk_weights.sum(dim=-1, keepdim=True) + 1e-20
            topk_weights /= denominator
        topk_weights = topk_weights * self.routed_scaling_factor
        return topk_indices, topk_weights

    def forward(self, hidden_states):
        residuals = hidden_states
        orig_shape = hidden_states.shape
        router_logits = self.gate(hidden_states)
        topk_indices, topk_weights = self.route_tokens_to_experts(router_logits)
        hidden_states = hidden_states.view(-1, hidden_states.shape[-1])
        hidden_states = self.experts(hidden_states, topk_indices, topk_weights).view(*orig_shape)
        if self.shared_experts is not None:
            hidden_states = hidden_states + self.shared_experts(residuals)
        return hidden_states


class Mistral4Attention(DeepseekV3Attention):
    def __init__(self, config: Mistral4Config, layer_idx: int):
        nn.Module.__init__(self)
        self.config = config
        self.layer_idx = layer_idx
        self.num_key_value_groups = config.num_attention_heads // config.num_key_value_heads
        self.attention_dropout = config.attention_dropout
        self.num_heads = config.num_attention_heads

        self.q_lora_rank = config.q_lora_rank
        self.qk_rope_head_dim = config.qk_rope_head_dim
        self.kv_lora_rank = config.kv_lora_rank
        self.v_head_dim = config.v_head_dim
        self.qk_nope_head_dim = config.qk_nope_head_dim
        self.qk_head_dim = config.qk_head_dim

        self.is_causal = True
        if self.q_lora_rank is None:
            self.q_proj = nn.Linear(config.hidden_size, self.num_heads * self.qk_head_dim, bias=False)
        else:
            self.q_a_proj = nn.Linear(config.hidden_size, config.q_lora_rank, bias=config.attention_bias)
            self.q_a_layernorm = Mistral4RMSNorm(config.q_lora_rank)
            self.q_b_proj = nn.Linear(config.q_lora_rank, self.num_heads * self.qk_head_dim, bias=False)

        self.kv_a_proj_with_mqa = nn.Linear(
            config.hidden_size,
            self.kv_lora_rank + self.qk_rope_head_dim,
            bias=config.attention_bias,
        )
        self.kv_a_layernorm = Mistral4RMSNorm(self.kv_lora_rank)
        self.kv_b_proj = nn.Linear(
            self.kv_lora_rank,
            self.num_heads * (self.qk_nope_head_dim + self.v_head_dim),
            bias=False,
        )

        self.o_proj = nn.Linear(
            self.num_heads * self.v_head_dim,
            config.hidden_size,
            bias=config.attention_bias,
        )

        self.scaling = self.qk_head_dim ** (-0.5)

    def forward(
        self,
        hidden_states: torch.Tensor,
        position_embeddings: tuple[torch.Tensor, torch.Tensor],
        attention_mask: torch.Tensor | None,
        position_ids: torch.Tensor,
        past_key_values: Cache | None = None,
        **kwargs,
    ) -> tuple[torch.Tensor, torch.Tensor | None, tuple[torch.Tensor] | None]:
        batch_size, seq_length = hidden_states.shape[:-1]
        query_shape = (batch_size, seq_length, -1, self.qk_head_dim)
        key_shape = (batch_size, seq_length, -1, self.qk_nope_head_dim + self.v_head_dim)

        if self.q_lora_rank is None:
            q_states = self.q_proj(hidden_states)
        else:
            q_states = self.q_b_proj(self.q_a_layernorm(self.q_a_proj(hidden_states)))
        q_states = q_states.view(query_shape).transpose(1, 2)
        q_pass, q_rot = torch.split(q_states, [self.qk_nope_head_dim, self.qk_rope_head_dim], dim=-1)

        compressed_kv = self.kv_a_proj_with_mqa(hidden_states)
        k_pass, k_rot = torch.split(compressed_kv, [self.kv_lora_rank, self.qk_rope_head_dim], dim=-1)

        k_pass = self.kv_b_proj(self.kv_a_layernorm(k_pass)).view(key_shape).transpose(1, 2)
        k_pass, value_states = torch.split(k_pass, [self.qk_nope_head_dim, self.v_head_dim], dim=-1)

        k_rot = k_rot.view(batch_size, 1, seq_length, self.qk_rope_head_dim)

        cos, sin = position_embeddings
        if self.config.rope_interleave:  # support using interleaved weights for efficiency
            q_rot, k_rot = apply_rotary_pos_emb_interleave(q_rot, k_rot, cos, sin)
        else:
            q_rot, k_rot = apply_rotary_pos_emb(q_rot, k_rot, cos, sin)
        k_rot = k_rot.expand(*k_pass.shape[:-1], -1)

        query_states = torch.cat((q_pass, q_rot), dim=-1)
        key_states = torch.cat((k_pass, k_rot), dim=-1)

        query_states = query_states * get_llama_4_attn_scale(
            position_ids,
            self.config.rope_parameters.get("llama_4_scaling_beta"),
            self.config.rope_parameters.get("original_max_position_embeddings"),
        ).to(query_states.dtype)

        if past_key_values is not None:
            key_states, value_states = past_key_values.update(key_states, value_states, self.layer_idx)

        attention_interface: Callable = ALL_ATTENTION_FUNCTIONS.get_interface(
            self.config._attn_implementation, eager_attention_forward
        )

        attn_output, attn_weights = attention_interface(
            self,
            query_states,
            key_states,
            value_states,
            attention_mask,
            dropout=0.0 if not self.training else self.attention_dropout,
            scaling=self.scaling,
            **kwargs,
        )

        attn_output = attn_output.reshape(batch_size, seq_length, -1).contiguous()
        attn_output = self.o_proj(attn_output)
        return attn_output, attn_weights


class Mistral4DecoderLayer(DeepseekV3DecoderLayer):
    def __init__(self, config: Mistral4Config, layer_idx: int):
        nn.Module.__init__(self)
        self.hidden_size = config.hidden_size

        self.self_attn = Mistral4Attention(config=config, layer_idx=layer_idx)

        if layer_idx >= config.first_k_dense_replace:
            self.mlp = Mistral4MoE(config)
        else:
            self.mlp = Mistral4MLP(config)

        self.input_layernorm = Mistral4RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
        self.post_attention_layernorm = Mistral4RMSNorm(config.hidden_size, eps=config.rms_norm_eps)


class Mistral4PreTrainedModel(PreTrainedModel):
    config: Mistral4Config
    base_model_prefix = "model"
    supports_gradient_checkpointing = True
    _no_split_modules = ["Mistral4DecoderLayer"]
    _skip_keys_device_placement = ["past_key_values"]
    _supports_flash_attn = True
    _supports_sdpa = True
    _supports_flex_attn = True

    _can_compile_fullgraph = True
    _supports_attention_backend = True
    _can_record_outputs = {
        "hidden_states": Mistral4DecoderLayer,
        "attentions": Mistral4Attention,
    }
    _keep_in_fp32_modules_strict = []
    _keys_to_ignore_on_load_unexpected = []

    @torch.no_grad()
    def _init_weights(self, module):
        super()._init_weights(module)
        if isinstance(module, Mistral4TopkRouter):
            module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
        elif isinstance(module, Mistral4NaiveMoe):
            module.gate_up_proj.data.normal_(mean=0.0, std=self.config.initializer_range)
            module.down_proj.normal_(mean=0.0, std=self.config.initializer_range)


class Mistral4Model(Mistral4PreTrainedModel):
    def __init__(self, config: Mistral4Config):
        super().__init__(config)
        self.padding_idx = config.pad_token_id
        self.vocab_size = config.vocab_size

        self.embed_tokens = nn.Embedding(config.vocab_size, config.hidden_size, self.padding_idx)
        self.layers = nn.ModuleList(
            [Mistral4DecoderLayer(config, layer_idx) for layer_idx in range(config.num_hidden_layers)]
        )
        self.norm = Mistral4RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
        self.rotary_emb = Mistral4RotaryEmbedding(config=config)
        self.gradient_checkpointing = False

        # Initialize weights and apply final processing
        self.post_init()

    def forward(
        self,
        input_ids: torch.LongTensor | None = None,
        attention_mask: torch.Tensor | None = None,
        position_ids: torch.LongTensor | None = None,
        past_key_values: Cache | None = None,
        inputs_embeds: torch.FloatTensor | None = None,
        use_cache: bool | None = None,
        **kwargs,
    ) :
        if (input_ids is None) ^ (inputs_embeds is not None):
            raise ValueError("You must specify exactly one of input_ids or inputs_embeds")

        if inputs_embeds is None:
            inputs_embeds: torch.Tensor = self.embed_tokens(input_ids)

        if use_cache and past_key_values is None:
            past_key_values = DynamicCache(config=self.config)

        if position_ids is None:
            past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0
            position_ids = torch.arange(inputs_embeds.shape[1], device=inputs_embeds.device) + past_seen_tokens
            position_ids = position_ids.unsqueeze(0)

        causal_mask = create_causal_mask(
            config=self.config,
            inputs_embeds=inputs_embeds,
            attention_mask=attention_mask,
            past_key_values=past_key_values,
            position_ids=position_ids,
        )

        hidden_states = inputs_embeds
        position_embeddings = self.rotary_emb(hidden_states, position_ids=position_ids)

        for decoder_layer in self.layers[: self.config.num_hidden_layers]:
            hidden_states = decoder_layer(
                hidden_states,
                attention_mask=causal_mask,
                position_embeddings=position_embeddings,
                position_ids=position_ids,
                past_key_values=past_key_values,
                use_cache=use_cache,
                **kwargs,
            )

        hidden_states = self.norm(hidden_states)
        return BaseModelOutputWithPast(
            last_hidden_state=hidden_states,
            past_key_values=past_key_values,
        )


class Mistral4ForCausalLM(Mistral4PreTrainedModel, GenerationMixin):
    _tied_weights_keys = {"lm_head.weight": "model.embed_tokens.weight"}
    _tp_plan = {"lm_head": "colwise_gather_output"}
    _pp_plan = {"lm_head": (["hidden_states"], ["logits"])}

    def __init__(self, config):
        super().__init__(config)
        self.model = Mistral4Model(config)
        self.vocab_size = config.vocab_size
        self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)

        # Initialize weights and apply final processing
        self.post_init()

    def forward(
        self,
        input_ids: torch.LongTensor | None = None,
        attention_mask: torch.Tensor | None = None,
        position_ids: torch.LongTensor | None = None,
        past_key_values: Cache | None = None,
        inputs_embeds: torch.FloatTensor | None = None,
        labels: torch.LongTensor | None = None,
        use_cache: bool | None = None,
        logits_to_keep: int | torch.Tensor = 0,
        **kwargs,
    ) -> CausalLMOutputWithPast:
        r"""
        Example:

        ```python
        >>> from transformers import AutoTokenizer, Mistral4ForCausalLM

        >>> model = Mistral4ForCausalLM.from_pretrained("meta-mistral4/Mistral4-2-7b-hf")
        >>> tokenizer = AutoTokenizer.from_pretrained("meta-mistral4/Mistral4-2-7b-hf")

        >>> prompt = "Hey, are you conscious? Can you talk to me?"
        >>> inputs = tokenizer(prompt, return_tensors="pt")

        >>> # Generate
        >>> generate_ids = model.generate(inputs.input_ids, max_length=30)
        >>> tokenizer.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]
        "Hey, are you conscious? Can you talk to me?\nI'm not conscious, but I can talk to you."
        ```"""
        outputs: BaseModelOutputWithPast = self.model(
            input_ids=input_ids,
            attention_mask=attention_mask,
            position_ids=position_ids,
            past_key_values=past_key_values,
            inputs_embeds=inputs_embeds,
            use_cache=use_cache,
            **kwargs,
        )

        hidden_states = outputs.last_hidden_state
        # Only compute necessary logits, and do not upcast them to float if we are not computing the loss
        slice_indices = slice(-logits_to_keep, None) if isinstance(logits_to_keep, int) else logits_to_keep
        logits = self.lm_head(hidden_states[:, slice_indices, :])

        loss = None
        if labels is not None:
            loss = self.loss_function(logits=logits, labels=labels, vocab_size=self.config.vocab_size, **kwargs)

        return CausalLMOutputWithPast(
            loss=loss,
            logits=logits,
            past_key_values=outputs.past_key_values,
            hidden_states=outputs.hidden_states,
            attentions=outputs.attentions,
        )




__all__ = [
    "Mistral4PreTrainedModel",
    "Mistral4Model",
    "Mistral4ForCausalLM",
]