diff --git a/src/llmcompressor/transformers/tracing/__init__.py b/src/llmcompressor/transformers/tracing/__init__.py index 93de3d08ab..c3874392aa 100644 --- a/src/llmcompressor/transformers/tracing/__init__.py +++ b/src/llmcompressor/transformers/tracing/__init__.py @@ -13,6 +13,13 @@ from .qwen2_5_vl import ( Qwen2_5_VLForConditionalGeneration as TraceableQwen2_5_VLForConditionalGeneration ) +from .mistral import ( + MistralForCausalLM as TraceableMistralForCausalLM +) +from .mistral3 import ( + Mistral3ForConditionalGeneration as TraceableMistral3ForConditionalGeneration +) + from .debug import get_model_class __all__ = [ @@ -22,4 +29,6 @@ "TraceableQwen2VLForConditionalGeneration", "TraceableIdefics3ForConditionalGeneration", "TraceableQwen2_5_VLForConditionalGeneration", + "TraceableMistralForCausalLM", + "TraceableMistral3ForConditionalGeneration", ] diff --git a/src/llmcompressor/transformers/tracing/mistral.py b/src/llmcompressor/transformers/tracing/mistral.py new file mode 100644 index 0000000000..cac2bb2f32 --- /dev/null +++ b/src/llmcompressor/transformers/tracing/mistral.py @@ -0,0 +1,1104 @@ +# 🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨 +# This file was automatically generated from src/transformers/models/mistral/modular_mistral.py. +# Do NOT edit this file manually as any edits will be overwritten by the generation of +# the file from the modular. If any change should be done, please apply the change to the +# modular_mistral.py file directly. One of our CI enforces this. +# 🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨 +from typing import Callable, List, Optional, Tuple, Union + +import torch +from torch import nn + +from transformers.activations import ACT2FN +from transformers.cache_utils import Cache, DynamicCache, SlidingWindowCache, StaticCache +from transformers.generation import GenerationMixin +from transformers.modeling_attn_mask_utils import AttentionMaskConverter +from transformers.modeling_flash_attention_utils import FlashAttentionKwargs +from transformers.modeling_outputs import ( + BaseModelOutputWithPast, + CausalLMOutputWithPast, + QuestionAnsweringModelOutput, + SequenceClassifierOutputWithPast, + TokenClassifierOutput, +) +from transformers.modeling_rope_utils import ROPE_INIT_FUNCTIONS, dynamic_rope_update +from transformers.modeling_utils import ALL_ATTENTION_FUNCTIONS, PreTrainedModel +from transformers.processing_utils import Unpack +from transformers.utils import ( + LossKwargs, + add_code_sample_docstrings, + add_start_docstrings, + add_start_docstrings_to_model_forward, + can_return_tuple, + logging, + replace_return_docstrings, +) +from transformers.utils.deprecation import deprecate_kwarg +from transformers.models.mistral.configuration_mistral import MistralConfig + + +logger = logging.get_logger(__name__) + +_CHECKPOINT_FOR_DOC = "mistralai/Mistral-7B-v0.1" +_CONFIG_FOR_DOC = "MistralConfig" + + +class MistralMLP(nn.Module): + def __init__(self, config): + super().__init__() + self.config = config + self.hidden_size = config.hidden_size + self.intermediate_size = config.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] + + def forward(self, x): + down_proj = self.down_proj(self.act_fn(self.gate_proj(x)) * self.up_proj(x)) + return down_proj + + +def rotate_half(x): + """Rotates half the hidden dims of the input.""" + x1 = x[..., : x.shape[-1] // 2] + x2 = x[..., x.shape[-1] // 2 :] + return torch.cat((-x2, x1), dim=-1) + + +def apply_rotary_pos_emb(q, k, cos, sin, position_ids=None, unsqueeze_dim=1): + """Applies Rotary Position Embedding to the query and key tensors. + + Args: + q (`torch.Tensor`): The query tensor. + k (`torch.Tensor`): The key tensor. + cos (`torch.Tensor`): The cosine part of the rotary embedding. + sin (`torch.Tensor`): The sine part of the rotary embedding. + position_ids (`torch.Tensor`, *optional*): + Deprecated and unused. + unsqueeze_dim (`int`, *optional*, defaults to 1): + The 'unsqueeze_dim' argument specifies the dimension along which to unsqueeze cos[position_ids] and + sin[position_ids] so that they can be properly broadcasted to the dimensions of q and k. For example, note + that cos[position_ids] and sin[position_ids] have the shape [batch_size, seq_len, head_dim]. Then, if q and + k have the shape [batch_size, heads, seq_len, head_dim], then setting unsqueeze_dim=1 makes + cos[position_ids] and sin[position_ids] broadcastable to the shapes of q and k. Similarly, if q and k have + the shape [batch_size, seq_len, heads, head_dim], then set unsqueeze_dim=2. + Returns: + `tuple(torch.Tensor)` comprising of the query and key tensors rotated using the Rotary Position Embedding. + """ + cos = cos.unsqueeze(unsqueeze_dim) + sin = sin.unsqueeze(unsqueeze_dim) + q_embed = (q * cos) + (rotate_half(q) * sin) + k_embed = (k * cos) + (rotate_half(k) * sin) + return q_embed, k_embed + + +def repeat_kv(hidden_states: torch.Tensor, n_rep: int) -> torch.Tensor: + """ + This is the equivalent of torch.repeat_interleave(x, dim=1, repeats=n_rep). The hidden states go from (batch, + num_key_value_heads, seqlen, head_dim) to (batch, num_attention_heads, seqlen, head_dim) + """ + batch, num_key_value_heads, slen, head_dim = hidden_states.shape + if n_rep == 1: + return hidden_states + hidden_states = hidden_states[:, :, None, :, :].expand(batch, num_key_value_heads, n_rep, slen, head_dim) + return hidden_states.reshape(batch, num_key_value_heads * n_rep, slen, head_dim) + + +def eager_attention_forward( + module: nn.Module, + query: torch.Tensor, + key: torch.Tensor, + value: torch.Tensor, + attention_mask: Optional[torch.Tensor], + scaling: float, + dropout: float = 0.0, + **kwargs, +): + key_states = repeat_kv(key, module.num_key_value_groups) + value_states = repeat_kv(value, module.num_key_value_groups) + + attn_weights = torch.matmul(query, key_states.transpose(2, 3)) * scaling + if attention_mask is not None: + causal_mask = attention_mask[:, :, :, : key_states.shape[-2]] + attn_weights = attn_weights + causal_mask + + attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype=torch.float32).to(query.dtype) + attn_weights = nn.functional.dropout(attn_weights, p=dropout, training=module.training) + attn_output = torch.matmul(attn_weights, value_states) + attn_output = attn_output.transpose(1, 2).contiguous() + + return attn_output, attn_weights + + +class MistralAttention(nn.Module): + """Multi-headed attention from 'Attention Is All You Need' paper""" + + def __init__(self, config: MistralConfig, layer_idx: int): + super().__init__() + self.config = config + self.layer_idx = layer_idx + self.head_dim = getattr(config, "head_dim", config.hidden_size // config.num_attention_heads) + self.num_key_value_groups = config.num_attention_heads // config.num_key_value_heads + self.scaling = self.head_dim**-0.5 + self.attention_dropout = config.attention_dropout + self.is_causal = True + self.q_proj = nn.Linear(config.hidden_size, config.num_attention_heads * self.head_dim, bias=False) + self.k_proj = nn.Linear(config.hidden_size, config.num_key_value_heads * self.head_dim, bias=False) + self.v_proj = nn.Linear(config.hidden_size, config.num_key_value_heads * self.head_dim, bias=False) + self.o_proj = nn.Linear(config.num_attention_heads * self.head_dim, config.hidden_size, bias=False) + + def forward( + self, + hidden_states: torch.Tensor, + position_embeddings: Tuple[torch.Tensor, torch.Tensor], + attention_mask: Optional[torch.Tensor], + past_key_value: Optional[Cache] = None, + cache_position: Optional[torch.LongTensor] = None, + **kwargs: Unpack[FlashAttentionKwargs], + ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]: + input_shape = hidden_states.shape[:-1] + hidden_shape = (*input_shape, -1, self.head_dim) + + query_states = self.q_proj(hidden_states).view(hidden_shape).transpose(1, 2) + key_states = self.k_proj(hidden_states).view(hidden_shape).transpose(1, 2) + value_states = self.v_proj(hidden_states).view(hidden_shape).transpose(1, 2) + + cos, sin = position_embeddings + query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin) + + if past_key_value is not None: + # sin and cos are specific to RoPE models; cache_position needed for the static cache + cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position} + key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs) + + attention_interface: Callable = eager_attention_forward + if self.config._attn_implementation != "eager": + if self.config._attn_implementation == "sdpa" and kwargs.get("output_attentions", False): + logger.warning_once( + "`torch.nn.functional.scaled_dot_product_attention` does not support `output_attentions=True`. Falling back to " + 'eager attention. This warning can be removed using the argument `attn_implementation="eager"` when loading the model.' + ) + else: + attention_interface = ALL_ATTENTION_FUNCTIONS[self.config._attn_implementation] + + 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, + sliding_window=getattr(self.config, "sliding_window", None), # main diff with Llama + **kwargs, + ) + + attn_output = attn_output.reshape(*input_shape, -1).contiguous() + attn_output = self.o_proj(attn_output) + return attn_output, attn_weights + + +class MistralRMSNorm(nn.Module): + def __init__(self, hidden_size, eps=1e-6): + """ + MistralRMSNorm is equivalent to T5LayerNorm + """ + super().__init__() + self.weight = nn.Parameter(torch.ones(hidden_size)) + self.variance_epsilon = eps + + def forward(self, hidden_states): + input_dtype = hidden_states.dtype + hidden_states = hidden_states.to(torch.float32) + variance = hidden_states.pow(2).mean(-1, keepdim=True) + hidden_states = hidden_states * torch.rsqrt(variance + self.variance_epsilon) + return self.weight * hidden_states.to(input_dtype) + + def extra_repr(self): + return f"{tuple(self.weight.shape)}, eps={self.variance_epsilon}" + + +class MistralDecoderLayer(nn.Module): + def __init__(self, config: MistralConfig, layer_idx: int): + super().__init__() + self.hidden_size = config.hidden_size + self.self_attn = MistralAttention(config=config, layer_idx=layer_idx) + self.mlp = MistralMLP(config) + self.input_layernorm = MistralRMSNorm(config.hidden_size, eps=config.rms_norm_eps) + self.post_attention_layernorm = MistralRMSNorm(config.hidden_size, eps=config.rms_norm_eps) + + def forward( + self, + hidden_states: torch.Tensor, + attention_mask: Optional[torch.Tensor] = None, + position_ids: Optional[torch.LongTensor] = None, + past_key_value: Optional[Cache] = None, + output_attentions: Optional[bool] = False, + use_cache: Optional[bool] = False, + cache_position: Optional[torch.LongTensor] = None, + position_embeddings: Optional[Tuple[torch.Tensor, torch.Tensor]] = None, # necessary, but kept here for BC + **kwargs: Unpack[FlashAttentionKwargs], + ) -> Tuple[torch.FloatTensor, Optional[Tuple[torch.FloatTensor, torch.FloatTensor]]]: + residual = hidden_states + + hidden_states = self.input_layernorm(hidden_states) + + # Self Attention + hidden_states, self_attn_weights = self.self_attn( + hidden_states=hidden_states, + attention_mask=attention_mask, + position_ids=position_ids, + past_key_value=past_key_value, + output_attentions=output_attentions, + use_cache=use_cache, + cache_position=cache_position, + position_embeddings=position_embeddings, + **kwargs, + ) + hidden_states = residual + hidden_states + + # Fully Connected + residual = hidden_states + hidden_states = self.post_attention_layernorm(hidden_states) + hidden_states = self.mlp(hidden_states) + hidden_states = residual + hidden_states + + outputs = (hidden_states,) + if output_attentions: + outputs += (self_attn_weights,) + + return outputs + + +class MistralRotaryEmbedding(nn.Module): + def __init__(self, config: MistralConfig, device=None): + super().__init__() + # BC: "rope_type" was originally "type" + if hasattr(config, "rope_scaling") and config.rope_scaling is not None: + self.rope_type = config.rope_scaling.get("rope_type", config.rope_scaling.get("type")) + else: + self.rope_type = "default" + self.max_seq_len_cached = config.max_position_embeddings + self.original_max_seq_len = config.max_position_embeddings + + self.config = config + self.rope_init_fn = ROPE_INIT_FUNCTIONS[self.rope_type] + + inv_freq, self.attention_scaling = self.rope_init_fn(self.config, device) + self.register_buffer("inv_freq", inv_freq, persistent=False) + self.original_inv_freq = self.inv_freq + + @torch.no_grad() + @dynamic_rope_update # power user: used with advanced RoPE types (e.g. dynamic rope) + def forward(self, x, position_ids): + inv_freq_expanded = self.inv_freq[None, :, None].float().expand(position_ids.shape[0], -1, 1).to(x.device) + position_ids_expanded = position_ids[:, None, :].float() + + device_type = x.device.type if isinstance(x.device.type, str) and x.device.type != "mps" else "cpu" + with torch.autocast(device_type=device_type, enabled=False): # Force float32 + freqs = (inv_freq_expanded.float() @ position_ids_expanded.float()).transpose(1, 2) + emb = torch.cat((freqs, freqs), dim=-1) + cos = emb.cos() * self.attention_scaling + sin = emb.sin() * self.attention_scaling + + return cos.to(dtype=x.dtype), sin.to(dtype=x.dtype) + + +MISTRAL_START_DOCSTRING = r""" + This model inherits from [`PreTrainedModel`]. Check the superclass documentation for the generic methods the + library implements for all its model (such as downloading or saving, resizing the input embeddings, pruning heads + etc.) + + This model is also a PyTorch [torch.nn.Module](https://pytorch.org/docs/stable/nn.html#torch.nn.Module) subclass. + Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage + and behavior. + + Parameters: + config ([`MistralConfig`]): + Model configuration class with all the parameters of the model. Initializing with a config file does not + load the weights associated with the model, only the configuration. Check out the + [`~PreTrainedModel.from_pretrained`] method to load the model weights. +""" + + +@add_start_docstrings( + "The bare Mistral Model outputting raw hidden-states without any specific head on top.", + MISTRAL_START_DOCSTRING, +) +class MistralPreTrainedModel(PreTrainedModel): + config_class = MistralConfig + base_model_prefix = "model" + supports_gradient_checkpointing = True + _no_split_modules = ["MistralDecoderLayer"] + _skip_keys_device_placement = ["past_key_values"] + _supports_flash_attn_2 = True + _supports_sdpa = True + _supports_flex_attn = True + _supports_cache_class = True + _supports_quantized_cache = True + _supports_static_cache = True + _supports_attention_backend = True + + def _init_weights(self, module): + std = self.config.initializer_range + if isinstance(module, nn.Linear): + module.weight.data.normal_(mean=0.0, std=std) + if module.bias is not None: + module.bias.data.zero_() + elif isinstance(module, nn.Embedding): + module.weight.data.normal_(mean=0.0, std=std) + if module.padding_idx is not None: + module.weight.data[module.padding_idx].zero_() + + +MISTRAL_INPUTS_DOCSTRING = r""" + Args: + input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`): + Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you provide + it. + + Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and + [`PreTrainedTokenizer.__call__`] for details. + + [What are input IDs?](../glossary#input-ids) + attention_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*): + Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`: + + - 1 for tokens that are **not masked**, + - 0 for tokens that are **masked**. + + [What are attention masks?](../glossary#attention-mask) + + Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and + [`PreTrainedTokenizer.__call__`] for details. + + If `past_key_values` is used, optionally only the last `input_ids` have to be input (see + `past_key_values`). + + If you want to change padding behavior, you should read [`modeling_opt._prepare_decoder_attention_mask`] + and modify to your needs. See diagram 1 in [the paper](https://arxiv.org/abs/1910.13461) for more + information on the default strategy. + + - 1 indicates the head is **not masked**, + - 0 indicates the head is **masked**. + position_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*): + Indices of positions of each input sequence tokens in the position embeddings. Selected in the range `[0, + config.n_positions - 1]`. + + [What are position IDs?](../glossary#position-ids) + past_key_values (`Cache`, *optional*): + Pre-computed hidden-states (key and values in the self-attention blocks and in the cross-attention + blocks) that can be used to speed up sequential decoding. This typically consists in the `past_key_values` + returned by the model at a previous stage of decoding, when `use_cache=True` or `config.use_cache=True`. + + It is a [`~cache_utils.Cache`] instance. For more details, see our [kv cache guide](https://huggingface.co/docs/transformers/en/kv_cache). + + If `past_key_values` are used, the user can optionally input only the last `input_ids` (those that don't + have their past key value states given to this model) of shape `(batch_size, 1)` instead of all `input_ids` + of shape `(batch_size, sequence_length)`. + inputs_embeds (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*): + Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation. This + is useful if you want more control over how to convert `input_ids` indices into associated vectors than the + model's internal embedding lookup matrix. + use_cache (`bool`, *optional*): + If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding (see + `past_key_values`). + output_attentions (`bool`, *optional*): + Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned + tensors for more detail. + output_hidden_states (`bool`, *optional*): + Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for + more detail. + return_dict (`bool`, *optional*): + Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple. + cache_position (`torch.LongTensor` of shape `(sequence_length)`, *optional*): + Indices depicting the position of the input sequence tokens in the sequence. Contrarily to `position_ids`, + this tensor is not affected by padding. It is used to update the cache in the correct position and to infer + the complete sequence length. +""" + +# +# TRACING: move _prepare_4d_causal_attention_mask_with_cache_position out of MistralModel class +# and wrap it to avoid tracing +# +@torch.fx.wrap +def _prepare_4d_causal_attention_mask_with_cache_position( + attention_mask: torch.Tensor, + sequence_length: int, + target_length: int, + dtype: torch.dtype, + device: torch.device, + cache_position: torch.Tensor, + batch_size: int, + config: MistralConfig, + past_key_values: Cache, +): + """ + Creates a causal 4D mask of shape `(batch_size, 1, query_length, key_value_length)` from a 2D mask of shape + `(batch_size, key_value_length)`, or if the input `attention_mask` is already 4D, do nothing. + + Args: + attention_mask (`torch.Tensor`): + A 2D attention mask of shape `(batch_size, key_value_length)` or a 4D attention mask of shape `(batch_size, 1, query_length, key_value_length)`. + sequence_length (`int`): + The sequence length being processed. + target_length (`int`): + The target length: when generating with static cache, the mask should be as long as the static cache, to account for the 0 padding, the part of the cache that is not filled yet. + dtype (`torch.dtype`): + The dtype to use for the 4D attention mask. + device (`torch.device`): + The device to place the 4D attention mask on. + cache_position (`torch.Tensor`): + Indices depicting the position of the input sequence tokens in the sequence. + batch_size (`torch.Tensor`): + Batch size. + config (`MistralConfig`): + The model's configuration class + past_key_values (`Cache`): + The cache class that is being used currently to generate + """ + if attention_mask is not None and attention_mask.dim() == 4: + # In this case we assume that the mask comes already in inverted form and requires no inversion or slicing. + causal_mask = attention_mask + else: + min_dtype = torch.finfo(dtype).min + causal_mask = torch.full( + (sequence_length, target_length), fill_value=min_dtype, dtype=dtype, device=device + ) + diagonal_attend_mask = torch.arange(target_length, device=device) > cache_position.reshape(-1, 1) + if config.sliding_window is not None: + # if we have sliding window, we should not attend to tokens beyond sliding window length, so we mask them out also + # the check is needed to verify is current checkpoint was trained with sliding window or not + if not isinstance(past_key_values, SlidingWindowCache) or sequence_length > target_length: + sliding_attend_mask = torch.arange(target_length, device=device) <= ( + cache_position.reshape(-1, 1) - config.sliding_window + ) + diagonal_attend_mask.bitwise_or_(sliding_attend_mask) + causal_mask *= diagonal_attend_mask + causal_mask = causal_mask[None, None, :, :].expand(batch_size, 1, -1, -1) + if attention_mask is not None: + causal_mask = causal_mask.clone() # copy to contiguous memory for in-place edit + if attention_mask.shape[-1] > target_length: + attention_mask = attention_mask[:, :target_length] + mask_length = attention_mask.shape[-1] + padding_mask = causal_mask[:, :, :, :mask_length] + attention_mask[:, None, None, :].to( + causal_mask.device + ) + padding_mask = padding_mask == 0 + causal_mask[:, :, :, :mask_length] = causal_mask[:, :, :, :mask_length].masked_fill( + padding_mask, min_dtype + ) + return causal_mask + + +@add_start_docstrings( + "The bare Mistral Model outputting raw hidden-states without any specific head on top.", + MISTRAL_START_DOCSTRING, +) +class MistralModel(MistralPreTrainedModel): + """ + Transformer decoder consisting of *config.num_hidden_layers* layers. Each layer is a [`MistralDecoderLayer`] + + Args: + config: MistralConfig + """ + + def __init__(self, config: MistralConfig): + 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( + [MistralDecoderLayer(config, layer_idx) for layer_idx in range(config.num_hidden_layers)] + ) + self.norm = MistralRMSNorm(config.hidden_size, eps=config.rms_norm_eps) + self.rotary_emb = MistralRotaryEmbedding(config=config) + self.gradient_checkpointing = False + + # Initialize weights and apply final processing + self.post_init() + + def get_input_embeddings(self): + return self.embed_tokens + + def set_input_embeddings(self, value): + self.embed_tokens = value + + @can_return_tuple + @add_start_docstrings_to_model_forward(MISTRAL_INPUTS_DOCSTRING) + def forward( + self, + input_ids: Optional[torch.LongTensor] = None, + attention_mask: Optional[torch.Tensor] = None, + position_ids: Optional[torch.LongTensor] = None, + past_key_values: Optional[Cache] = None, + inputs_embeds: Optional[torch.FloatTensor] = None, + use_cache: Optional[bool] = None, + output_attentions: Optional[bool] = None, + output_hidden_states: Optional[bool] = None, + cache_position: Optional[torch.LongTensor] = None, + **flash_attn_kwargs: Unpack[FlashAttentionKwargs], + ) -> BaseModelOutputWithPast: + output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions + output_hidden_states = ( + output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states + ) + use_cache = use_cache if use_cache is not None else self.config.use_cache + + if (input_ids is None) ^ (inputs_embeds is not None): + raise ValueError("You must specify exactly one of input_ids or inputs_embeds") + + if self.gradient_checkpointing and self.training and use_cache: + logger.warning_once( + "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`." + ) + use_cache = False + + # TODO (joao): remove this exception in v4.56 -- it exists for users that try to pass a legacy cache + if not isinstance(past_key_values, (type(None), Cache)): + raise ValueError("The `past_key_values` should be either a `Cache` object or `None`.") + + if inputs_embeds is None: + inputs_embeds = self.embed_tokens(input_ids) + + if use_cache and past_key_values is None: + past_key_values = DynamicCache() + + if cache_position is None: + past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0 + cache_position = torch.arange( + past_seen_tokens, past_seen_tokens + inputs_embeds.shape[1], device=inputs_embeds.device + ) + + if position_ids is None: + position_ids = cache_position.unsqueeze(0) + + causal_mask = self._update_causal_mask( + attention_mask, inputs_embeds, cache_position, past_key_values, output_attentions + ) + + hidden_states = inputs_embeds + + # create position embeddings to be shared across the decoder layers + position_embeddings = self.rotary_emb(hidden_states, position_ids) + + # decoder layers + all_hidden_states = () if output_hidden_states else None + all_self_attns = () if output_attentions else None + + for decoder_layer in self.layers[: self.config.num_hidden_layers]: + if output_hidden_states: + all_hidden_states += (hidden_states,) + + if self.gradient_checkpointing and self.training: + layer_outputs = self._gradient_checkpointing_func( + partial(decoder_layer.__call__, **flash_attn_kwargs), + hidden_states, + causal_mask, + position_ids, + past_key_values, + output_attentions, + use_cache, + cache_position, + position_embeddings, + ) + else: + layer_outputs = decoder_layer( + hidden_states, + attention_mask=causal_mask, + position_ids=position_ids, + past_key_value=past_key_values, + output_attentions=output_attentions, + use_cache=use_cache, + cache_position=cache_position, + position_embeddings=position_embeddings, + **flash_attn_kwargs, + ) + + hidden_states = layer_outputs[0] + + if output_attentions: + all_self_attns += (layer_outputs[1],) + + hidden_states = self.norm(hidden_states) + + # add hidden states from the last decoder layer + if output_hidden_states: + all_hidden_states += (hidden_states,) + + return BaseModelOutputWithPast( + last_hidden_state=hidden_states, + past_key_values=past_key_values if use_cache else None, + hidden_states=all_hidden_states, + attentions=all_self_attns, + ) + + def _update_causal_mask( + self, + attention_mask: torch.Tensor, + input_tensor: torch.Tensor, + cache_position: torch.Tensor, + past_key_values: Cache, + output_attentions: bool = False, + ): + if self.config._attn_implementation == "flash_attention_2": + if attention_mask is not None and past_key_values is not None: + is_padding_right = attention_mask[:, -1].sum().item() != input_tensor.size()[0] + if is_padding_right: + raise ValueError( + "You are attempting to perform batched generation with padding_side='right'" + " this may lead to unexpected behaviour for Flash Attention version of Mistral. Make sure to " + " call `tokenizer.padding_side = 'left'` before tokenizing the input. " + ) + if attention_mask is not None and 0.0 in attention_mask: + return attention_mask + return None + + # For SDPA, when possible, we will rely on its `is_causal` argument instead of its `attn_mask` argument, in + # order to dispatch on Flash Attention 2. This feature is not compatible with static cache, as SDPA will fail + # to infer the attention mask. + past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0 + using_static_cache = isinstance(past_key_values, StaticCache) + using_sliding_window_cache = isinstance(past_key_values, SlidingWindowCache) + + # When output attentions is True, sdpa implementation's forward method calls the eager implementation's forward + if ( + self.config._attn_implementation == "sdpa" + and not (using_static_cache or using_sliding_window_cache) + and not output_attentions + ): + if AttentionMaskConverter._ignore_causal_mask_sdpa( + attention_mask, + inputs_embeds=input_tensor, + past_key_values_length=past_seen_tokens, + sliding_window=self.config.sliding_window, + is_training=self.training, + ): + return None + + dtype, device = input_tensor.dtype, input_tensor.device + min_dtype = torch.finfo(dtype).min + sequence_length = input_tensor.shape[1] + # SlidingWindowCache or StaticCache + if using_sliding_window_cache or using_static_cache: + target_length = past_key_values.get_max_cache_shape() + # DynamicCache or no cache + else: + target_length = ( + attention_mask.shape[-1] + if isinstance(attention_mask, torch.Tensor) + else past_seen_tokens + sequence_length + 1 + ) + + # In case the provided `attention` mask is 2D, we generate a causal mask here (4D). + causal_mask = _prepare_4d_causal_attention_mask_with_cache_position( + attention_mask, + sequence_length=sequence_length, + target_length=target_length, + dtype=dtype, + device=device, + cache_position=cache_position, + batch_size=input_tensor.shape[0], + config=self.config, + past_key_values=past_key_values, + ) + + if ( + self.config._attn_implementation == "sdpa" + and attention_mask is not None + and attention_mask.device.type in ["cuda", "xpu"] + and not output_attentions + ): + # Attend to all tokens in fully masked rows in the causal_mask, for example the relevant first rows when + # using left padding. This is required by F.scaled_dot_product_attention memory-efficient attention path. + # Details: https://github.com/pytorch/pytorch/issues/110213 + causal_mask = AttentionMaskConverter._unmask_unattended(causal_mask, min_dtype) + + return causal_mask + + +class KwargsForCausalLM(FlashAttentionKwargs, LossKwargs): ... + + +class MistralForCausalLM(MistralPreTrainedModel, GenerationMixin): + _tied_weights_keys = ["lm_head.weight"] + _tp_plan = {"lm_head": "colwise_rep"} + _pp_plan = {"lm_head": (["hidden_states"], ["logits"])} + + def __init__(self, config): + super().__init__(config) + self.model = MistralModel(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 get_input_embeddings(self): + return self.model.embed_tokens + + def set_input_embeddings(self, value): + self.model.embed_tokens = value + + def get_output_embeddings(self): + return self.lm_head + + def set_output_embeddings(self, new_embeddings): + self.lm_head = new_embeddings + + def set_decoder(self, decoder): + self.model = decoder + + def get_decoder(self): + return self.model + + @can_return_tuple + @deprecate_kwarg("num_logits_to_keep", version="4.50", new_name="logits_to_keep") + @add_start_docstrings_to_model_forward(MISTRAL_INPUTS_DOCSTRING) + @replace_return_docstrings(output_type=CausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC) + def forward( + self, + input_ids: Optional[torch.LongTensor] = None, + attention_mask: Optional[torch.Tensor] = None, + position_ids: Optional[torch.LongTensor] = None, + past_key_values: Optional[Cache] = None, + inputs_embeds: Optional[torch.FloatTensor] = None, + labels: Optional[torch.LongTensor] = None, + use_cache: Optional[bool] = None, + output_attentions: Optional[bool] = None, + output_hidden_states: Optional[bool] = None, + cache_position: Optional[torch.LongTensor] = None, + logits_to_keep: Union[int, torch.Tensor] = 0, + **kwargs: Unpack[KwargsForCausalLM], + ) -> CausalLMOutputWithPast: + r""" + labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*): + Labels for computing the masked language modeling loss. Indices should either be in `[0, ..., + config.vocab_size]` or -100 (see `input_ids` docstring). Tokens with indices set to `-100` are ignored + (masked), the loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`. + + logits_to_keep (`int` or `torch.Tensor`, *optional*): + If an `int`, compute logits for the last `logits_to_keep` tokens. If `0`, calculate logits for all + `input_ids` (special case). Only last token logits are needed for generation, and calculating them only for that + token can save memory, which becomes pretty significant for long sequences or large vocabulary size. + If a `torch.Tensor`, must be 1D corresponding to the indices to keep in the sequence length dimension. + This is useful when using packed tensor format (single dimension for batch and sequence length). + + Returns: + + Example: + + ```python + >>> from transformers import AutoTokenizer, MistralForCausalLM + + >>> model = MistralForCausalLM.from_pretrained("meta-mistral/Mistral-2-7b-hf") + >>> tokenizer = AutoTokenizer.from_pretrained("meta-mistral/Mistral-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." + ```""" + output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions + output_hidden_states = ( + output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states + ) + + # decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn) + 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, + output_attentions=output_attentions, + output_hidden_states=output_hidden_states, + cache_position=cache_position, + **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, + ) + + +@add_start_docstrings( + """ + The Mistral Model transformer with a token classification head on top (a linear layer on top of the hidden-states + output) e.g. for Named-Entity-Recognition (NER) tasks. + """, + MISTRAL_START_DOCSTRING, +) +class MistralForTokenClassification(MistralPreTrainedModel): + def __init__(self, config): + super().__init__(config) + self.num_labels = config.num_labels + self.model = MistralModel(config) + if getattr(config, "classifier_dropout", None) is not None: + classifier_dropout = config.classifier_dropout + elif getattr(config, "hidden_dropout", None) is not None: + classifier_dropout = config.hidden_dropout + else: + classifier_dropout = 0.1 + self.dropout = nn.Dropout(classifier_dropout) + self.score = nn.Linear(config.hidden_size, config.num_labels) + + # Initialize weights and apply final processing + self.post_init() + + def get_input_embeddings(self): + return self.model.embed_tokens + + def set_input_embeddings(self, value): + self.model.embed_tokens = value + + @can_return_tuple + @add_start_docstrings_to_model_forward(MISTRAL_INPUTS_DOCSTRING) + @add_code_sample_docstrings( + checkpoint=_CHECKPOINT_FOR_DOC, + output_type=TokenClassifierOutput, + config_class=_CONFIG_FOR_DOC, + ) + def forward( + self, + input_ids: Optional[torch.LongTensor] = None, + attention_mask: Optional[torch.Tensor] = None, + position_ids: Optional[torch.LongTensor] = None, + past_key_values: Optional[Cache] = None, + inputs_embeds: Optional[torch.FloatTensor] = None, + labels: Optional[torch.LongTensor] = None, + use_cache: Optional[bool] = None, + output_attentions: Optional[bool] = None, + output_hidden_states: Optional[bool] = None, + ) -> TokenClassifierOutput: + r""" + labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*): + Labels for computing the sequence classification/regression loss. Indices should be in `[0, ..., + config.num_labels - 1]`. If `config.num_labels == 1` a regression loss is computed (Mean-Square loss), If + `config.num_labels > 1` a classification loss is computed (Cross-Entropy). + """ + + outputs: BaseModelOutputWithPast = self.model( + input_ids, + attention_mask=attention_mask, + position_ids=position_ids, + past_key_values=past_key_values, + inputs_embeds=inputs_embeds, + use_cache=use_cache, + output_attentions=output_attentions, + output_hidden_states=output_hidden_states, + ) + sequence_output = outputs.last_hidden_state + sequence_output = self.dropout(sequence_output) + logits = self.score(sequence_output) + + loss = None + if labels is not None: + loss = self.loss_function(logits, labels, self.config) + + return TokenClassifierOutput( + loss=loss, + logits=logits, + hidden_states=outputs.hidden_states, + attentions=outputs.attentions, + ) + + +@add_start_docstrings( + """ + The Mistral Model transformer with a sequence classification head on top (linear layer). + + [`MistralForSequenceClassification`] uses the last token in order to do the classification, as other causal models + (e.g. GPT-2) do. + + Since it does classification on the last token, it requires to know the position of the last token. If a + `pad_token_id` is defined in the configuration, it finds the last token that is not a padding token in each row. If + no `pad_token_id` is defined, it simply takes the last value in each row of the batch. Since it cannot guess the + padding tokens when `inputs_embeds` are passed instead of `input_ids`, it does the same (take the last value in + each row of the batch). + """, + MISTRAL_START_DOCSTRING, +) +class MistralForSequenceClassification(MistralPreTrainedModel): + def __init__(self, config): + super().__init__(config) + self.num_labels = config.num_labels + self.model = MistralModel(config) + self.score = nn.Linear(config.hidden_size, self.num_labels, bias=False) + + # Initialize weights and apply final processing + self.post_init() + + def get_input_embeddings(self): + return self.model.embed_tokens + + def set_input_embeddings(self, value): + self.model.embed_tokens = value + + @can_return_tuple + @add_start_docstrings_to_model_forward(MISTRAL_INPUTS_DOCSTRING) + def forward( + self, + input_ids: Optional[torch.LongTensor] = None, + attention_mask: Optional[torch.Tensor] = None, + position_ids: Optional[torch.LongTensor] = None, + past_key_values: Optional[Cache] = None, + inputs_embeds: Optional[torch.FloatTensor] = None, + labels: Optional[torch.LongTensor] = None, + use_cache: Optional[bool] = None, + output_attentions: Optional[bool] = None, + output_hidden_states: Optional[bool] = None, + ) -> SequenceClassifierOutputWithPast: + r""" + labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*): + Labels for computing the sequence classification/regression loss. Indices should be in `[0, ..., + config.num_labels - 1]`. If `config.num_labels == 1` a regression loss is computed (Mean-Square loss), If + `config.num_labels > 1` a classification loss is computed (Cross-Entropy). + """ + + transformer_outputs: BaseModelOutputWithPast = self.model( + input_ids, + attention_mask=attention_mask, + position_ids=position_ids, + past_key_values=past_key_values, + inputs_embeds=inputs_embeds, + use_cache=use_cache, + output_attentions=output_attentions, + output_hidden_states=output_hidden_states, + ) + hidden_states = transformer_outputs.last_hidden_state + logits = self.score(hidden_states) + + if input_ids is not None: + batch_size = input_ids.shape[0] + else: + batch_size = inputs_embeds.shape[0] + + if self.config.pad_token_id is None and batch_size != 1: + raise ValueError("Cannot handle batch sizes > 1 if no padding token is defined.") + if self.config.pad_token_id is None: + last_non_pad_token = -1 + elif input_ids is not None: + # To handle both left- and right- padding, we take the rightmost token that is not equal to pad_token_id + non_pad_mask = (input_ids != self.config.pad_token_id).to(logits.device, torch.int32) + token_indices = torch.arange(input_ids.shape[-1], device=logits.device, dtype=torch.int32) + last_non_pad_token = (token_indices * non_pad_mask).argmax(-1) + else: + last_non_pad_token = -1 + logger.warning_once( + f"{self.__class__.__name__} will not detect padding tokens in `inputs_embeds`. Results may be " + "unexpected if using padding tokens in conjunction with `inputs_embeds.`" + ) + + pooled_logits = logits[torch.arange(batch_size, device=logits.device), last_non_pad_token] + + loss = None + if labels is not None: + loss = self.loss_function(logits=logits, labels=labels, pooled_logits=pooled_logits, config=self.config) + + return SequenceClassifierOutputWithPast( + loss=loss, + logits=pooled_logits, + past_key_values=transformer_outputs.past_key_values, + hidden_states=transformer_outputs.hidden_states, + attentions=transformer_outputs.attentions, + ) + + +@add_start_docstrings( + """ +The Mistral Model transformer with a span classification head on top for extractive question-answering tasks like +SQuAD (a linear layer on top of the hidden-states output to compute `span start logits` and `span end logits`). + """, + MISTRAL_START_DOCSTRING, +) +class MistralForQuestionAnswering(MistralPreTrainedModel): + base_model_prefix = "model" + + def __init__(self, config): + super().__init__(config) + self.qa_outputs = nn.Linear(config.hidden_size, 2) + self.model = MistralModel(config) # diff with Llama: transformer->model + + # Initialize weights and apply final processing + self.post_init() + + def get_input_embeddings(self): + return self.model.embed_tokens + + def set_input_embeddings(self, value): + self.model.embed_tokens = value + + @can_return_tuple + @add_start_docstrings_to_model_forward(MISTRAL_INPUTS_DOCSTRING) + def forward( + self, + input_ids: Optional[torch.LongTensor] = None, + attention_mask: Optional[torch.FloatTensor] = None, + position_ids: Optional[torch.LongTensor] = None, + past_key_values: Optional[Union[Cache, List[torch.FloatTensor]]] = None, + inputs_embeds: Optional[torch.FloatTensor] = None, + start_positions: Optional[torch.LongTensor] = None, + end_positions: Optional[torch.LongTensor] = None, + output_attentions: Optional[bool] = None, + output_hidden_states: Optional[bool] = None, + **kwargs, + ) -> QuestionAnsweringModelOutput: + r""" + start_positions (`torch.LongTensor` of shape `(batch_size,)`, *optional*): + Labels for position (index) of the start of the labelled span for computing the token classification loss. + Positions are clamped to the length of the sequence (`sequence_length`). Position outside of the sequence + are not taken into account for computing the loss. + end_positions (`torch.LongTensor` of shape `(batch_size,)`, *optional*): + Labels for position (index) of the end of the labelled span for computing the token classification loss. + Positions are clamped to the length of the sequence (`sequence_length`). Position outside of the sequence + are not taken into account for computing the loss. + """ + + outputs: BaseModelOutputWithPast = self.model( + input_ids, + attention_mask=attention_mask, + position_ids=position_ids, + past_key_values=past_key_values, + inputs_embeds=inputs_embeds, + output_attentions=output_attentions, + output_hidden_states=output_hidden_states, + ) + + sequence_output = outputs.last_hidden_state + + logits = self.qa_outputs(sequence_output) + start_logits, end_logits = logits.split(1, dim=-1) + start_logits = start_logits.squeeze(-1).contiguous() + end_logits = end_logits.squeeze(-1).contiguous() + + loss = None + if start_positions is not None and end_positions is not None: + loss = self.loss_function(start_logits, end_logits, start_positions, end_positions, **kwargs) + + return QuestionAnsweringModelOutput( + loss=loss, + start_logits=start_logits, + end_logits=end_logits, + hidden_states=outputs.hidden_states, + attentions=outputs.attentions, + ) diff --git a/src/llmcompressor/transformers/tracing/mistral3.py b/src/llmcompressor/transformers/tracing/mistral3.py new file mode 100644 index 0000000000..2acb48b8e1 --- /dev/null +++ b/src/llmcompressor/transformers/tracing/mistral3.py @@ -0,0 +1,559 @@ +# 🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨 +# This file was automatically generated from src/transformers/models/mistral3/modular_mistral3.py. +# Do NOT edit this file manually as any edits will be overwritten by the generation of +# the file from the modular. If any change should be done, please apply the change to the +# modular_mistral3.py file directly. One of our CI enforces this. +# 🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨 +# coding=utf-8 +# Copyright 2025 HuggingFace Inc. team. All rights reserved. +# +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +from dataclasses import dataclass +from typing import List, Optional, Tuple, Union + +import torch +from torch import nn + +from transformers.activations import ACT2FN +from transformers.generation import GenerationMixin +from transformers.modeling_outputs import ModelOutput +from transformers.modeling_utils import PreTrainedModel +from transformers.utils import ( + add_start_docstrings, + add_start_docstrings_to_model_forward, + is_torchdynamo_compiling, + replace_return_docstrings, +) +from transformers.utils.deprecation import deprecate_kwarg +from transformers.models.auto import AutoModel#, AutoModelForCausalLM +from transformers.models.mistral3.configuration_mistral3 import Mistral3Config +from llmcompressor.transformers.tracing import TraceableMistralForCausalLM + + +_CONFIG_FOR_DOC = "Mistral3Config" + + +class Mistral3RMSNorm(nn.Module): + def __init__(self, hidden_size, eps=1e-6): + """ + Mistral3RMSNorm is equivalent to T5LayerNorm + """ + super().__init__() + self.weight = nn.Parameter(torch.ones(hidden_size)) + self.variance_epsilon = eps + + def forward(self, hidden_states): + input_dtype = hidden_states.dtype + hidden_states = hidden_states.to(torch.float32) + variance = hidden_states.pow(2).mean(-1, keepdim=True) + hidden_states = hidden_states * torch.rsqrt(variance + self.variance_epsilon) + return self.weight * hidden_states.to(input_dtype) + + def extra_repr(self): + return f"{tuple(self.weight.shape)}, eps={self.variance_epsilon}" + + +class Mistral3PatchMerger(nn.Module): + """ + Learned merging of spatial_merge_size ** 2 patches + """ + + def __init__(self, config: Mistral3Config): + super().__init__() + self.config = config + + hidden_size = config.vision_config.hidden_size + self.spatial_merge_size = config.spatial_merge_size + self.patch_size = self.config.vision_config.patch_size + self.merging_layer = nn.Linear(hidden_size * self.spatial_merge_size**2, hidden_size, bias=False) + + def forward(self, image_features: torch.Tensor, image_sizes: torch.Tensor) -> torch.Tensor: + image_sizes = [ + (image_size[0] // self.patch_size, image_size[1] // self.patch_size) for image_size in image_sizes + ] + + tokens_per_image = [h * w for h, w in image_sizes] + d = image_features.shape[-1] + + permuted_tensor = [] + for image_index, image_tokens in enumerate(image_features.split(tokens_per_image)): + # Reshape image_tokens into a 2D grid + h, w = image_sizes[image_index] + image_grid = image_tokens.view(h, w, d).permute(2, 0, 1).unsqueeze(0) + grid = torch.nn.functional.unfold( + image_grid, kernel_size=self.spatial_merge_size, stride=self.spatial_merge_size + ) + grid = grid.view(d * self.spatial_merge_size**2, -1).t() + permuted_tensor.append(grid) + + image_features = torch.cat(permuted_tensor, dim=0) + image_features = self.merging_layer(image_features) + return image_features + + +class Mistral3MultiModalProjector(nn.Module): + def __init__(self, config: Mistral3Config): + super().__init__() + self.norm = Mistral3RMSNorm(config.vision_config.hidden_size) + self.patch_merger = Mistral3PatchMerger(config) + # We have hidden_size * the number of vision feature layers + num_feature_layers = 1 if isinstance(config.vision_feature_layer, int) else len(config.vision_feature_layer) + self.linear_1 = nn.Linear( + config.vision_config.hidden_size * num_feature_layers, + config.text_config.hidden_size, + bias=config.multimodal_projector_bias, + ) + self.act = ACT2FN[config.projector_hidden_act] + self.linear_2 = nn.Linear( + config.text_config.hidden_size, config.text_config.hidden_size, bias=config.multimodal_projector_bias + ) + + def forward(self, image_features: torch.Tensor, image_sizes: torch.Tensor): + image_features = self.norm(image_features) + image_features = self.patch_merger(image_features, image_sizes) + hidden_states = self.linear_1(image_features) + hidden_states = self.act(hidden_states) + hidden_states = self.linear_2(hidden_states) + return hidden_states + + +@dataclass +class Mistral3CausalLMOutputWithPast(ModelOutput): + """ + Base class for Mistral3 causal language model (or autoregressive) outputs. + + Args: + loss (`torch.FloatTensor` of shape `(1,)`, *optional*, returned when `labels` is provided): + Language modeling loss (for next-token prediction). + logits (`torch.FloatTensor` of shape `(batch_size, sequence_length, config.vocab_size)`): + Prediction scores of the language modeling head (scores for each vocabulary token before SoftMax). + past_key_values (`tuple(tuple(torch.FloatTensor))`, *optional*, returned when `use_cache=True` is passed or when `config.use_cache=True`): + Tuple of `tuple(torch.FloatTensor)` of length `config.n_layers`, with each tuple having 2 tensors of shape + `(batch_size, num_heads, sequence_length, embed_size_per_head)`) + + Contains pre-computed hidden-states (key and values in the self-attention blocks) that can be used (see + `past_key_values` input) to speed up sequential decoding. + hidden_states (`tuple(torch.FloatTensor)`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`): + Tuple of `torch.FloatTensor` (one for the output of the embeddings, if the model has an embedding layer, + + one for the output of each layer) of shape `(batch_size, sequence_length, hidden_size)`. + + Hidden-states of the model at the output of each layer plus the optional initial embedding outputs. + attentions (`tuple(torch.FloatTensor)`, *optional*, returned when `output_attentions=True` is passed or when `config.output_attentions=True`): + Tuple of `torch.FloatTensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length, + sequence_length)`. + + Attentions weights after the attention softmax, used to compute the weighted average in the self-attention + heads. + image_hidden_states (`torch.FloatTensor`, *optional*): + A `torch.FloatTensor` of size (batch_size, num_images, sequence_length, hidden_size)`. + image_hidden_states of the model produced by the vision encoder and after projecting the last hidden state. + """ + + loss: Optional[torch.FloatTensor] = None + logits: Optional[torch.FloatTensor] = None + past_key_values: Optional[List[torch.FloatTensor]] = None + hidden_states: Optional[Tuple[torch.FloatTensor]] = None + attentions: Optional[Tuple[torch.FloatTensor]] = None + image_hidden_states: Optional[torch.FloatTensor] = None + + +MISTRAL3_START_DOCSTRING = r""" + This model inherits from [`PreTrainedModel`]. Check the superclass documentation for the generic methods the + library implements for all its model (such as downloading or saving, resizing the input embeddings, pruning heads + etc.) + + This model is also a PyTorch [torch.nn.Module](https://pytorch.org/docs/stable/nn.html#torch.nn.Module) subclass. + Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage + and behavior. + + Parameters: + config ([`Mistral3Config`] or [`Mistral3VisionConfig`]): + Model configuration class with all the parameters of the model. Initializing with a config file does not + load the weights associated with the model, only the configuration. Check out the + [`~PreTrainedModel.from_pretrained`] method to load the model weights. +""" + + +@add_start_docstrings( + "The bare LLaMA Model outputting raw hidden-states without any specific head on top.", + MISTRAL3_START_DOCSTRING, +) +class Mistral3PreTrainedModel(PreTrainedModel): + config_class = Mistral3Config + base_model_prefix = "model" + supports_gradient_checkpointing = True + _no_split_modules = ["Mistral3VisionAttention"] + _skip_keys_device_placement = "past_key_values" + _supports_cache_class = True + _supports_flash_attn_2 = True + _supports_sdpa = True + _supports_quantized_cache = True + _supports_static_cache = True + + def _init_weights(self, module): + # important: this ported version of Mistral3 isn't meant for training from scratch - only + # inference and fine-tuning - so the proper init weights code has been removed - the original codebase + # https://github.com/haotian-liu/Mistral3/tree/main/mistral3 should serve for that purpose + std = ( + self.config.initializer_range + if hasattr(self.config, "initializer_range") + else self.config.text_config.initializer_range + ) + + if hasattr(module, "class_embedding"): + module.class_embedding.data.normal_(mean=0.0, std=std) + + if isinstance(module, (nn.Linear, nn.Conv2d)): + module.weight.data.normal_(mean=0.0, std=std) + if module.bias is not None: + module.bias.data.zero_() + elif isinstance(module, nn.Embedding): + module.weight.data.normal_(mean=0.0, std=std) + if module.padding_idx is not None: + module.weight.data[module.padding_idx].zero_() + + +MISTRAL3_INPUTS_DOCSTRING = r""" + Args: + input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`): + Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you provide + it. + + Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and + [`PreTrainedTokenizer.__call__`] for details. + + [What are input IDs?](../glossary#input-ids) + pixel_values (`torch.FloatTensor` of shape `(batch_size, num_channels, image_size, image_size)): + The tensors corresponding to the input images. Pixel values can be obtained using + [`AutoImageProcessor`]. See [`CLIPImageProcessor.__call__`] for details ([]`Mistral3Processor`] uses + [`CLIPImageProcessor`] for processing images). + attention_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*): + Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`: + + - 1 for tokens that are **not masked**, + - 0 for tokens that are **masked**. + + [What are attention masks?](../glossary#attention-mask) + + Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and + [`PreTrainedTokenizer.__call__`] for details. + + If `past_key_values` is used, optionally only the last `decoder_input_ids` have to be input (see + `past_key_values`). + + If you want to change padding behavior, you should read [`modeling_opt._prepare_decoder_attention_mask`] + and modify to your needs. See diagram 1 in [the paper](https://arxiv.org/abs/1910.13461) for more + information on the default strategy. + + - 1 indicates the head is **not masked**, + - 0 indicates the head is **masked**. + position_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*): + Indices of positions of each input sequence tokens in the position embeddings. Selected in the range `[0, + config.n_positions - 1]`. [What are position IDs?](../glossary#position-ids) + past_key_values (`tuple(tuple(torch.FloatTensor))`, *optional*, returned when `use_cache=True` is passed or when `config.use_cache=True`): + Tuple of `tuple(torch.FloatTensor)` of length `config.n_layers`, with each tuple having 2 tensors of shape + `(batch_size, num_heads, sequence_length, embed_size_per_head)`) and 2 additional tensors of shape + `(batch_size, num_heads, encoder_sequence_length, embed_size_per_head)`. + + Contains pre-computed hidden-states (key and values in the self-attention blocks and in the cross-attention + blocks) that can be used (see `past_key_values` input) to speed up sequential decoding. + + If `past_key_values` are used, the user can optionally input only the last `decoder_input_ids` (those that + don't have their past key value states given to this model) of shape `(batch_size, 1)` instead of all + `decoder_input_ids` of shape `(batch_size, sequence_length)`. + inputs_embeds (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*): + Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation. This + is useful if you want more control over how to convert `input_ids` indices into associated vectors than the + model's internal embedding lookup matrix. + vision_feature_layer (`Union[int, List[int]], *optional*, defaults to -2`): + The index of the layer to select the vision feature. If multiple indices are provided, + the vision feature of the corresponding indices will be concatenated to form the + vision features. + vision_feature_select_strategy (`str`, *optional*, defaults to `"default"`): + The feature selection strategy used to select the vision feature from the vision backbone. + Can be one of `"default"` or `"full"`. + use_cache (`bool`, *optional*): + If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding (see + `past_key_values`). + output_attentions (`bool`, *optional*): + Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned + tensors for more detail. + output_hidden_states (`bool`, *optional*): + Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for + more detail. + return_dict (`bool`, *optional*): + Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple. + cache_position (`torch.LongTensor` of shape `(sequence_length)`, *optional*): + Indices depicting the position of the input sequence tokens in the sequence. Contrarily to `position_ids`, + this tensor is not affected by padding. It is used to update the cache in the correct position and to infer + the complete sequence length. +""" + + +@add_start_docstrings( + """The MISTRAL3 model which consists of a vision backbone and a language model.""", + MISTRAL3_START_DOCSTRING, +) +class Mistral3ForConditionalGeneration(Mistral3PreTrainedModel, GenerationMixin): + def __init__(self, config: Mistral3Config): + super().__init__(config) + self.vision_tower = AutoModel.from_config(config.vision_config) + + self.multi_modal_projector = Mistral3MultiModalProjector(config) + self.vocab_size = config.text_config.vocab_size + # + # TRACING: Set language_model to tracing-enabled version of MistralForCausalLM + # + self.language_model = TraceableMistralForCausalLM(config.text_config) + if self.language_model._tied_weights_keys is not None: + self._tied_weights_keys = [f"language_model.{k}" for k in self.language_model._tied_weights_keys] + + self.pad_token_id = self.config.pad_token_id if self.config.pad_token_id is not None else -1 + + self.post_init() + + def get_input_embeddings(self): + return self.language_model.get_input_embeddings() + + def set_input_embeddings(self, value): + self.language_model.set_input_embeddings(value) + + def get_output_embeddings(self): + return self.language_model.get_output_embeddings() + + def set_output_embeddings(self, new_embeddings): + self.language_model.set_output_embeddings(new_embeddings) + + def set_decoder(self, decoder): + self.language_model.set_decoder(decoder) + + def get_decoder(self): + return self.language_model.get_decoder() + + def get_image_features( + self, + pixel_values: torch.FloatTensor, + vision_feature_layer: Union[int, List[int]], + image_sizes: torch.Tensor, + **kwargs, + ): + """ + Obtains image last hidden states from the vision tower and apply multimodal projection. + + Args: + pixel_values (`torch.FloatTensor]` of shape `(batch_size, channels, height, width)`): + The tensors corresponding to the input images. + vision_feature_layer (`Union[int, List[int]]`): + The index of the layer to select the vision feature. If multiple indices are provided, + the vision feature of the corresponding indices will be concatenated to form the + vision features. + image_sizes (`torch.Tensor`): + Tensor containing the image sizes as returned by the processor. + Returns: + image_features (`torch.Tensor`): Image feature tensor of shape `(num_images, image_length, embed_dim)`). + """ + kwargs = {k: v for k, v in kwargs.items() if v is not None} + # this is not memory efficient at all (output_hidden_states=True) will save all the hidden states. + image_outputs = self.vision_tower(pixel_values, image_sizes=image_sizes, output_hidden_states=True, **kwargs) + # If we have one vision feature layer, return the corresponding hidden states, + # otherwise, select the hidden states of each feature layer and concatenate them + if isinstance(vision_feature_layer, int): + selected_image_feature = image_outputs.hidden_states[vision_feature_layer] + else: + hs_pool = [image_outputs.hidden_states[layer_idx] for layer_idx in vision_feature_layer] + selected_image_feature = torch.cat(hs_pool, dim=-1) + + image_features = self.multi_modal_projector(selected_image_feature.squeeze(0), image_sizes) + return image_features + + @deprecate_kwarg("num_logits_to_keep", version="4.50", new_name="logits_to_keep") + @add_start_docstrings_to_model_forward(MISTRAL3_INPUTS_DOCSTRING) + @replace_return_docstrings(output_type=Mistral3CausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC) + def forward( + self, + input_ids: Optional[torch.LongTensor] = None, + pixel_values: Optional[torch.FloatTensor] = None, + attention_mask: Optional[torch.Tensor] = None, + position_ids: Optional[torch.LongTensor] = None, + past_key_values: Optional[List[torch.FloatTensor]] = None, + inputs_embeds: Optional[torch.FloatTensor] = None, + vision_feature_layer: Optional[Union[int, List[int]]] = None, + labels: Optional[torch.LongTensor] = None, + use_cache: Optional[bool] = None, + output_attentions: Optional[bool] = None, + output_hidden_states: Optional[bool] = None, + return_dict: Optional[bool] = None, + cache_position: Optional[torch.LongTensor] = None, + logits_to_keep: Union[int, torch.Tensor] = 0, + image_sizes: Optional[torch.Tensor] = None, + **lm_kwargs, + ) -> Union[Tuple, Mistral3CausalLMOutputWithPast]: + r""" + labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*): + Labels for computing the masked language modeling loss. Indices should either be in `[0, ..., + config.vocab_size]` or -100 (see `input_ids` docstring). Tokens with indices set to `-100` are ignored + (masked), the loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`. + + logits_to_keep (`int` or `torch.Tensor`, *optional*): + If an `int`, compute logits for the last `logits_to_keep` tokens. If `0`, calculate logits for all + `input_ids` (special case). Only last token logits are needed for generation, and calculating them only for that + token can save memory, which becomes pretty significant for long sequences or large vocabulary size. + If a `torch.Tensor`, must be 1D corresponding to the indices to keep in the sequence length dimension. + This is useful when using packed tensor format (single dimension for batch and sequence length). + + + Returns: + + Example: + + ```python + >>> from PIL import Image + >>> import requests + >>> from transformers import AutoProcessor, Mistral3ForConditionalGeneration + + >>> model = Mistral3ForConditionalGeneration.from_pretrained("mistralai/Mistral-Small-3.1-24B-Instruct-2503") + >>> processor = AutoProcessor.from_pretrained("mistralai/Mistral-Small-3.1-24B-Instruct-2503") + + >>> prompt = "[INST][IMG]What is the image?[/INST]" + >>> url = "http://images.cocodataset.org/val2017/000000039769.jpg" + >>> image = Image.open(requests.get(url, stream=True).raw) + + >>> inputs = processor(images=image, text=prompt, return_tensors="pt") + + >>> # Generate + >>> generate_ids = model.generate(**inputs, max_new_tokens=15) + >>> processor.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0] + "What is the image?The image depicts two cats lying on a pink blanket." + ```""" + + output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions + output_hidden_states = ( + output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states + ) + return_dict = return_dict if return_dict is not None else self.config.use_return_dict + vision_feature_layer = ( + vision_feature_layer if vision_feature_layer is not None else self.config.vision_feature_layer + ) + + if (input_ids is None) ^ (inputs_embeds is not None): + raise ValueError("You must specify exactly one of input_ids or inputs_embeds") + + if pixel_values is not None and inputs_embeds is not None: + raise ValueError( + "You cannot specify both pixel_values and inputs_embeds at the same time, and must specify either one" + ) + + if inputs_embeds is None: + inputs_embeds = self.get_input_embeddings()(input_ids) + + if pixel_values is not None: + image_features = self.get_image_features( + pixel_values=pixel_values, + vision_feature_layer=vision_feature_layer, + image_sizes=image_sizes, + ) + + special_image_mask = (input_ids == self.config.image_token_index).unsqueeze(-1) + special_image_mask = special_image_mask.expand_as(inputs_embeds).to(inputs_embeds.device) + # + # TRACING: Skip check to enable tracing + # + # if not is_torchdynamo_compiling() and inputs_embeds[special_image_mask].numel() != image_features.numel(): + # n_image_tokens = (input_ids == self.config.image_token_index).sum() + # n_image_features = image_features.shape[0] * image_features.shape[1] + # raise ValueError( + # f"Image features and image tokens do not match: tokens: {n_image_tokens}, features {n_image_features}" + # ) + image_features = image_features.to(inputs_embeds.device, inputs_embeds.dtype) + inputs_embeds = inputs_embeds.masked_scatter(special_image_mask, image_features) + + outputs = self.language_model( + attention_mask=attention_mask, + position_ids=position_ids, + past_key_values=past_key_values, + inputs_embeds=inputs_embeds, + use_cache=use_cache, + output_attentions=output_attentions, + output_hidden_states=output_hidden_states, + return_dict=return_dict, + cache_position=cache_position, + logits_to_keep=logits_to_keep, + **lm_kwargs, + ) + + logits = outputs[0] + + loss = None + if labels is not None: + # Shift so that tokens < n predict n + if attention_mask is not None: + # we use the input attention mask to shift the logits and labels, because it is 2D. + # we also crop attn mask in case it is longer, which happens in PrefixTuning with peft + shift_attention_mask = attention_mask[:, -(logits.shape[1] - 1) :].to(logits.device) + shift_logits = logits[..., :-1, :][shift_attention_mask.to(logits.device) != 0].contiguous() + shift_labels = labels[..., 1:][shift_attention_mask.to(labels.device) != 0].contiguous() + else: + shift_logits = logits[..., :-1, :].contiguous() + shift_labels = labels[..., 1:].contiguous() + # Flatten the tokens + loss_fct = nn.CrossEntropyLoss() + loss = loss_fct( + shift_logits.view(-1, shift_logits.size(-1)), shift_labels.view(-1).to(shift_logits.device) + ) + + if not return_dict: + output = (logits,) + outputs[1:] + return (loss,) + output if loss is not None else output + + return Mistral3CausalLMOutputWithPast( + loss=loss, + logits=logits, + past_key_values=outputs.past_key_values, + hidden_states=outputs.hidden_states, + attentions=outputs.attentions, + image_hidden_states=image_features if pixel_values is not None else None, + ) + + def prepare_inputs_for_generation( + self, + input_ids, + past_key_values=None, + inputs_embeds=None, + pixel_values=None, + attention_mask=None, + cache_position=None, + logits_to_keep=None, + **kwargs, + ): + # Overwritten -- in specific circumstances we don't want to forward image inputs to the model + + model_inputs = self.language_model.prepare_inputs_for_generation( + input_ids, + past_key_values=past_key_values, + inputs_embeds=inputs_embeds, + attention_mask=attention_mask, + cache_position=cache_position, + logits_to_keep=logits_to_keep, + **kwargs, + ) + + if cache_position[0] == 0: + # If we're in cached decoding stage, pixel values should be None because input ids do not contain special image token anymore + # Otherwise we need pixel values to be passed to model + model_inputs["pixel_values"] = pixel_values + + return model_inputs + + +__all__ = ["Mistral3PreTrainedModel", "Mistral3ForConditionalGeneration"]