Initial model upload with custom code

modeling_qwen2.py

@@ -1,3 +1,8 @@
+import logging
+from transformers import GenerationMixin
+import torch
+from typing import Optional, Union, List
+from transformers.modeling_outputs import CausalLMOutputWithPast
 # Copyright 2024 The Qwen team, Alibaba Group and The HuggingFace Inc. team. All rights reserved.
 # Copyright 2025 Bytedance Ltd. and/or its affiliates
 #
@@ -38,20 +43,14 @@ from transformers.utils import (
     replace_return_docstrings,
 )
 
-from veomni.models.transformers.qwen2.generation_utils import MDMGenerationMixin
 
-from ....data.constants import IGNORE_INDEX
-from ....distributed.parallel_state import get_parallel_state
-from ....distributed.sequence_parallel import (
     gather_heads_scatter_seq,
     gather_seq_scatter_heads,
     reduce_sequence_parallel_loss,
 )
-from ....utils import logging
-from ....utils.import_utils import is_liger_kernel_available
 
 
-if is_liger_kernel_available():
+if False:
     from liger_kernel.transformers import LigerFusedLinearCrossEntropyLoss  # type: ignore
     from liger_kernel.transformers.rms_norm import LigerRMSNorm
     from liger_kernel.transformers.rope import liger_rotary_pos_emb
@@ -183,7 +182,7 @@ class Qwen2Attention(nn.Module):
         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)
-        if get_parallel_state().ulysses_enabled:
+        if False:
             query_states = gather_seq_scatter_heads(query_states, seq_dim=2, head_dim=1)
             key_states = gather_seq_scatter_heads(key_states, seq_dim=2, head_dim=1)
             value_states = gather_seq_scatter_heads(value_states, seq_dim=2, head_dim=1)
@@ -229,7 +228,7 @@ class Qwen2Attention(nn.Module):
         )
 
         attn_output = attn_output.reshape(bsz, full_q_len, -1, self.head_dim).contiguous()
-        if get_parallel_state().ulysses_enabled:
+        if False:
             attn_output = gather_heads_scatter_seq(attn_output, head_dim=2, seq_dim=1)
 
         attn_output = attn_output.reshape(bsz, q_len, self.config.hidden_size).contiguous()
@@ -533,115 +532,79 @@ class Qwen2Model(Qwen2PreTrainedModel):
     def set_input_embeddings(self, value):
         self.embed_tokens = value
 
+   
     @add_start_docstrings_to_model_forward(QWEN2_INPUTS_DOCSTRING)
+    @replace_return_docstrings(output_type=CausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC)
     def forward(
         self,
         input_ids: torch.LongTensor = None,
         attention_mask: Optional[torch.Tensor] = None,
         position_ids: Optional[torch.LongTensor] = None,
-        past_key_values: Optional[Cache] = None,
+        past_key_values: Optional[Union[Cache, List[torch.FloatTensor]]] = 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,
         return_dict: Optional[bool] = None,
         cache_position: Optional[torch.LongTensor] = None,
         is_causal: bool = True,
-        **flash_attn_kwargs: Unpack[FlashAttentionKwargs],
-    ) -> Union[Tuple, BaseModelOutputWithPast]:
+        **kwargs,
+    ) -> Union[Tuple, CausalLMOutputWithPast]:
+        r\"\"\"
+        Args:
+            labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
+                Labels for computing the masked language modeling loss. Indices should be in `[0, ...,
+                config.vocab_size - 1]`. Tokens with indices set to `-100` are ignored (masked), the loss is only
+                computed for the tokens with labels in `[0, ..., config.vocab_size - 1]`.
+        \"\"\"
         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
         return_dict = return_dict if return_dict is not None else self.config.use_return_dict
 
-        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
-
-        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
+        outputs = 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,
+            return_dict=return_dict,
+            cache_position=cache_position,
+            is_causal=is_causal,
+            **kwargs,
         )
 
-        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(
-                    decoder_layer.__call__,
-                    hidden_states,
-                    causal_mask,
-                    position_ids,
-                    past_key_values,
-                    output_attentions,
-                    use_cache,
-                    cache_position,
-                    position_embeddings,
-                    is_causal,
-                )
-            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,
-                    is_causal=is_causal,
-                    **flash_attn_kwargs,
-                )
-
-            hidden_states = layer_outputs[0]
-
-            if output_attentions:
-                all_self_attns += (layer_outputs[1],)
-
-        hidden_states = self.norm(hidden_states)
+        hidden_states = outputs[0]
+        logits = self.lm_head(hidden_states)
+        logits = logits.float()
+        loss = None
+        
+        if labels is not None:
+            # Maintained for compatibility with Trainer API, but not essential for pure inference
+            shift_logits = logits[..., :-1, :].contiguous()
+            shift_labels = labels[..., 1:].contiguous()
+            loss_fct = torch.nn.CrossEntropyLoss()
+            shift_logits = shift_logits.view(-1, self.config.vocab_size)
+            shift_labels = shift_labels.view(-1)
+            shift_labels = shift_labels.to(shift_logits.device)
+            loss = loss_fct(shift_logits, shift_labels)
 
-        # add hidden states from the last decoder layer
-        if output_hidden_states:
-            all_hidden_states += (hidden_states,)
+        if not return_dict:
+            output = (logits,) + outputs[1:]
+            return (loss,) + output if loss is not None else output
 
-        output = 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,
+        return CausalLMOutputWithPast(
+            loss=loss,
+            logits=logits,
+            past_key_values=outputs.past_key_values,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
         )
-        return output if return_dict else output.to_tuple()
 
     def _update_causal_mask(
         self,
@@ -799,7 +762,7 @@ class Qwen2Model(Qwen2PreTrainedModel):
 class KwargsForCausalLM(FlashAttentionKwargs, ): ...
 
 
-class Qwen2ForCausalLM(Qwen2PreTrainedModel,  MDMGenerationMixin):
+class Qwen2ForCausalLM(Qwen2PreTrainedModel,  GenerationMixin):
     _tied_weights_keys = ["lm_head.weight"]
     _tp_plan = {"lm_head": "colwise_rep"}
     _pp_plan = {"lm_head": (["hidden_states"], ["logits"])}
@@ -831,6 +794,7 @@ class Qwen2ForCausalLM(Qwen2PreTrainedModel,  MDMGenerationMixin):
     def get_decoder(self):
         return self.model
 
+   
     @add_start_docstrings_to_model_forward(QWEN2_INPUTS_DOCSTRING)
     @replace_return_docstrings(output_type=CausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC)
     def forward(
@@ -841,76 +805,27 @@ class Qwen2ForCausalLM(Qwen2PreTrainedModel,  MDMGenerationMixin):
         past_key_values: Optional[Union[Cache, List[torch.FloatTensor]]] = None,
         inputs_embeds: Optional[torch.FloatTensor] = None,
         labels: Optional[torch.LongTensor] = None,
-        mask_ratio: Optional[torch.FloatTensor]=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,
         is_causal: bool = True,
-        **kwargs: Unpack[KwargsForCausalLM],
+        **kwargs,
     ) -> Union[Tuple, CausalLMOutputWithPast]:
-        r"""
+        r\"\"\"
         Args:
             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, Qwen2ForCausalLM
-
-        >>> model = Qwen2ForCausalLM.from_pretrained("meta-qwen2/Qwen2-2-7b-hf")
-        >>> tokenizer = AutoTokenizer.from_pretrained("meta-qwen2/Qwen2-2-7b-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."
-        ```"""
+                Labels for computing the masked language modeling loss. Indices should be in `[0, ...,
+                config.vocab_size - 1]`. Tokens with indices set to `-100` are ignored (masked), the loss is only
+                computed for the tokens with labels in `[0, ..., config.vocab_size - 1]`.
+        \"\"\"
         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
 
-        if not get_parallel_state().sp_enabled and labels is not None:
-            # Shift so that tokens < n predict n
-            labels = labels[..., 1:].contiguous()
-            labels = labels.view(-1)
-            if (
-                position_ids is not None
-                and position_ids.size(0) == 1
-                and not (torch.diff(position_ids, dim=-1) >= 0).all()
-            ):
-                position_ids_ = position_ids.flatten()
-                indices_q = torch.arange(position_ids_.size(0), device=position_ids_.device, dtype=torch.int32)
-                cu_seq_lens = torch.cat(
-                    (
-                        indices_q[position_ids_ == 0],
-                        torch.tensor(position_ids_.size(), device=position_ids_.device, dtype=torch.int32),
-                    )
-                )
-                labels[cu_seq_lens[1:-1] - 1] = IGNORE_INDEX
-        if mask_ratio is not None:
-            is_causal = False
-            mask_ratio = mask_ratio[..., 1:].contiguous()
-        # decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn)
         outputs = self.model(
             input_ids=input_ids,
             attention_mask=attention_mask,
@@ -927,64 +842,19 @@ class Qwen2ForCausalLM(Qwen2PreTrainedModel,  MDMGenerationMixin):
         )
 
         hidden_states = outputs[0]
-        # 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
-        hidden_states = hidden_states[:, slice_indices, :]
-
+        logits = self.lm_head(hidden_states)
+        logits = logits.float()
         loss = None
-        logits = None
+        
         if labels is not None:
-            labels = labels.view(-1)  # flatten label
-            if is_liger_kernel_available():
-                if mask_ratio is not None:
-                    loss_fct = LigerFusedLinearCrossEntropyLoss(reduction="none",ignore_index=IGNORE_INDEX)
-                    if not get_parallel_state().sp_enabled:
-                        # Shift so that tokens < n predict n
-                        hidden_states = hidden_states[..., :-1, :].contiguous()
-                    loss = loss_fct(
-                        self.lm_head.weight, 
-                        hidden_states.view(-1, self.config.hidden_size),
-                        labels
-                    )
-                    path_loss = (-loss).exp().detach() * loss
-                    loss = loss + path_loss
-                    loss_mask = labels != IGNORE_INDEX
-                    loss = (loss * loss_mask * (1/mask_ratio)).sum() / (loss_mask.sum() + 1e-8)
-                else:
-                    loss_fct = LigerFusedLinearCrossEntropyLoss(reduction="mean")
-                    if not get_parallel_state().sp_enabled:
-                        # Shift so that tokens < n predict n
-                        hidden_states = hidden_states[..., :-1, :].contiguous()
-                    hidden_states = hidden_states.view(-1, self.config.hidden_size)
-                    loss = loss_fct(self.lm_head.weight, hidden_states, labels)
-            else:
-                if mask_ratio is not None:
-                    loss_fct = torch.nn.CrossEntropyLoss(reduction="none")
-                    logits = self.lm_head(hidden_states)
-                    logits = logits.view(-1, self.vocab_size)
-                    loss = loss_fct(logits, labels.view(-1))
-                    path_loss = (-loss).exp().detach() * loss
-                    loss = loss + path_loss
-                    loss_mask = labels != IGNORE_INDEX
-                    loss = (loss * loss_mask * (1/mask_ratio)).sum() / (loss_mask.sum() + 1e-8)
-                else:
-                    loss_fct = torch.nn.CrossEntropyLoss(reduction="mean")
-                    logits = self.lm_head(hidden_states)
-                    # Upcast to float if we need to compute the loss to avoid potential precision issues
-                    logits = logits.float()
-                    if not get_parallel_state().sp_enabled:
-                        # Shift so that tokens < n predict n
-                        logits = logits[..., :-1, :].contiguous()
-
-                    # Flatten the tokens
-                    logits = logits.view(-1, self.vocab_size)
-                    loss = loss_fct(logits, labels)
-
-            if get_parallel_state().sp_enabled:
-                num_valid_tokens = (labels != IGNORE_INDEX).sum()
-                loss = reduce_sequence_parallel_loss(loss, num_valid_tokens)
-        else:
-            logits = self.lm_head(hidden_states)
+            # Maintained for compatibility with Trainer API, but not essential for pure inference
+            shift_logits = logits[..., :-1, :].contiguous()
+            shift_labels = labels[..., 1:].contiguous()
+            loss_fct = torch.nn.CrossEntropyLoss()
+            shift_logits = shift_logits.view(-1, self.config.vocab_size)
+            shift_labels = shift_labels.view(-1)
+            shift_labels = shift_labels.to(shift_logits.device)
+            loss = loss_fct(shift_logits, shift_labels)
 
         if not return_dict:
             output = (logits,) + outputs[1:]
@@ -997,88 +867,3 @@ class Qwen2ForCausalLM(Qwen2PreTrainedModel,  MDMGenerationMixin):
             hidden_states=outputs.hidden_states,
             attentions=outputs.attentions,
         )
-    
-
-
-
-import torch
-from tqdm import tqdm
-from typing import Callable, Tuple, Any
-
-
-def topk_masking(scores: torch.Tensor, cutoff_len: torch.Tensor, mode: str = "lowest") -> torch.Tensor:
-    """Generate a mask selecting the top-k lowest or highest elements per row."""
-    sorted_scores = scores.sort(dim=-1, descending=(mode == "highest")).values
-    cutoff = sorted_scores.gather(dim=-1, index=cutoff_len)
-    return (scores >= cutoff) if mode == "highest" else (scores < cutoff)
-
-
-def sample_categorical(
-    logits: torch.Tensor, temperature: float = 1.0, noise_scale: float = 1.0
-) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
-    """
-    Sample from a categorical distribution with optional Gumbel noise.
-    Returns sampled tokens, their scores, and the noised logits.
-    """
-    logits = logits.to(torch.float64)
-    if temperature > 0:
-        gumbel_noise = -torch.log(-torch.log(torch.rand_like(logits) + 1e-8) + 1e-8)
-        logits = logits / temperature + noise_scale * gumbel_noise
-    log_probs = logits.log_softmax(dim=-1)
-    scores, tokens = log_probs.max(dim=-1)
-    return tokens, scores.to(logits.dtype), logits.to(logits.dtype)
-
-
-@torch.inference_mode()
-@torch.amp.autocast(device_type="cuda", dtype=torch.float16)
-def p2_sampling(
-    xt: torch.Tensor,
-    model: Any,
-    mask_id: int,
-    num_steps: int,
-    tau: float = 1.0,
-    kappa_fn: Callable[[float], float] = lambda t: t,
-    eta: float = 1.0,
-    **kwargs
-) -> torch.Tensor:
-    """
-    P2 Sampling implementation for discrete diffusion models.
-    Reference: https://arxiv.org/pdf/2502.03540
-    """
-    dt = 1 / num_steps
-    fix_mask = (xt != mask_id)
-
-    for i in tqdm(range(1, num_steps + 1)):
-        t = i * dt
-        kappa_t = kappa_fn(t)
-
-        logits = model(xt).double()
-        last_mask = (xt == mask_id)
-        unmask_t = ~last_mask & ~fix_mask
-
-        x0, score, _ = sample_categorical(logits, temperature=tau)
-        score = score.masked_fill(fix_mask, float("inf"))
-        score[unmask_t] *= eta
-
-        num_to_mask = ((~fix_mask).sum(dim=1, keepdim=True).float() * (1 - kappa_t)).long()
-        to_mask = topk_masking(score, num_to_mask, mode="lowest")
-
-        xt[to_mask] = mask_id
-        mask_2_x0 = last_mask & ~to_mask
-        xt[mask_2_x0] = x0[mask_2_x0]
-
-    xt[xt == mask_id] = x0[xt == mask_id]
-    return xt
-
-
-
-if is_liger_kernel_available():
-    apply_rotary_pos_emb = liger_rotary_pos_emb
-    Qwen2RMSNorm = LigerRMSNorm
-    Qwen2MLP = LigerSwiGLUMLP
-    logger.info_rank0("Apply liger kernel to Qwen2.")
-
-
-ModelClass = Qwen2ForCausalLM
-
-__all__ = ["Qwen2ForCausalLM", "Qwen2Model", "Qwen2PreTrainedModel"]