Upload folder using huggingface_hub

.gitattributes

@@ -33,3 +33,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
+tokenizer.json filter=lfs diff=lfs merge=lfs -text

.ruff_cache/.gitignore

@@ -0,0 +1,2 @@
+# Automatically created by ruff.
+*

.ruff_cache/CACHEDIR.TAG

@@ -0,0 +1 @@
+Signature: 8a477f597d28d172789f06886806bc55

README.md

@@ -0,0 +1,75 @@
+---
+
+library_name: transformers
+tags:
+  - custom_generate
+---
+
+## Description
+Implementation of [Contrastive Search](https://huggingface.co/blog/introducing-csearch), a decoding strategy that jointly optimizes model confidence and a degeneration penalty to produce fluent, coherent, and low-repetition text. At each step, the model considers the top-k candidate tokens and selects the one maximizing:
+
+score(v) = (1 - alpha) * p(v | context) - alpha * max_cosine_similarity(h_v, H_context)
+
+where `alpha` is the trade-off between confidence and the cosine-similarity-based penalty.
+
+This strategy typically:
+
+- Reduces repetition compared to greedy/beam search
+- Preserves semantic coherence better than pure sampling
+
+---
+
+## Base model
+
+- `Qwen/Qwen2.5-0.5B-Instruct` (example)
+
+---
+
+## Model compatibility
+
+- Decoder-only transformer models for causal LM
+
+---
+
+## Additional Arguments
+
+- `top_k` (int): Number of candidate tokens to consider each step (e.g., 4)
+- `penalty_alpha` (float): Weight of the degeneration penalty (e.g., 0.6)
+
+Tips:
+- Larger `top_k` explores more candidates but increases compute
+- `penalty_alpha` in [0.3, 0.8] often works well; `0.0` reduces to greedy
+
+---
+
+## Output Type changes
+
+(none) — returns the same structure as standard `transformers` generation
+
+---
+
+## Example usage
+
+```py
+from transformers import AutoModelForCausalLM, AutoTokenizer, infer_device
+
+device = infer_device()
+
+model_id = "Qwen/Qwen2.5-0.5B-Instruct"
+tokenizer = AutoTokenizer.from_pretrained(model_id)
+model = AutoModelForCausalLM.from_pretrained(model_id, device_map="auto").to(device)
+
+inputs = tokenizer(["DeepMind Company is"], return_tensors="pt").to(device)
+
+# Contrastive search
+gen_out = model.generate(
+    **inputs,
+    custom_generate="contrastive_search",
+    penalty_alpha=0.6,
+    top_k=4,
+    max_new_tokens=128,
+    trust_remote_code=True,
+)
+
+print(tokenizer.batch_decode(gen_out, skip_special_tokens=True))
+```

config.json

@@ -0,0 +1,30 @@
+{
+    "architectures": [
+      "Qwen3ForCausalLM"
+    ],
+    "attention_bias": false,
+    "attention_dropout": 0.0,
+    "bos_token_id": 151643,
+    "eos_token_id": 151645,
+    "head_dim": 128,
+    "hidden_act": "silu",
+    "hidden_size": 1024,
+    "initializer_range": 0.02,
+    "intermediate_size": 3072,
+    "max_position_embeddings": 40960,
+    "max_window_layers": 28,
+    "model_type": "qwen3",
+    "num_attention_heads": 16,
+    "num_hidden_layers": 28,
+    "num_key_value_heads": 8,
+    "rms_norm_eps": 1e-06,
+    "rope_scaling": null,
+    "rope_theta": 1000000,
+    "sliding_window": null,
+    "tie_word_embeddings": true,
+    "torch_dtype": "bfloat16",
+    "transformers_version": "4.56.0",
+    "use_cache": true,
+    "use_sliding_window": false,
+    "vocab_size": 151936
+  }

custom_generate/generate.py

@@ -0,0 +1,608 @@
+from typing import Union, Optional, TYPE_CHECKING
+import torch
+from transformers import LogitsProcessorList, StoppingCriteriaList, GenerationConfig
+from transformers.generation.utils import (
+    GenerateNonBeamOutput,
+    GenerateDecoderOnlyOutput,
+)
+from transformers.cache_utils import Cache, EncoderDecoderCache, DynamicCache
+from transformers.modeling_outputs import CausalLMOutputWithPast, Seq2SeqLMOutput
+from transformers.generation.utils import GenerateEncoderDecoderOutput, ALL_CACHE_NAMES
+from transformers.utils import ModelOutput
+from transformers.configuration_utils import PretrainedConfig
+import torch.nn as nn
+import logging
+
+if TYPE_CHECKING:
+    from transformers.generation.streamers import BaseStreamer
+
+logger = logging.getLogger(__name__)
+
+
+def stack_model_outputs(
+    model_outputs: list[ModelOutput], config: PretrainedConfig
+) -> ModelOutput:
+    """
+    Stack a list of ModelOutput objects (or its subclasses) along the batch_size dimension. The function infers the
+    specific ModelOutput subclass from the list provided.
+    """
+    if not model_outputs:
+        raise ValueError("Input list is empty.")
+
+    # Infer the class from the first object in the list
+    model_output_cls = type(model_outputs[0])
+
+    # Ensure all objects are of the same type
+    if not all(isinstance(obj, model_output_cls) for obj in model_outputs):
+        raise ValueError("All elements in the list should be of the same type.")
+
+    # Helper function to concat tensors or tuples of tensors
+    def _concat(data):
+        """
+        Reverse of `_split` function above.
+        """
+        if any(data is None for data in data):
+            return None
+        if isinstance(data[0], torch.Tensor):
+            return torch.cat(data, dim=0)
+        elif isinstance(data[0], tuple):
+            # If the elements of the tuple are also tuples (e.g., past_key_values in our earlier example)
+            if isinstance(data[0][0], tuple):
+                return tuple(
+                    tuple(
+                        torch.cat([attr[i][j] for attr in data], dim=0)
+                        for j in range(len(data[0][0]))
+                    )
+                    for i in range(len(data[0]))
+                )
+            else:
+                return tuple(
+                    torch.cat([attr[i] for attr in data], dim=0)
+                    for i in range(len(data[0]))
+                )
+        elif isinstance(data[0], (int, float)):
+            # If the elements are integers or floats, return a tensor
+            return torch.tensor(data)
+        else:
+            raise TypeError(f"Unexpected attribute type: {type(data[0])}")
+
+    # Use a dictionary comprehension to gather attributes from all objects and concatenate them
+    concatenated_data = {
+        k: _concat([getattr(model_output, k) for model_output in model_outputs])
+        for k in model_output_cls.__dataclass_fields__
+    }
+
+    # Return a new object of the inferred class with the concatenated attributes
+    return model_output_cls(**concatenated_data)
+
+
+def _ranking_fast(
+    context_hidden: torch.FloatTensor,
+    next_hidden: torch.FloatTensor,
+    next_top_k_probs: torch.FloatTensor,
+    cosine_matrix_mask: torch.LongTensor,
+    alpha: float,
+    beam_width: int,
+) -> torch.FloatTensor:
+    """
+    Reranks the top_k candidates based on a degeneration penalty (cosine similarity with previous tokens), as described
+    in the paper "A Contrastive Framework for Neural Text Generation". Returns the index of the best candidate for each
+    row in the batch.
+    """
+    norm_context_hidden = context_hidden / context_hidden.norm(dim=2, keepdim=True)
+    norm_next_hidden = next_hidden / next_hidden.norm(dim=2, keepdim=True)
+    cosine_matrix = torch.matmul(
+        norm_context_hidden, norm_next_hidden.transpose(1, 2)
+    ).squeeze(-1)  # [B*K, S]
+
+    # Penalize cosine_matrix based on the cosine_matrix_mask (ignore padding positions)
+    # Using a large negative value for masked positions
+    cosine_matrix_mask = cosine_matrix_mask.to(dtype=cosine_matrix.dtype)
+    cosine_matrix_mask = (1 - cosine_matrix_mask) * torch.finfo(cosine_matrix.dtype).min
+    cosine_matrix = cosine_matrix + cosine_matrix_mask
+
+    degeneration_penalty, _ = torch.max(cosine_matrix, dim=-1)  # [B*K]
+    next_top_k_probs = next_top_k_probs.view(-1)  # [B*K]
+    contrastive_score = (1.0 - alpha) * next_top_k_probs - alpha * degeneration_penalty
+    contrastive_score = torch.stack(
+        torch.split(contrastive_score, beam_width)
+    )  # [B, K]
+    _, selected_idx = contrastive_score.max(dim=-1)  # [B]
+    return selected_idx
+
+
+@torch.no_grad()
+def _contrastive_search(
+    model,
+    input_ids: torch.LongTensor,
+    logits_processor: LogitsProcessorList,
+    stopping_criteria: StoppingCriteriaList,
+    generation_config: GenerationConfig,
+    synced_gpus: bool,
+    streamer: Optional["BaseStreamer"],
+    **model_kwargs,
+) -> Union[GenerateNonBeamOutput, torch.LongTensor]:
+    r"""
+    Generates sequences of token ids for models with a language modeling head using **contrastive search** and can
+    be used for text-decoder, text-to-text, speech-to-text, and vision-to-text models.
+
+    Parameters:
+        input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`):
+            The sequence used as a prompt for the generation.
+        logits_processor (`LogitsProcessorList`):
+            An instance of [`LogitsProcessorList`]. List of instances of class derived from [`LogitsProcessor`]
+            used to modify the prediction scores of the language modeling head applied at each generation step.
+        stopping_criteria (`StoppingCriteriaList`):
+            An instance of [`StoppingCriteriaList`]. List of instances of class derived from [`StoppingCriteria`]
+            used to tell if the generation loop should stop.
+        generation_config ([`~generation.GenerationConfig`]):
+            The generation configuration to be used as parametrization of the decoding method.
+        synced_gpus (`bool`):
+            Whether to continue running the while loop until max_length (needed to avoid deadlocking with
+            `FullyShardedDataParallel` and DeepSpeed ZeRO Stage 3).
+        streamer (`BaseStreamer`, *optional*):
+            Streamer object that will be used to stream the generated sequences. Generated tokens are passed
+            through `streamer.put(token_ids)` and the streamer is responsible for any further processing.
+        model_kwargs:
+            Additional model specific keyword arguments will be forwarded to the `forward` function of the model.
+            If model is an encoder-decoder model the kwargs should include `encoder_outputs`.
+
+    Return:
+        [`~generation.GenerateDecoderOnlyOutput`], [`~generation.GenerateEncoderDecoderOutput`]
+        or `torch.LongTensor`: A `torch.LongTensor` containing the generated tokens (default behaviour) or a
+        [`~generation.GenerateDecoderOnlyOutput`] if `model.config.is_encoder_decoder=False` and
+        `return_dict_in_generate=True` or a [`~generation.GenerateEncoderDecoderOutput`] if
+        `model.config.is_encoder_decoder=True`.
+    """
+    if not model_kwargs["use_cache"]:
+        raise ValueError("Contrastive search requires `use_cache=True`")
+    if model._is_stateful:
+        # Just like assisted generation, we need to be able to rollback to a previous state (see comment above)
+        raise ValueError(
+            f"contrastive search is not supported with stateful models, such as {model.__class__.__name__}"
+        )
+    # init values
+    has_eos_stopping_criteria = any(
+        hasattr(criteria, "eos_token_id") for criteria in stopping_criteria
+    )
+    top_k = generation_config.top_k
+    penalty_alpha = generation_config.penalty_alpha
+    pad_token_id = generation_config._pad_token_tensor
+    output_attentions = generation_config.output_attentions
+    output_hidden_states = generation_config.output_hidden_states
+    output_scores = generation_config.output_scores
+    output_logits = generation_config.output_logits
+    return_dict_in_generate = generation_config.return_dict_in_generate
+    sequential = generation_config.low_memory
+
+    # init attention / hidden states / scores tuples
+    raw_logits = () if (return_dict_in_generate and output_logits) else None
+    scores = () if (return_dict_in_generate and output_scores) else None
+    decoder_attentions = () if (return_dict_in_generate and output_attentions) else None
+    cross_attentions = () if (return_dict_in_generate and output_attentions) else None
+    decoder_hidden_states = (
+        () if (return_dict_in_generate and output_hidden_states) else None
+    )
+
+    # if model is an encoder-decoder, retrieve encoder attention weights and hidden states
+    if return_dict_in_generate and model.config.is_encoder_decoder:
+        encoder_attentions = (
+            model_kwargs["encoder_outputs"].get("attentions")
+            if output_attentions
+            else None
+        )
+        encoder_hidden_states = (
+            model_kwargs["encoder_outputs"].get("hidden_states")
+            if output_hidden_states
+            else None
+        )
+
+    # keep track of which sequences are already finished
+    batch_size, cur_len = input_ids.shape[:2]
+    unfinished_sequences = torch.ones(
+        batch_size, dtype=torch.long, device=input_ids.device
+    )
+    model_kwargs = model._get_initial_cache_position(
+        cur_len, input_ids.device, model_kwargs
+    )
+
+    # Create cosine_matrix_mask based on the attention_mask
+    cosine_matrix_mask = torch.ones_like(input_ids, dtype=torch.long)
+    if model.config.is_encoder_decoder:
+        if (
+            "decoder_attention_mask" in model_kwargs
+            and model_kwargs["decoder_attention_mask"] is not None
+        ):
+            cosine_matrix_mask = model_kwargs["decoder_attention_mask"]
+    else:
+        cosine_matrix_mask = model_kwargs["attention_mask"]
+    cosine_matrix_mask = cosine_matrix_mask.repeat_interleave(top_k, dim=0)
+
+    this_peer_finished = False
+
+    while model._has_unfinished_sequences(
+        this_peer_finished, synced_gpus, device=input_ids.device
+    ):
+        # if the first step in the loop, encode all the prefix and obtain: (1) past_key_values;
+        # (2) last_hidden_states; (3) logit_for_next_step; (4) update model kwargs for the next step
+        if model_kwargs.get("past_key_values") is None or (
+            isinstance(model_kwargs["past_key_values"], (Cache, EncoderDecoderCache))
+            and model_kwargs["past_key_values"].get_seq_length() == 0
+        ):
+            # prepare inputs
+            model_kwargs["use_cache"] = True
+            model_inputs = model.prepare_inputs_for_generation(input_ids, **model_kwargs)
+
+            # encode the given prefix and prepare model inputs; encoder-decoder model process the prefix and save
+            # the `encoder_outputs`
+            outputs = model(
+                **model_inputs,
+                return_dict=True,
+                output_hidden_states=True,
+                output_attentions=output_attentions,
+            )
+
+            # last decoder hidden states will be used to compute the degeneration penalty (cosine similarity with
+            # previous tokens)
+            if model.config.is_encoder_decoder:
+                last_hidden_states = outputs.decoder_hidden_states[-1]
+            else:
+                last_hidden_states = outputs.hidden_states[-1]
+
+            # next logit for contrastive search to select top-k candidate tokens
+            # Copy is needed to avoid keeping a hanging ref to outputs.logits which may be very large for this first iteration
+            # (the clone itmodel is always small)
+            # torch.float32 is needed to retain precision for later logits manipulations
+            logit_for_next_step = outputs.logits[:, -1, :].to(
+                copy=True, dtype=torch.float32, device=input_ids.device
+            )
+
+            model_kwargs = model._update_model_kwargs_for_generation(
+                outputs,
+                model_kwargs,
+                is_encoder_decoder=model.config.is_encoder_decoder,
+            )
+
+            if not sequential:
+                # Expands model inputs top_k times, for batched forward passes (akin to beam search).
+                # input_ids is required for expanding visual inputs in qwen2vl
+                _, model_kwargs = model._expand_inputs_for_generation(
+                    input_ids=input_ids,
+                    expand_size=top_k,
+                    is_encoder_decoder=model.config.is_encoder_decoder,
+                    **model_kwargs,
+                )
+
+            past_key_values = model_kwargs.get("past_key_values")
+            if past_key_values is None:
+                raise ValueError(
+                    f"{model.__class__.__name__} does not support caching and therefore **can't** be used "
+                    "for contrastive search."
+                )
+            elif (
+                not isinstance(past_key_values[0], (tuple, torch.Tensor))
+                or past_key_values[0][0].shape[0] != batch_size
+            ):
+                raise ValueError(
+                    f"{model.__class__.__name__} does not have a standard cache format and therefore **can't** be "
+                    "used for contrastive search without further modifications."
+                )
+
+        # contrastive_search main logic start:
+        # contrastive search decoding consists of two steps: (1) candidate tokens recall; (2) candidate re-rank by
+        # degeneration penalty
+        processed_logit_for_next_step = logits_processor(input_ids, logit_for_next_step)
+        next_probs = nn.functional.softmax(processed_logit_for_next_step, dim=-1)
+
+        top_k_probs, top_k_ids = torch.topk(next_probs, dim=-1, k=top_k)
+
+        # Store scores, attentions and hidden_states when required
+        if return_dict_in_generate:
+            if output_logits:
+                raw_logits += (logit_for_next_step,)
+            if output_scores:
+                scores += (processed_logit_for_next_step,)
+            if output_attentions:
+                decoder_attentions += (
+                    (outputs.decoder_attentions,)
+                    if model.config.is_encoder_decoder
+                    else (outputs.attentions,)
+                )
+                if model.config.is_encoder_decoder:
+                    cross_attentions += (outputs.cross_attentions,)
+
+            if output_hidden_states:
+                decoder_hidden_states += (
+                    (outputs.decoder_hidden_states,)
+                    if model.config.is_encoder_decoder
+                    else (outputs.hidden_states,)
+                )
+
+        # This is needed to properly delete outputs.logits which may be very large for this first iteration
+        # Otherwise a reference to outputs.logits is kept all along until after the next call to model.forward()
+        del outputs
+
+        if not sequential:
+            # Replicates the new past_key_values to match the `top_k` candidates
+            past = model_kwargs["past_key_values"]
+            # If it is a static cache, modify it in-place layer after layer to save memory
+            if isinstance(past, DynamicCache) or (
+                isinstance(past, EncoderDecoderCache)
+                and isinstance(past.model_attention_cache, DynamicCache)
+            ):
+                past.batch_repeat_interleave(top_k)
+            else:
+                new_key_values = []
+                for layer in past:
+                    items = []
+                    # item is either the key or the value matrix
+                    for item in layer:
+                        items.append(item.repeat_interleave(top_k, dim=0))
+                    new_key_values.append(tuple(items))
+
+                past = tuple(new_key_values)
+
+            model_kwargs["past_key_values"] = past
+
+        if sequential:
+            all_outputs = []
+            for i in range(top_k):
+                # compute the candidate tokens by the language model and collect their hidden_states
+                next_model_inputs = model.prepare_inputs_for_generation(
+                    top_k_ids[:, i].view(-1, 1), **model_kwargs
+                )
+
+                outputs = model(
+                    **next_model_inputs,
+                    return_dict=True,
+                    output_hidden_states=True,
+                    output_attentions=output_attentions,
+                )
+                if isinstance(outputs["past_key_values"], DynamicCache) or (
+                    isinstance(outputs["past_key_values"], EncoderDecoderCache)
+                    and isinstance(
+                        outputs["past_key_values"].model_attention_cache, DynamicCache
+                    )
+                ):
+                    # Remove past K-V from output since we don't need to stack later
+                    outputs["past_key_values"] = None
+                    # Remove last token from past K-V since we don't want to append it at this point
+                    model_kwargs["past_key_values"].crop(-1)
+
+                all_outputs.append(outputs)
+            outputs = stack_model_outputs(all_outputs, model.config.get_text_config())
+
+        else:
+            # compute the candidate tokens by the language model and collect their hidden_states
+            # assembles top_k_ids into batch of size k
+            next_model_inputs = model.prepare_inputs_for_generation(
+                top_k_ids.view(-1, 1), **model_kwargs
+            )
+
+            outputs = model(
+                **next_model_inputs,
+                return_dict=True,
+                output_hidden_states=True,
+                output_attentions=output_attentions,
+            )
+
+        # This is essential to avoid having a last reference to the big past K-V and double the necessary memory
+        # in the next loop
+        del next_model_inputs
+
+        # name is different for encoder-decoder and decoder-only models
+        if model.config.is_encoder_decoder:
+            next_hidden = outputs.decoder_hidden_states[-1]
+            full_hidden_states = outputs.decoder_hidden_states
+        else:
+            next_hidden = outputs.hidden_states[-1]
+            full_hidden_states = outputs.hidden_states
+
+        # .float() is needed to retain precision for later logits manipulations
+        logits = outputs.logits[:, -1, :].float()
+        context_hidden = last_hidden_states.repeat_interleave(top_k, dim=0)
+
+        # compute the degeneration penalty and re-rank the candidates based on the degeneration penalty and the
+        # model confidence. Keeping `selected_idx` on CPU enables multi-device contrastive search and doesn't
+        # introduce (noticeable) slowdowns on single-device runs.
+        selected_idx = _ranking_fast(
+            context_hidden,
+            next_hidden,
+            top_k_probs,
+            cosine_matrix_mask,
+            penalty_alpha,
+            top_k,
+        )
+        cosine_matrix_mask = torch.cat(
+            [
+                cosine_matrix_mask,
+                cosine_matrix_mask.new_ones((cosine_matrix_mask.shape[0], 1)),
+            ],
+            dim=-1,
+        )
+        selected_idx = selected_idx.to("cpu")
+
+        # This will be used instead of the previous inneficient torch.stack(torch.split())
+        augmented_idx = torch.tensor(
+            [x + i * top_k for i, x in enumerate(selected_idx)]
+        )
+
+        # prepare for the next step: (1) next token_id; (2) past_key_values; (3) last_hidden_states for computing
+        # the degeneration penalty; (4) logits for selecting next top-k candidates; (5) selected tokens scores
+        # (model confidence minus degeneration penalty); (6) decoder hidden_states
+        next_tokens = top_k_ids[range(len(top_k_ids)), selected_idx]
+        next_hidden = torch.stack(torch.split(next_hidden.squeeze(dim=1), top_k))
+        next_hidden = next_hidden[range(batch_size), selected_idx, :]
+        last_hidden_states = torch.cat(
+            [last_hidden_states, next_hidden.unsqueeze(1)], dim=1
+        )
+
+        next_decoder_hidden_states = ()
+        for layer in full_hidden_states:
+            layer = torch.stack(torch.split(layer, top_k))[
+                range(batch_size), selected_idx, :
+            ]
+            next_decoder_hidden_states += (layer,)
+
+        # generate past_key_values cache of only the selected token
+        if sequential:
+            next_model_input = model.prepare_inputs_for_generation(
+                top_k_ids[:, selected_idx].view(-1, 1), **model_kwargs
+            )
+
+            selected_outputs = model(
+                **next_model_input,
+                return_dict=True,
+                output_hidden_states=False,
+                output_attentions=False,
+            )
+            next_past_key_values = selected_outputs["past_key_values"]
+
+        else:
+            next_past_key_values = None
+            for possible_cache_name in ALL_CACHE_NAMES:
+                next_past_key_values = next_past_key_values or getattr(
+                    outputs, possible_cache_name, None
+                )
+            # Do it in-place layer per layer to save memory
+            if isinstance(next_past_key_values, DynamicCache) or (
+                isinstance(next_past_key_values, EncoderDecoderCache)
+                and isinstance(next_past_key_values.model_attention_cache, DynamicCache)
+            ):
+                next_past_key_values.batch_select_indices(augmented_idx)
+            else:
+                new_key_values = []
+                for layer in next_past_key_values:
+                    items = []
+                    # item is either the key or the value matrix
+                    for item in layer:
+                        items.append(item[augmented_idx, ...])
+                    new_key_values.append(tuple(items))
+
+                next_past_key_values = tuple(new_key_values)
+
+        logit_for_next_step = torch.stack(torch.split(logits, top_k))[
+            range(batch_size), selected_idx, :
+        ]
+        logit_for_next_step = logit_for_next_step.to(input_ids.device)
+
+        # Rebuilds the relevant parts of the model output for the selected token, for use in the next iteration
+        if model.config.is_encoder_decoder:
+            next_step_cross_attentions = ()
+            next_step_decoder_attentions = ()
+            if output_attentions:
+                for layer in outputs.cross_attentions:
+                    layer = torch.stack(torch.split(layer, top_k, dim=0))[
+                        range(batch_size), selected_idx, ...
+                    ]
+                    next_step_cross_attentions += (layer,)
+                for layer in outputs.decoder_attentions:
+                    layer = torch.stack(torch.split(layer, top_k, dim=0))[
+                        range(batch_size), selected_idx, ...
+                    ]
+                    next_step_decoder_attentions += (layer,)
+            outputs = Seq2SeqLMOutput(
+                past_key_values=next_past_key_values,
+                decoder_hidden_states=next_decoder_hidden_states,
+                decoder_attentions=next_step_decoder_attentions or None,
+                cross_attentions=next_step_cross_attentions or None,
+            )
+        else:
+            next_step_attentions = ()
+            if output_attentions:
+                for layer in outputs.attentions:
+                    layer = torch.stack(torch.split(layer, top_k, dim=0))[
+                        range(batch_size), selected_idx, ...
+                    ]
+                    next_step_attentions += (layer,)
+            outputs = CausalLMOutputWithPast(
+                past_key_values=next_past_key_values,
+                hidden_states=next_decoder_hidden_states,
+                attentions=next_step_attentions or None,
+            )
+        # contrastive_search main logic end
+
+        # synced_gpus: don't waste resources running the code we don't need; kwargs must be updated before skipping
+        model_kwargs = model._update_model_kwargs_for_generation(
+            outputs,
+            model_kwargs,
+            is_encoder_decoder=model.config.is_encoder_decoder,
+        )
+        if synced_gpus and this_peer_finished:
+            continue
+
+        # finished sentences should have their next token be a padding token
+        if has_eos_stopping_criteria:
+            next_tokens = next_tokens * unfinished_sequences + pad_token_id * (
+                1 - unfinished_sequences
+            )
+
+        # update generated ids, model inputs, and length for next step
+        input_ids = torch.cat([input_ids, next_tokens[:, None]], dim=-1)
+        if streamer is not None:
+            streamer.put(next_tokens.cpu())
+
+        # stop when each sentence is finished
+        unfinished_sequences = unfinished_sequences & ~stopping_criteria(
+            input_ids, scores
+        )
+        this_peer_finished = unfinished_sequences.max() == 0
+
+    if streamer is not None:
+        streamer.end()
+
+    if return_dict_in_generate:
+        # Contrastive search works by forward looking at the next token, so we need to exclude it from
+        # `past_key_values` to be consistent with the other decoding methods
+        if model_kwargs.get("past_key_values") is not None:
+            if isinstance(model_kwargs["past_key_values"], DynamicCache) or (
+                isinstance(model_kwargs["past_key_values"], EncoderDecoderCache)
+                and isinstance(
+                    model_kwargs["past_key_values"].model_attention_cache, DynamicCache
+                )
+            ):
+                model_kwargs["past_key_values"].crop(-1)
+            else:
+                past_key_values = []
+                for layer in model_kwargs["past_key_values"]:
+                    layer_past_key_values = []
+                    for item in layer:
+                        layer_past_key_values.append(item[..., :-1, :])
+                    past_key_values.append(tuple(layer_past_key_values))
+                model_kwargs["past_key_values"] = tuple(past_key_values)
+
+        if model.config.is_encoder_decoder:
+            return GenerateEncoderDecoderOutput(
+                sequences=input_ids,
+                scores=scores,
+                logits=raw_logits,
+                encoder_attentions=encoder_attentions,
+                encoder_hidden_states=encoder_hidden_states,
+                decoder_attentions=decoder_attentions,
+                cross_attentions=cross_attentions,
+                decoder_hidden_states=decoder_hidden_states,
+                past_key_values=model_kwargs.get("past_key_values"),
+            )
+        else:
+            return GenerateDecoderOnlyOutput(
+                sequences=input_ids,
+                scores=scores,
+                logits=raw_logits,
+                attentions=decoder_attentions,
+                hidden_states=decoder_hidden_states,
+                past_key_values=model_kwargs.get("past_key_values"),
+            )
+    else:
+        return input_ids
+
+
+def generate(model, *args, **kwargs):
+    """Custom generate function for Contrastive Search decoding.
+    Args:
+        model (`PreTrainedModel`):
+            The model to generate from.
+        penalty_alpha (`float`): The alpha value for the degeneration penalty.
+        top_k (`int`): The number of candidates to consider at each step.
+    """
+    generation_outputs = model.generate(*args, custom_generate=_contrastive_search, **kwargs)
+    return generation_outputs

generation_config.json

@@ -0,0 +1,13 @@
+{
+    "bos_token_id": 151643,
+    "do_sample": true,
+    "eos_token_id": [
+        151645,
+        151643
+    ],
+    "pad_token_id": 151643,
+    "temperature": 0.6,
+    "top_k": 20,
+    "top_p": 0.95,
+    "transformers_version": "4.56.0"
+}

model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f47f71177f32bcd101b7573ec9171e6a57f4f4d31148d38e382306f42996874b
+size 1503300328

tokenizer.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:aeb13307a71acd8fe81861d94ad54ab689df773318809eed3cbe794b4492dae4
+size 11422654

tokenizer_config.json

@@ -0,0 +1,239 @@
+{
+  "add_bos_token": false,
+  "add_prefix_space": false,
+  "added_tokens_decoder": {
+    "151643": {
+      "content": "<|endoftext|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151644": {
+      "content": "<|im_start|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151645": {
+      "content": "<|im_end|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151646": {
+      "content": "<|object_ref_start|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151647": {
+      "content": "<|object_ref_end|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151648": {
+      "content": "<|box_start|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151649": {
+      "content": "<|box_end|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151650": {
+      "content": "<|quad_start|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151651": {
+      "content": "<|quad_end|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151652": {
+      "content": "<|vision_start|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151653": {
+      "content": "<|vision_end|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151654": {
+      "content": "<|vision_pad|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151655": {
+      "content": "<|image_pad|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151656": {
+      "content": "<|video_pad|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151657": {
+      "content": "<tool_call>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151658": {
+      "content": "</tool_call>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151659": {
+      "content": "<|fim_prefix|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151660": {
+      "content": "<|fim_middle|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151661": {
+      "content": "<|fim_suffix|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151662": {
+      "content": "<|fim_pad|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151663": {
+      "content": "<|repo_name|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151664": {
+      "content": "<|file_sep|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151665": {
+      "content": "<tool_response>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151666": {
+      "content": "</tool_response>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151667": {
+      "content": "<think>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151668": {
+      "content": "</think>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    }
+  },
+  "additional_special_tokens": [
+    "<|im_start|>",
+    "<|im_end|>",
+    "<|object_ref_start|>",
+    "<|object_ref_end|>",
+    "<|box_start|>",
+    "<|box_end|>",
+    "<|quad_start|>",
+    "<|quad_end|>",
+    "<|vision_start|>",
+    "<|vision_end|>",
+    "<|vision_pad|>",
+    "<|image_pad|>",
+    "<|video_pad|>"
+  ],
+  "bos_token": null,
+  "chat_template": "{%- if tools %}\n    {{- '<|im_start|>system\\n' }}\n    {%- if messages[0].role == 'system' %}\n        {{- messages[0].content + '\\n\\n' }}\n    {%- endif %}\n    {{- \"# Tools\\n\\nYou may call one or more functions to assist with the user query.\\n\\nYou are provided with function signatures within <tools></tools> XML tags:\\n<tools>\" }}\n    {%- for tool in tools %}\n        {{- \"\\n\" }}\n        {{- tool | tojson }}\n    {%- endfor %}\n    {{- \"\\n</tools>\\n\\nFor each function call, return a json object with function name and arguments within <tool_call></tool_call> XML tags:\\n<tool_call>\\n{\\\"name\\\": <function-name>, \\\"arguments\\\": <args-json-object>}\\n</tool_call><|im_end|>\\n\" }}\n{%- else %}\n    {%- if messages[0].role == 'system' %}\n        {{- '<|im_start|>system\\n' + messages[0].content + '<|im_end|>\\n' }}\n    {%- endif %}\n{%- endif %}\n{%- set ns = namespace(multi_step_tool=true, last_query_index=messages|length - 1) %}\n{%- for message in messages[::-1] %}\n    {%- set index = (messages|length - 1) - loop.index0 %}\n    {%- if ns.multi_step_tool and message.role == \"user\" and message.content is string and not(message.content.startswith('<tool_response>') and message.content.endswith('</tool_response>')) %}\n        {%- set ns.multi_step_tool = false %}\n        {%- set ns.last_query_index = index %}\n    {%- endif %}\n{%- endfor %}\n{%- for message in messages %}\n    {%- if message.content is string %}\n        {%- set content = message.content %}\n    {%- else %}\n        {%- set content = '' %}\n    {%- endif %}\n    {%- if (message.role == \"user\") or (message.role == \"system\" and not loop.first) %}\n        {{- '<|im_start|>' + message.role + '\\n' + content + '<|im_end|>' + '\\n' }}\n    {%- elif message.role == \"assistant\" %}\n        {%- set reasoning_content = '' %}\n        {%- if message.reasoning_content is string %}\n            {%- set reasoning_content = message.reasoning_content %}\n        {%- else %}\n            {%- if '</think>' in content %}\n                {%- set reasoning_content = content.split('</think>')[0].rstrip('\\n').split('<think>')[-1].lstrip('\\n') %}\n                {%- set content = content.split('</think>')[-1].lstrip('\\n') %}\n            {%- endif %}\n        {%- endif %}\n        {%- if loop.index0 > ns.last_query_index %}\n            {%- if loop.last or (not loop.last and reasoning_content) %}\n                {{- '<|im_start|>' + message.role + '\\n<think>\\n' + reasoning_content.strip('\\n') + '\\n</think>\\n\\n' + content.lstrip('\\n') }}\n            {%- else %}\n                {{- '<|im_start|>' + message.role + '\\n' + content }}\n            {%- endif %}\n        {%- else %}\n            {{- '<|im_start|>' + message.role + '\\n' + content }}\n        {%- endif %}\n        {%- if message.tool_calls %}\n            {%- for tool_call in message.tool_calls %}\n                {%- if (loop.first and content) or (not loop.first) %}\n                    {{- '\\n' }}\n                {%- endif %}\n                {%- if tool_call.function %}\n                    {%- set tool_call = tool_call.function %}\n                {%- endif %}\n                {{- '<tool_call>\\n{\"name\": \"' }}\n                {{- tool_call.name }}\n                {{- '\", \"arguments\": ' }}\n                {%- if tool_call.arguments is string %}\n                    {{- tool_call.arguments }}\n                {%- else %}\n                    {{- tool_call.arguments | tojson }}\n                {%- endif %}\n                {{- '}\\n</tool_call>' }}\n            {%- endfor %}\n        {%- endif %}\n        {{- '<|im_end|>\\n' }}\n    {%- elif message.role == \"tool\" %}\n        {%- if loop.first or (messages[loop.index0 - 1].role != \"tool\") %}\n            {{- '<|im_start|>user' }}\n        {%- endif %}\n        {{- '\\n<tool_response>\\n' }}\n        {{- content }}\n        {{- '\\n</tool_response>' }}\n        {%- if loop.last or (messages[loop.index0 + 1].role != \"tool\") %}\n            {{- '<|im_end|>\\n' }}\n        {%- endif %}\n    {%- endif %}\n{%- endfor %}\n{%- if add_generation_prompt %}\n    {{- '<|im_start|>assistant\\n' }}\n    {%- if enable_thinking is defined and enable_thinking is false %}\n        {{- '<think>\\n\\n</think>\\n\\n' }}\n    {%- endif %}\n{%- endif %}",
+  "clean_up_tokenization_spaces": false,
+  "eos_token": "<|im_end|>",
+  "errors": "replace",
+  "model_max_length": 131072,
+  "pad_token": "<|endoftext|>",
+  "split_special_tokens": false,
+  "tokenizer_class": "Qwen2Tokenizer",
+  "unk_token": null
+}