project-llm-trainer 0.12.3__py3-none-any.whl

llm_trainer/generate_utils.py

@@ -0,0 +1,450 @@
+from typing import Union, Optional, List
+import torch
+from llm_model import VlmModel, KVCache
+from .tools import TrainerTools
+from .utils import (
+    autocast,
+    batch_repeat_image_tok,
+    calc_position_ids
+)
+
+
+def _suppress_warper(logits: torch.Tensor, suppress_tokens: List[int]) -> torch.Tensor:
+    """
+    抑制特殊token输出
+    :param logits:
+    :param suppress_tokens:
+    :return:
+    """
+    suppress_tokens = torch.tensor(suppress_tokens, device=logits.device)
+    vocab_tensor = torch.arange(logits.shape[-1], device=logits.device)
+    suppress_token_mask = torch.isin(vocab_tensor, suppress_tokens)
+    logits = torch.where(suppress_token_mask, -float("inf"), logits)
+
+    return logits
+
+
+def _temperature_warper(logits: torch.Tensor, temperature: float) -> torch.Tensor:
+    """
+    应用temperature
+    :param logits:
+    :param temperature:
+    :return:
+    """
+    logits = logits / temperature
+    return logits
+
+
+def _top_k_warper(logits: torch.Tensor, k: int, device: Union[str, torch.device, int] = None) -> torch.Tensor:
+    """
+    top k采样
+    :param logits:
+    :param k:
+    :param device:
+    :return:
+    """
+    # [batch, k]
+    topk_logits, _ = torch.topk(logits, k=k)
+    # []
+    min_val: torch.Tensor = topk_logits[:, -1]
+    logits = torch.where(logits < min_val.unsqueeze(-1), torch.tensor(-torch.inf).to(device), logits)
+    return logits
+
+
+def _top_p_warper(logits: torch.Tensor, p: float, min_tokens_to_keep: int = 1) -> torch.Tensor:
+    """
+    top p 核采样
+    :param logits:
+    :param p:
+    :param min_tokens_to_keep:
+    :return:
+    """
+    # 正序排列 eg: [0.1, 0.2, 0.3]
+    sorted_logits, sorted_indices = torch.sort(logits, dim=-1, descending=False)
+    # cumsum求和, 每一个元素的值都是与之前元素的求和
+    # 例如：torch.cumsum(torch.tensor([[0.1, 0.2, 0.3]]), dim=-1) 结果是: [0.1, 0.3, 0.6]
+    cumulative_probs = sorted_logits.softmax(dim=-1).cumsum(dim=-1)
+    # 删除累积概率<=1-p的部分, 因为cumulative_probs是正序排列的，所以要用1-p
+    # 例如：
+    #   假设 p=0.9，并且经过排序和计算后，我们有以下的 cumulative_probs
+    #       cumulative_probs = [0.1, 0.3, 0.7, 0.92, 0.98]
+    #       那么 (1 - p) 就是 0.1
+    #       执行 cumulative_probs <= (1 - p) 后，得到的 sorted_indices_to_remove 就是
+    #       sorted_indices_to_remove = [True, False, False, False, False]
+    #       这意味着，累积概率小于等于 0.1 的词（也就是第一个词）应该被移除
+    #   为什么是 (1 - p)？
+    #       这里使用 (1 - p) 的原因是为了方便后续的处理。在实际的代码中，
+    #       通常会将 sorted_indices_to_remove 向右移动一位，并将第一个元素设置为 False。
+    #       这样做是为了保留至少一个词，即使第一个词的概率非常小。
+    #       通过使用 (1 - p)，我们可以直接使用 cumulative_probs 进行比较，而不需要额外的步骤来处理第一个词
+    sorted_indices_to_remove = cumulative_probs <= (1 - p)
+    # 保证至少有min_tokens_to_keep个token保留
+    # 例如:
+    #   sorted_indices_to_remove=[True, True, True]，min_tokens_to_keep=1时
+    #   该操作后sorted_indices_to_remove=[True, True, False]
+    sorted_indices_to_remove[..., -min_tokens_to_keep:] = 0
+    # 下面步骤是将排序后确定要删除的元素index映射回非排序的元素index
+    #  scatter函数根据 index 中提供的索引，将 src 中的值复制到 tensor 中。
+    # 举例说明, 假设我们有一个 batch，词汇表大小为 5，并且有以下数据
+    #   sorted_indices = [[2, 0, 4, 1, 3]] (排序后的索引)
+    #   sorted_indices_to_remove = [[False, True, False, True, False]] (排序后的移除掩码)
+    # 执行sorted_indices_to_remove.scatter(1, sorted_indices, sorted_indices_to_remove) 后，得到的 indices_to_remove 将是
+    #   indices_to_remove = [[True, True, False, False, False]]
+    indices_to_remove = sorted_indices_to_remove.scatter(1, index=sorted_indices, src=sorted_indices_to_remove)
+
+    # 将需要移除的元素的值设置为-inf
+    scores_processed = logits.masked_fill_(indices_to_remove, -float("Inf"))
+
+    return scores_processed
+
+
+def _generate(
+        model: torch.nn.Module,
+        *,
+        tokens: torch.Tensor,
+        max_new_tokens: int,
+        temperature: Optional[float],
+        k: Optional[int],
+        p: Optional[float],
+        pixel_values: Optional[torch.Tensor] = None,
+        tokens_per_image: int = -1,
+        suppress_tokens: Optional[List[int]] = None,
+        device: Union[str, torch.device, int]
+):
+    """
+    :param model:
+    :param tokens:
+    :param max_new_tokens:
+    :param temperature: 设置None不不生效temperature
+    :param k: top k参数，设置为None或者0不生效topk
+    :param p: top p参数，设置为None不生效top p
+    :param suppress_tokens: 要抑制的tokens
+    :param device:
+
+    如果内容质量底，需要减小temperature、k、p
+    如果temperature很大但内容单一，需要增大k、p
+    """
+    use_kv_cache = True
+
+    # 确保输入维度是 [Batch, Seq]
+    if tokens.dim() == 1:
+        tokens = tokens.unsqueeze(0)
+
+    attention_mask = torch.ones_like(tokens, device=device, dtype=torch.long)
+
+    if isinstance(model, VlmModel):
+        tokens = batch_repeat_image_tok(tokens, tokens_per_image)
+
+    if pixel_values is not None:
+        pixel_values = pixel_values.to(device)
+
+    kv_cache: Optional[KVCache] = None
+    generate_tokens = tokens.clone()
+
+    with torch.inference_mode():
+        for _ in range(max_new_tokens):
+            t = tokens
+            with autocast(TrainerTools().parallel.device_type):
+                result = model(
+                    t,
+                    attention_mask=attention_mask,
+                    past_key_values=kv_cache,
+                    use_cache=use_kv_cache,
+                    pixel_values=pixel_values
+                )
+
+                # logits (batch, seq_len, vocab_size)
+                logits = result['logits']
+                kv_cache = result['past_key_values']
+
+            # (batch, vocab_size)
+            logits = logits[:, -1, :]
+
+            # 抑制特殊token输出
+            if suppress_tokens and len(suppress_tokens) != 0:
+                logits = _suppress_warper(logits, suppress_tokens)
+
+            multinomial = False
+            if temperature and temperature > 0:
+                multinomial = True
+                logits = _temperature_warper(logits, temperature)
+
+            if k and k != 0:
+                logits = _top_k_warper(logits, k, device)
+
+            if p and 0 < p <= 1:
+                logits = _top_p_warper(logits, p)
+
+            if multinomial:
+                prob = logits.softmax(dim=-1)
+                # 返回下标
+                next_token = torch.multinomial(prob, num_samples=1)
+            else:
+                # 返回下标
+                next_token = logits.argmax(dim=-1, keepdim=True)
+
+            # token, is_full_result
+            yield next_token, False
+
+            if use_kv_cache:
+                tokens = next_token
+                generate_tokens = torch.cat((generate_tokens, next_token), dim=-1)
+            else:
+                tokens = torch.cat((tokens, next_token), dim=-1)
+
+            # [关键修复] 更新 mask：追加 1，让 Position ID 继续增长
+            new_mask_bit = torch.ones((tokens.shape[0], 1), device=device, dtype=attention_mask.dtype)
+            attention_mask = torch.cat((attention_mask, new_mask_bit), dim=-1)
+
+            if next_token.item() == TrainerTools().tokenizer.end:
+                break
+
+    # token, is_full_result
+    yield tokens if not use_kv_cache else generate_tokens, True
+
+
+def _streaming_generate(
+        model: torch.nn.Module,
+        *,
+        prompt: Union[str, torch.Tensor],
+        max_new_tokens: int,
+        temperature: Optional[float] = 1.0,
+        k: Optional[int] = None,
+        p: Optional[float] = None,
+        pixel_values: Optional[torch.Tensor] = None,
+        tokens_per_image: int = -1,
+        suppress_tokens: Optional[List[int]] = None,
+        device: Union[str, torch.device, int] = None
+):
+    device = TrainerTools().parallel.device if not device else device
+
+    if isinstance(prompt, torch.Tensor):
+        encoded_tokens = prompt.to(device)
+    else:
+        encoded_tokens = TrainerTools().tokenizer.encode(prompt, unsqueeze=True, covert_tensor=True).to(device)
+
+    generate_text_iterator = _generate(
+        model=model,
+        tokens=encoded_tokens,
+        max_new_tokens=max_new_tokens,
+        temperature=temperature,
+        k=k,
+        p=p,
+        pixel_values=pixel_values,
+        tokens_per_image=tokens_per_image,
+        suppress_tokens=suppress_tokens,
+        device=device
+    )
+
+    for (token, is_full_result) in generate_text_iterator:
+        yield token, is_full_result
+
+
+def streaming_generate(
+        model: torch.nn.Module,
+        *,
+        prompt: Union[str, torch.Tensor],
+        max_new_tokens: int,
+        temperature: Optional[float] = 1.0,
+        k: Optional[int] = None,
+        p: Optional[float] = None,
+        pixel_values: Optional[torch.Tensor] = None,
+        tokens_per_image: int = -1,
+        suppress_tokens: Optional[List[int]] = None,
+        device: Union[str, torch.device, int] = None,
+        return_token: bool = False
+):
+    text_iterator = _streaming_generate(
+        model=model,
+        prompt=prompt,
+        max_new_tokens=max_new_tokens,
+        temperature=temperature,
+        k=k,
+        p=p,
+        pixel_values=pixel_values,
+        tokens_per_image=tokens_per_image,
+        suppress_tokens=suppress_tokens,
+        device=device
+    )
+
+    for (token, is_full_result) in text_iterator:
+        if not is_full_result:
+            if return_token:
+                yield token.squeeze(0)
+            else:
+                yield TrainerTools().tokenizer.decode(token.squeeze(0))
+
+
+def generate(
+        model: torch.nn.Module,
+        *,
+        prompt: Union[str, torch.Tensor],
+        max_new_tokens: int,
+        temperature: Optional[float] = 1.0,
+        k: Optional[int] = None,
+        p: Optional[float] = None,
+        pixel_values: Optional[torch.Tensor] = None,
+        tokens_per_image: int = -1,
+        suppress_tokens: Optional[List[int]] = None,
+        device: Union[str, torch.device, int] = None,
+        return_token: bool = False
+):
+    text_iterator = _streaming_generate(
+        model=model,
+        prompt=prompt,
+        max_new_tokens=max_new_tokens,
+        temperature=temperature,
+        k=k,
+        p=p,
+        suppress_tokens=suppress_tokens,
+        pixel_values=pixel_values,
+        tokens_per_image=tokens_per_image,
+        device=device
+    )
+
+    for (token, is_full_result) in text_iterator:
+        if is_full_result:
+            if return_token:
+                return token.squeeze(0)
+            else:
+                return TrainerTools().tokenizer.decode(token.squeeze(0))
+
+    return None
+
+
+def batch_generate(
+        model: torch.nn.Module,
+        *,
+        tokens: torch.Tensor,
+        attention_mask: torch.Tensor,
+        max_new_tokens: int,
+        temperature: Optional[float] = None,
+        k: Optional[int] = None,
+        p: Optional[float] = None,
+        pixel_values: Optional[torch.Tensor] = None,
+        tokens_per_image: int = -1,
+        suppress_tokens: Optional[List[int]] = None,
+        device: Union[str, torch.device, int],
+        return_logits: bool = True
+):
+    use_kv_cache = True
+    end_token = TrainerTools().tokenizer.end
+    pad_token_id = TrainerTools().tokenizer.pad
+
+    if isinstance(model, VlmModel):
+        tokens = batch_repeat_image_tok(tokens, tokens_per_image)
+
+    if pixel_values is not None:
+        pixel_values = pixel_values.to(device)
+
+    orig_tokens = tokens.clone()
+    full_attention_mask = attention_mask.clone()
+
+    # 初始化 position_ids，处理 left padding
+    position_ids = calc_position_ids(full_attention_mask)
+
+    kv_cache: Optional[KVCache] = None
+    batch_size = tokens.shape[0]
+
+    # 预分配最大长度，避免循环中 cat 造成内存碎片
+    generated_tokens_buffer = torch.full(
+        (batch_size, max_new_tokens),
+        pad_token_id,
+        dtype=torch.long,
+        device=device
+    )
+
+    done = torch.zeros(batch_size, dtype=torch.bool, device=device)
+    current_tokens = tokens
+
+    padded_logits = None
+    actual_gen_len = 0
+
+    pad_token_tensor = torch.tensor(pad_token_id, device=device, dtype=torch.long)
+
+    with torch.inference_mode():
+        for i in range(max_new_tokens):
+            if done.all():
+                break
+
+            actual_gen_len = i + 1
+
+            if current_tokens.dtype != torch.long:
+                current_tokens = current_tokens.long()
+
+            if kv_cache is None:
+                current_position_ids = position_ids
+            else:
+                # 下一个位置ID基于当前mask序列的最后一个有效位置
+                # 如果kv_cache有效，当前token是上一步生成的，位置是前一个位置+1
+                current_position_ids = position_ids[:, -1:] + 1
+                position_ids = torch.cat((position_ids, current_position_ids), dim=-1)
+
+            with autocast(TrainerTools().parallel.device_type):
+                result = model(
+                    current_tokens,
+                    attention_mask=full_attention_mask,
+                    position_ids=current_position_ids,
+                    past_key_values=kv_cache,
+                    use_cache=use_kv_cache,
+                    pixel_values=pixel_values
+                )
+                logits = result['logits']
+                kv_cache = result['past_key_values']
+
+            logits = logits[:, -1, :]
+
+            if return_logits:
+                if padded_logits is None:
+                    vocab_size = logits.shape[-1]
+                    padded_logits = torch.zeros(
+                        (batch_size, max_new_tokens, vocab_size),
+                        dtype=logits.dtype,
+                        device=device
+                    )
+                padded_logits[:, i, :] = logits
+
+            if suppress_tokens:
+                logits = _suppress_warper(logits, suppress_tokens)
+
+            multinomial = False
+            if temperature and temperature > 0:
+                multinomial = True
+                logits = _temperature_warper(logits, temperature)
+            if k and k != 0:
+                logits = _top_k_warper(logits, k, device)
+            if p and 0 < p <= 1:
+                logits = _top_p_warper(logits, p)
+
+            if multinomial:
+                prob = logits.softmax(dim=-1)
+                next_token_active = torch.multinomial(prob, num_samples=1)
+            else:
+                next_token_active = logits.argmax(dim=-1, keepdim=True)
+
+            next_token = torch.where(
+                done.unsqueeze(1),
+                pad_token_tensor,
+                next_token_active
+            )
+
+            generated_tokens_buffer[:, i] = next_token.squeeze(-1)
+
+            new_done = (next_token.squeeze(-1) == end_token)
+            done = done | new_done
+
+            current_tokens = next_token
+
+            new_mask = (~done).long().to(full_attention_mask.dtype)
+            full_attention_mask = torch.cat((full_attention_mask, new_mask.unsqueeze(-1)), dim=-1)
+
+    final_generated_tokens = generated_tokens_buffer[:, :actual_gen_len]
+
+    if padded_logits is not None:
+        final_padded_logits = padded_logits[:, :actual_gen_len, :]
+    else:
+        final_padded_logits = None
+
+    final_full_sequences = torch.cat((orig_tokens, final_generated_tokens), dim=1)
+
+    return final_full_sequences, final_padded_logits