[Feature] Support LLaMA-3 CPT and ST (#5619)

* support LLaMA-3

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Run pre-commit

---------

Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>

applications/Colossal-LLaMA/colossal_llama/utils/__init__.py

@@ -0,0 +1,2 @@
+#!/usr/bin/env python3
+# -*- coding: utf-8 -*-

applications/Colossal-LLaMA/colossal_llama/utils/ckpt_io.py

@@ -0,0 +1,88 @@
+#!/usr/bin/env python3
+# -*- coding: utf-8 -*-
+
+"""
+Helper functions for IO
+"""
+
+import json
+import os
+from typing import Any, Dict, Tuple, Union
+
+import torch
+from torch.optim.lr_scheduler import _LRScheduler
+from torch.optim.optimizer import Optimizer
+
+from colossalai.booster import Booster
+from colossalai.cluster import DistCoordinator
+
+
+def load_json(file_path: Union[str, os.PathLike]) -> Dict[str, Any]:
+    """
+    Load file in JSON format
+    """
+    with open(file=file_path, mode="r", encoding="utf-8") as fp:
+        return json.load(fp)
+
+
+def save_json(data: Dict[str, Any], file_path: Union[str, os.PathLike]) -> None:
+    """
+    Save as JSON format
+    """
+    with open(file=file_path, mode="w", encoding="utf-8") as fp:
+        json.dump(data, fp=fp, ensure_ascii=False, indent=4)
+
+
+def save_checkpoint(
+    save_dir: Union[str, os.PathLike],
+    booster: Booster,
+    model: torch.nn.Module,
+    optimizer: Optimizer,
+    lr_scheduler: _LRScheduler,
+    epoch: int,
+    step: int,
+    batch_size: int,
+    coordinator: DistCoordinator,
+) -> None:
+    """
+    Save model checkpoint, optimizer, LR scheduler and intermedidate running states.
+    """
+
+    save_dir = os.path.join(save_dir, f"epoch-{epoch}_step-{step}")
+    os.makedirs(os.path.join(save_dir, "modeling"), exist_ok=True)
+
+    booster.save_model(model, os.path.join(save_dir, "modeling"), shard=True)
+
+    booster.save_optimizer(optimizer, os.path.join(save_dir, "optimizer"), shard=True)
+    booster.save_lr_scheduler(lr_scheduler, os.path.join(save_dir, "lr_scheduler"))
+    running_states = {
+        "epoch": epoch,
+        "step": step,
+        "sample_start_index": step * batch_size,
+    }
+    if coordinator.is_master():
+        save_json(running_states, os.path.join(save_dir, "running_states.json"))
+
+
+def load_checkpoint(
+    load_dir: Union[str, os.PathLike],
+    booster: Booster,
+    model: torch.nn.Module,
+    optimizer: Optimizer,
+    lr_scheduler: _LRScheduler,
+) -> Tuple[int, int, int]:
+    """
+    Load model checkpoint, optimizer, LR scheduler and intermedidate running states.
+    """
+
+    # Update booster params states.
+    booster.load_model(model=model, checkpoint=os.path.join(load_dir, "modeling"))
+    booster.load_optimizer(optimizer=optimizer, checkpoint=os.path.join(load_dir, "optimizer"))
+    booster.load_lr_scheduler(lr_scheduler=lr_scheduler, checkpoint=os.path.join(load_dir, "lr_scheduler"))
+
+    running_states = load_json(file_path=os.path.join(load_dir, "running_states.json"))
+    return (
+        running_states["epoch"],
+        running_states["step"],
+        running_states["sample_start_index"],
+    )

applications/Colossal-LLaMA/colossal_llama/utils/flash_attention_patch.py

@@ -0,0 +1,352 @@
+#!/usr/bin/env python3
+# -*- coding: utf-8 -*-
+
+import math
+from types import MethodType
+from typing import Optional, Tuple
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from einops import rearrange
+from transformers.models.llama.configuration_llama import LlamaConfig
+from transformers.models.llama.modeling_llama import (
+    LlamaAttention,
+    LlamaForCausalLM,
+    LlamaModel,
+    LlamaRMSNorm,
+    apply_rotary_pos_emb,
+    repeat_kv,
+)
+
+from colossalai.accelerator import get_accelerator
+from colossalai.logging import get_dist_logger
+
+logger = get_dist_logger()
+
+if get_accelerator().name == "cuda":
+    from flash_attn.bert_padding import pad_input, unpad_input
+    from flash_attn.flash_attn_interface import flash_attn_func, flash_attn_varlen_kvpacked_func
+    from flash_attn.ops.rms_norm import rms_norm
+
+    def _prepare_decoder_attention_mask(
+        self: LlamaModel,
+        attention_mask: torch.BoolTensor,
+        input_shape: torch.Size,
+        inputs_embeds: torch.Tensor,
+        past_key_values_length: int,
+    ) -> Optional[torch.Tensor]:
+        """
+        Decoder attetion mask
+        """
+        if past_key_values_length > 0 and attention_mask is not None:
+            attention_mask = torch.cat(
+                tensors=(
+                    torch.full(
+                        size=(input_shape[0], past_key_values_length),
+                        fill_value=True,
+                        dtype=attention_mask.dtype,
+                        device=attention_mask.device,
+                    ),
+                    attention_mask,
+                ),
+                dim=-1,
+            )  # (bsz, past_key_values_length + q_len)
+        if attention_mask is not None and torch.all(attention_mask):
+            return None  # Faster
+        return attention_mask
+
+    def attention_forward(
+        self: LlamaAttention,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_value: Optional[Tuple[torch.Tensor]] = None,
+        output_attentions: bool = False,
+        use_cache: bool = False,
+        **kwargs,
+    ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
+        """
+        Re-define LLaMA-2 `LlamaAttention` forward method using flash-attention.
+        """
+        if output_attentions:
+            logger.warning(
+                "Argument `output_attentions` is not supported for flash-attention patched `LlamaAttention`, "
+                "return `None` instead."
+            )
+
+        bsz, q_len, _ = hidden_states.size()
+
+        if self.config.pretraining_tp > 1:
+            q_slicing, kv_slicing = (
+                dim // self.config.pretraining_tp
+                for dim in (
+                    self.num_heads * self.head_dim,
+                    self.num_key_value_heads * self.head_dim,
+                )
+            )  # `Tuple[int, int]`
+            q_slices, k_slices, v_slices = (
+                proj.weight.split(slicing, dim=0)
+                for proj, slicing in (
+                    (self.q_proj, q_slicing),
+                    (self.k_proj, kv_slicing),
+                    (self.v_proj, kv_slicing),
+                )
+            )  # Tuple[Tuple[torch.Tensor], Tuple[torch.Tensor], Tuple[torch.Tensor]]
+            q, k, v = (
+                torch.cat(
+                    [F.linear(hidden_states, slices[i]) for i in range(self.config.pretraining_tp)],
+                    dim=-1,
+                )
+                for slices in (q_slices, k_slices, v_slices)
+            )
+            # `Tuple[torch.Tensor, torch.Tensor, torch.Tensor]` of shape:
+            # (bsz, q_len, num_heads * head_dim),
+            # (bsz, q_len, num_key_value_heads * head_dim),
+            # (bsz, q_len, num_key_value_heads * head_dim)
+        else:
+            q, k, v = (proj(hidden_states) for proj in (self.q_proj, self.k_proj, self.v_proj))
+            # `Tuple[torch.Tensor, torch.Tensor, torch.Tensor]` of shape:
+            # (bsz, q_len, num_heads * head_dim),
+            # (bsz, q_len, num_key_value_heads * head_dim),
+            # (bsz, q_len, num_key_value_heads * head_dim)
+
+        # (bsz, q_len, num_heads * head_dim) -> (bsz, num_heads, q_len, head_dim);
+        # (bsz, q_len, num_key_value_heads * head_dim) -> (bsz, num_key_value_heads, q_len, head_dim);
+        # (bsz, q_len, num_key_value_heads * head_dim) -> (bsz, num_key_value_heads, q_len, head_dim)
+        q, k, v = (
+            states.view(bsz, q_len, num_heads, self.head_dim).transpose(1, 2)
+            for states, num_heads in (
+                (q, self.num_heads),
+                (k, self.num_key_value_heads),
+                (v, self.num_key_value_heads),
+            )
+        )
+        kv_len = k.shape[-2]  # initially, `kv_len` == `q_len`
+        past_kv_len = 0
+        if past_key_value is not None:
+            # if `past_key_value` is not None, `kv_len` > `q_len`.
+            past_kv_len = past_key_value[0].shape[-2]
+            kv_len += past_kv_len
+
+        # two `torch.Tensor` objs of shape (1, 1, kv_len, head_dim)
+        cos, sin = self.rotary_emb(v, seq_len=kv_len)
+        # (bsz, num_heads, q_len, head_dim), (bsz, num_key_value_heads, q_len, head_dim)
+        q, k = apply_rotary_pos_emb(q=q, k=k, cos=cos, sin=sin, position_ids=position_ids)
+        if past_key_value is not None:
+            # reuse k, v, self_attention
+            k = torch.cat([past_key_value[0], k], dim=2)
+            v = torch.cat([past_key_value[1], v], dim=2)
+
+        past_key_value = (k, v) if use_cache else None
+
+        # repeat k/v heads if n_kv_heads < n_heads
+        k = repeat_kv(hidden_states=k, n_rep=self.num_key_value_groups)
+        # (bsz, num_key_value_heads, q_len, head_dim) -> (bsz, num_heads, q_len, head_dim)
+        v = repeat_kv(hidden_states=v, n_rep=self.num_key_value_groups)
+        # (bsz, num_key_value_heads, q_len, head_dim) -> (bsz, num_heads, q_len, head_dim)
+
+        key_padding_mask = attention_mask
+        # (bsz, num_heads, q_len, head_dim) -> (bsz, q_len, num_heads, head_dim)
+        q, k, v = (states.transpose(1, 2) for states in (q, k, v))
+
+        if past_kv_len > 0:
+            q = torch.cat(
+                tensors=(
+                    torch.full(
+                        size=(bsz, past_kv_len, self.num_heads, self.head_dim),
+                        fill_value=0.0,
+                        dtype=q.dtype,
+                        device=q.device,
+                    ),
+                    q,
+                ),
+                dim=1,
+            )  # (bsz, past_kv_len + q_len, num_heads, head_dim)
+
+        if key_padding_mask is None:
+            # (bsz, past_kv_len + q_len, num_heads, head_dim)
+            output = flash_attn_func(q=q, k=k, v=v, dropout_p=0.0, softmax_scale=None, causal=True)  # (bsz, )
+            output = rearrange(
+                output, pattern="... h d -> ... (h d)"
+            )  # (bsz, past_kv_len + q_len, num_heads * head_dim)
+        else:
+            q, indices, cu_q_lens, max_q_len = unpad_input(hidden_states=q, attention_mask=key_padding_mask)
+            kv, _, cu_kv_lens, max_kv_len = unpad_input(
+                hidden_states=torch.stack(tensors=(k, v), dim=2),
+                attention_mask=key_padding_mask,
+            )
+            output_unpad = flash_attn_varlen_kvpacked_func(
+                q=q,
+                kv=kv,
+                cu_seqlens_q=cu_q_lens,
+                cu_seqlens_k=cu_kv_lens,
+                max_seqlen_q=max_q_len,
+                max_seqlen_k=max_kv_len,
+                dropout_p=0.0,
+                softmax_scale=None,
+                causal=True,
+            )
+            output = pad_input(
+                hidden_states=rearrange(output_unpad, pattern="nnz h d -> nnz (h d)"),
+                indices=indices,
+                batch=bsz,
+                seqlen=past_kv_len + q_len,
+            )  # (bsz, past_kv_len + q_len, num_heads * head_dim)
+
+        if past_kv_len > 0:
+            # Strip off the zero query outputs.
+            output = output[:, past_kv_len:, ...]  # (bsz, q_len, num_heads * head_dim)
+        output = self.o_proj(output)  # (bsz, q_len, hidden_size)
+        return output, None, past_key_value
+
+    def rms_norm_forward(self: LlamaRMSNorm, hidden_states: torch.Tensor) -> torch.Tensor:
+        """
+        Formard function for RMS Norm
+        """
+        return rms_norm(x=hidden_states, weight=self.weight, epsilon=self.variance_epsilon)
+
+    def replace_with_flash_attention(model: LlamaForCausalLM) -> None:
+        for name, module in model.named_modules():
+            if isinstance(module, LlamaAttention):
+                module.forward = MethodType(attention_forward, module)
+            if isinstance(module, LlamaModel):
+                module._prepare_decoder_attention_mask = MethodType(_prepare_decoder_attention_mask, module)
+            if isinstance(module, LlamaRMSNorm):
+                module.forward = MethodType(rms_norm_forward, module)
+
+elif get_accelerator().name == "npu":
+    import torch_npu
+
+    class NPULlamaAttention(LlamaAttention):
+        use_flash: bool = True
+
+        def __init__(self, config: LlamaConfig):
+            super().__init__(config)
+            self.setup()
+
+        def setup(self):
+            self._softmax_scale = 1 / math.sqrt(self.head_dim)
+
+        def forward(
+            self,
+            hidden_states: torch.Tensor,
+            attention_mask: Optional[torch.Tensor] = None,
+            position_ids: Optional[torch.LongTensor] = None,
+            past_key_value: Optional[Tuple[torch.Tensor]] = None,
+            output_attentions: bool = False,
+            use_cache: bool = False,
+        ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
+            bsz, q_len, _ = hidden_states.size()
+
+            if self.config.pretraining_tp > 1:
+                key_value_slicing = (self.num_key_value_heads * self.head_dim) // self.config.pretraining_tp
+                query_slices = self.q_proj.weight.split(
+                    (self.num_heads * self.head_dim) // self.config.pretraining_tp, dim=0
+                )
+                key_slices = self.k_proj.weight.split(key_value_slicing, dim=0)
+                value_slices = self.v_proj.weight.split(key_value_slicing, dim=0)
+
+                query_states = [F.linear(hidden_states, query_slices[i]) for i in range(self.config.pretraining_tp)]
+                query_states = torch.cat(query_states, dim=-1)
+
+                key_states = [F.linear(hidden_states, key_slices[i]) for i in range(self.config.pretraining_tp)]
+                key_states = torch.cat(key_states, dim=-1)
+
+                value_states = [F.linear(hidden_states, value_slices[i]) for i in range(self.config.pretraining_tp)]
+                value_states = torch.cat(value_states, dim=-1)
+
+            else:
+                query_states = self.q_proj(hidden_states)
+                key_states = self.k_proj(hidden_states)
+                value_states = self.v_proj(hidden_states)
+
+            query_states = query_states.view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
+            key_states = key_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
+            value_states = value_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
+
+            kv_seq_len = key_states.shape[-2]
+            if past_key_value is not None:
+                kv_seq_len += past_key_value[0].shape[-2]
+            cos, sin = self.rotary_emb(value_states, seq_len=kv_seq_len)
+            query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin, position_ids)
+
+            if past_key_value is not None:
+                # reuse k, v, self_attention
+                key_states = torch.cat([past_key_value[0], key_states], dim=2)
+                value_states = torch.cat([past_key_value[1], value_states], dim=2)
+
+            past_key_value = (key_states, value_states) if use_cache else None
+
+            key_states = repeat_kv(key_states, self.num_key_value_groups)
+            value_states = repeat_kv(value_states, self.num_key_value_groups)
+
+            if not self.use_flash:
+                attn_weights = torch.matmul(query_states, key_states.transpose(2, 3)) / math.sqrt(self.head_dim)
+
+                if attn_weights.size() != (bsz, self.num_heads, q_len, kv_seq_len):
+                    raise ValueError(
+                        f"Attention weights should be of size {(bsz, self.num_heads, q_len, kv_seq_len)}, but is"
+                        f" {attn_weights.size()}"
+                    )
+
+                if attention_mask is not None:
+                    if attention_mask.size() != (bsz, 1, q_len, kv_seq_len):
+                        raise ValueError(
+                            f"Attention mask should be of size {(bsz, 1, q_len, kv_seq_len)}, but is {attention_mask.size()}"
+                        )
+                    attn_weights = attn_weights + attention_mask
+
+                # upcast attention to fp32
+                attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype=torch.float32).to(query_states.dtype)
+                attn_output = torch.matmul(attn_weights, value_states)
+            else:
+                attn_output, *_ = torch_npu.npu_fusion_attention(
+                    query_states,
+                    key_states,
+                    value_states,
+                    self.num_heads,
+                    "BNSD",
+                    atten_mask=attention_mask.bool(),
+                    scale=self._softmax_scale,
+                    padding_mask=None,
+                    pre_tockens=65535,
+                    next_tockens=0,
+                    keep_prob=1.0,
+                    inner_precise=0,
+                )
+
+            if attn_output.size() != (bsz, self.num_heads, q_len, self.head_dim):
+                raise ValueError(
+                    f"`attn_output` should be of size {(bsz, self.num_heads, q_len, self.head_dim)}, but is"
+                    f" {attn_output.size()}"
+                )
+
+            attn_output = attn_output.transpose(1, 2).contiguous()
+            attn_output = attn_output.reshape(bsz, q_len, self.hidden_size)
+
+            if self.config.pretraining_tp > 1:
+                attn_output = attn_output.split(self.hidden_size // self.config.pretraining_tp, dim=2)
+                o_proj_slices = self.o_proj.weight.split(self.hidden_size // self.config.pretraining_tp, dim=1)
+                attn_output = sum(
+                    [F.linear(attn_output[i], o_proj_slices[i]) for i in range(self.config.pretraining_tp)]
+                )
+            else:
+                attn_output = self.o_proj(attn_output)
+
+            if not output_attentions:
+                attn_weights = None
+
+            return attn_output, attn_weights, past_key_value
+
+    class NPURMSNorm(LlamaRMSNorm):
+        def forward(self, hidden_states):
+            return torch_npu.npu_rms_norm(hidden_states, self.weight, epsilon=self.variance_epsilon)[0]
+
+    def replace_with_flash_attention(model: LlamaForCausalLM) -> None:
+        for name, module in model.named_modules():
+            if isinstance(module, LlamaAttention):
+                module.__class__ = NPULlamaAttention
+                module.setup()
+            if isinstance(module, LlamaRMSNorm):
+                module.__class__ = NPURMSNorm

applications/Colossal-LLaMA/colossal_llama/utils/froze.py

@@ -0,0 +1,18 @@
+#!/usr/bin/env python3
+# -*- coding: utf-8 -*-
+
+from transformers.models.llama import LlamaForCausalLM
+
+
+def freeze_non_embeds_parameters(model: LlamaForCausalLM) -> None:
+    """Freeze all parameters except embeddings."""
+    for name, params in model.named_parameters():
+        if "embed_tokens" not in name and "lm_head" not in name:
+            params.requires_grad = False
+        else:
+            params.requires_grad = True
+
+
+def unfreeze_parameters(model: LlamaForCausalLM) -> None:
+    for name, params in model.named_parameters():
+        params.requires_grad = False

applications/Colossal-LLaMA/colossal_llama/utils/neftune_patch.py

@@ -0,0 +1,72 @@
+#    Copyright 2023 The Hugging Face team
+#
+#    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.
+
+import torch
+
+
+def unwrap(model):
+    if hasattr(model, "module"):
+        return model.unwrap()
+    else:
+        return model
+
+
+def neftune_post_forward_hook(module, input, output):
+    """
+    Implements the NEFTune forward pass for the model using forward hooks. Note this works only for torch.nn.Embedding
+    layers. This method is slightly adapted from the original source code that can be found here:
+    https://github.com/neelsjain/NEFTune Simply add it to your model as follows:
+    ```python
+    model = ...
+    model.embed_tokens.neftune_noise_alpha = 0.1
+    model.embed_tokens.register_forward_hook(neftune_post_forward_hook)
+    ```
+    Args:
+        module (`torch.nn.Module`):
+            The embedding module where the hook is attached. Note that you need to set `module.neftune_noise_alpha` to
+            the desired noise alpha value.
+        input (`torch.Tensor`):
+            The input tensor to the model.
+        output (`torch.Tensor`):
+            The output tensor of the model (i.e. the embeddings).
+    """
+    if module.training:
+        dims = torch.tensor(output.size(1) * output.size(2))
+        mag_norm = module.neftune_noise_alpha / torch.sqrt(dims)
+        output = output + torch.zeros_like(output).uniform_(-mag_norm, mag_norm)
+    return output
+
+
+def activate_neftune(model, neftune_noise_alpha=0.1):
+    r"""
+    Activates the neftune as presented in this code: https://github.com/neelsjain/NEFTune and paper:
+    https://arxiv.org/abs/2310.05914
+    """
+    embeddings = unwrap(model).get_input_embeddings()
+
+    embeddings.neftune_noise_alpha = neftune_noise_alpha
+    hook_handle = embeddings.register_forward_hook(neftune_post_forward_hook)
+    neftune_hook_handle = hook_handle
+
+    return model, neftune_hook_handle
+
+
+def deactivate_neftune(model, neftune_hook_handle):
+    """
+    Deactivates the neftune method. Make sure to call `_activate_neftune` first.
+    """
+    embeddings = unwrap(model).get_input_embeddings()
+
+    neftune_hook_handle.remove()
+    del embeddings.neftune_noise_alpha

applications/Colossal-LLaMA/colossal_llama/utils/stream_chat_patch.py

@@ -0,0 +1,252 @@
+from copy import deepcopy
+from typing import Any, Callable, Dict, List, Optional, Tuple
+
+import torch
+from torch import nn
+from transformers import PreTrainedTokenizer
+from transformers.generation.utils import GenerationConfig, LogitsProcessorList, StoppingCriteriaList
+from transformers.utils import logging
+
+logger = logging.get_logger(__name__)
+
+
+def get_prompt_template(
+    input_query: str,
+    history: List[Dict] = None,
+    roles: list = ["", "Human", "Assistant"],
+) -> str:
+    """
+    Generates a prompt template for chat models based on input and history.
+
+    Args:
+        input_query (str): User's current input query.
+        history (List[Dict], optional): List of past conversations, each a dict with 'role' and 'message'.
+        roles (list): Specifies the roles in the conversation, defaults to ["", "Human", "Assistant"].
+
+    Returns:
+        str: A formatted prompt including the input query and history.
+    """
+    prompt = ""
+    if history is None:
+        new_history = []
+    else:
+        new_history = deepcopy(history)
+
+    new_history.append({"role": roles[1], "message": input_query.strip()})
+    new_history.append({"role": roles[2], "message": None})
+
+    for _, item in enumerate(new_history):
+        role = item.get("role")
+        message = item.get("message")
+        if role == roles[0]:
+            prompt += f"<s>{message}\n\n"
+        else:
+            if message:
+                prompt += f"{role}: <s>{message}</s>"
+            else:
+                prompt += f"{role}: <s>"
+    return prompt
+
+
+@torch.inference_mode()
+def streaming_chat(
+    model: Any,
+    tokenizer: PreTrainedTokenizer,
+    input_query: str,
+    history: List[Dict] = None,
+    roles: list = ["", "Human", "Assistant"],
+    past_key_values: Tuple[Tuple[torch.FloatTensor, Any], Any] = None,
+    temperature: float = 0.8,
+    top_p: float = 0.95,
+    top_k: int = 50,
+    do_sample: bool = True,
+    length_penalty: float = 1.2,
+    max_new_tokens: int = 512,
+    logits_processor: LogitsProcessorList = None,
+    return_past_key_values: bool = False,
+    **kwargs,
+):
+    """
+    Streaming chat responses generation with a given model and tokenizer.
+
+    Args:
+        model (Any): The language model to generate responses.
+        tokenizer (PreTrainedTokenizer): Tokenizer compatible with the model, used for encoding inputs and decoding responses.
+        input_query (str): The current user input to respond to.
+        history (List[Dict], optional): A list of past conversations, where each conversation is a dictionary with keys 'role' and 'message'.
+        roles (list): Roles involved in the conversation, defaults to ["", "Human", "Assistant"].
+        past_key_values (Tuple[Tuple[torch.FloatTensor, Any], Any], optional): Past key values for incremental decoding.
+        temperature (float): The temperature value for token sampling, defaults to 0.8.
+        top_p (float): Nucleus sampling probability threshold, defaults to 0.95.
+        top_k (int): Top-K filtering threshold, defaults to 50.
+        do_sample (bool): Whether to sample responses, defaults to True.
+        length_penalty (float): Penalty for response length, defaults to 1.2.
+        max_new_tokens (int): Maximum number of new tokens to generate, defaults to 512.
+        logits_processor (LogitsProcessorList, optional): Custom logits processors, defaults to None.
+        return_past_key_values (bool): Whether to return past key values for further incremental decoding, defaults to False.
+        **kwargs: Additional keyword arguments for generation.
+
+    Yields:
+        Tuple[str, List[Dict], Optional[Tuple[Tuple[torch.FloatTensor, Any], Any]]]: A tuple containing the generated response, updated history, and
+        optionally the updated past key values if `return_past_key_values` is True.
+
+    Ensures padding is on the left side for the tokenizer.
+    """
+    assert tokenizer.padding_side == "left", "Current generation only supports left padding."
+    if history is None:
+        history = []
+    if logits_processor is None:
+        logits_processor = LogitsProcessorList()
+
+    generation_kwargs = {
+        "temperature": temperature,
+        "top_p": top_p,
+        "top_k": top_k,
+        "do_sample": do_sample,
+        "max_new_tokens": max_new_tokens,
+        "length_penalty": length_penalty,
+        "use_cache": True,
+        **kwargs,
+    }
+
+    prompt_str = get_prompt_template(input_query, history=history, roles=roles)
+
+    eos_token_id = [tokenizer.eos_token_id]
+    inputs = tokenizer(prompt_str, return_tensors="pt").to(model.device)
+    history.append({"role": roles[1], "message": input_query.strip()})
+    history.append({"role": roles[2], "message": None})
+
+    for outputs in stream_generate(
+        model,
+        **inputs,
+        past_key_values=past_key_values,
+        eos_token_id=eos_token_id,
+        return_past_key_values=return_past_key_values,
+        **generation_kwargs,
+    ):
+        if return_past_key_values:
+            outputs, past_key_values = outputs
+
+        outputs = outputs.tolist()[0][len(inputs["input_ids"][0]) : -1]
+        response = tokenizer.decode(outputs)
+
+        history[-1]["message"] = response.strip()
+        if return_past_key_values:
+            yield response, history, past_key_values
+        else:
+            yield response, history
+
+
+@torch.inference_mode()
+def stream_generate(
+    model: Any,
+    input_ids: torch.Tensor,
+    generation_config: Optional[GenerationConfig] = None,
+    logits_processor: Optional[LogitsProcessorList] = None,
+    stopping_criteria: Optional[StoppingCriteriaList] = None,
+    prefix_allowed_tokens_fn: Optional[Callable[[int, torch.Tensor], List[int]]] = None,
+    return_past_key_values: bool = False,
+    **kwargs,
+):
+    """
+    Generates sequences of token ids using the specified model and generation parameters.
+    Adapted from https://huggingface.co/THUDM/chatglm3-6b/blob/main/modeling_chatglm.py
+
+    Args:
+        model (Any): The model used for generating sequences of token ids.
+        input_ids (torch.Tensor): The sequence used as a prompt for the generation or as model inputs to the encoder.
+        generation_config (Optional[GenerationConfig]): The generation configuration to be used as base parametrization for the generation call.
+        logits_processor (Optional[LogitsProcessorList]): Custom logits processors that complement the default logits processors built from arguments
+        and generation config.
+        stopping_criteria (Optional[StoppingCriteriaList]): Custom stopping criteria that complement the default stopping criteria built from arguments
+        and a generation config.
+        prefix_allowed_tokens_fn (Optional[Callable[[int, torch.Tensor], List[int]]]): Function to constrain token generation.
+        return_past_key_values (bool): Whether to return past key values for further incremental decoding, defaults to False.
+        **kwargs: Additional parameters for model generation.
+
+    Yields:
+        torch.Tensor: The generated token IDs, updated after each generation step.
+        Optional[Tuple[Tuple[torch.FloatTensor, Any], Any]]: The past key values, returned if `return_past_key_values` is True, defaults to False.
+    """
+    input_ids_len = input_ids.size(1)
+
+    if generation_config is None:
+        generation_config = model.generation_config
+    generation_config = deepcopy(generation_config)
+    model_kwargs = generation_config.update(**kwargs)
+
+    eos_token_id = generation_config.eos_token_id
+    if isinstance(eos_token_id, int):
+        eos_token_id = [eos_token_id]
+    eos_token_id_tensor = torch.tensor(eos_token_id).to(input_ids.device) if eos_token_id is not None else None
+
+    if generation_config.max_new_tokens is not None:
+        generation_config.max_length = generation_config.max_new_tokens + input_ids_len
+
+    if input_ids_len >= generation_config.max_length:
+        input_ids_string = "decoder_input_ids" if model.config.is_encoder_decoder else "input_ids"
+        logger.warning(
+            f"Input length of {input_ids_string} is {input_ids_len}, but `max_length` is set to"
+            f" {generation_config.max_length}. This can lead to unexpected behavior. You should consider"
+            " increasing `max_new_tokens`."
+        )
+    logits_processor = logits_processor if logits_processor is not None else LogitsProcessorList()
+    stopping_criteria = stopping_criteria if stopping_criteria is not None else StoppingCriteriaList()
+
+    # prepare distribution pre_processing samplers
+    logits_processor = model._get_logits_processor(
+        generation_config=generation_config,
+        input_ids_seq_length=input_ids_len,
+        encoder_input_ids=input_ids,
+        prefix_allowed_tokens_fn=prefix_allowed_tokens_fn,
+        logits_processor=logits_processor,
+    )
+
+    # prepare stopping criteria
+    stopping_criteria = model._get_stopping_criteria(
+        generation_config=generation_config, stopping_criteria=stopping_criteria
+    )
+
+    logits_warper = model._get_logits_warper(generation_config)
+    unfinished_sequences = input_ids.new(input_ids.shape[0]).fill_(1)
+    scores = None
+
+    while True:
+        model_inputs = model.prepare_inputs_for_generation(input_ids, **model_kwargs)
+        # forward pass to get next token
+        outputs = model(
+            **model_inputs,
+            return_dict=True,
+            output_attentions=False,
+            output_hidden_states=False,
+        )
+
+        # NOTE: this is correct only in left padding mode
+        # pre-process distribution
+        next_token_logits = outputs.logits[:, -1, :]
+        next_token_scores = logits_processor(input_ids, next_token_logits)
+        next_token_scores = logits_warper(input_ids, next_token_scores)
+
+        # sample
+        probs = nn.functional.softmax(next_token_scores, dim=-1)
+        if generation_config.do_sample:
+            next_tokens = torch.multinomial(probs, num_samples=1).squeeze(1)
+        else:
+            next_tokens = torch.argmax(probs, dim=-1)
+
+        # update generated ids, model inputs, and length for next step
+        input_ids = torch.cat([input_ids, next_tokens[:, None]], dim=-1)
+        model_kwargs = model._update_model_kwargs_for_generation(
+            outputs, model_kwargs, is_encoder_decoder=model.config.is_encoder_decoder
+        )
+        unfinished_sequences = unfinished_sequences.mul(
+            next_tokens.tile(eos_token_id_tensor.shape[0], 1).ne(eos_token_id_tensor.unsqueeze(1)).prod(dim=0)
+        )
+
+        if return_past_key_values:
+            yield input_ids, outputs.past_key_values
+        else:
+            yield input_ids
+        # stop when each sentence is finished, or if exceed the maximum length
+        if unfinished_sequences.max() == 0 or stopping_criteria(input_ids, scores):
+            break