Merge branch 'main' into main

Former-commit-id: 7be442f37d53a0c6324728fa1fa8e2c84d7f0fa5

src/llamafactory/model/model_utils/attention.py

@@ -1,7 +1,22 @@
+# Copyright 2024 the LlamaFactory 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.
+
 from typing import TYPE_CHECKING
 
+from transformers.utils import is_flash_attn_2_available, is_torch_sdpa_available
+
 from ...extras.logging import get_logger
-from ...extras.packages import is_flash_attn2_available, is_sdpa_available
 
 
 if TYPE_CHECKING:
@@ -13,21 +28,33 @@ if TYPE_CHECKING:
 logger = get_logger(__name__)
 
 
-def configure_attn_implementation(config: "PretrainedConfig", model_args: "ModelArguments") -> None:
+def configure_attn_implementation(
+    config: "PretrainedConfig", model_args: "ModelArguments", is_trainable: bool
+) -> None:
+    if getattr(config, "model_type", None) == "gemma2" and is_trainable:  # gemma2 adopts soft-cap attention
+        if model_args.flash_attn == "auto":
+            logger.warning("Gemma-2 models should use eager attention in training, change `flash_attn` to disabled.")
+            model_args.flash_attn = "disabled"
+        elif model_args.flash_attn != "disabled":
+            logger.warning(
+                "Gemma-2 models should use eager attention in training, but you set `flash_attn: {}`. "
+                "Will proceed at your own risk.".format(model_args.flash_attn)
+            )
+
     if model_args.flash_attn == "auto":
         return
 
-    elif model_args.flash_attn == "off":
+    elif model_args.flash_attn == "disabled":
         requested_attn_implementation = "eager"
 
     elif model_args.flash_attn == "sdpa":
-        if not is_sdpa_available():
+        if not is_torch_sdpa_available():
             logger.warning("torch>=2.1.1 is required for SDPA attention.")
             return
 
         requested_attn_implementation = "sdpa"
     elif model_args.flash_attn == "fa2":
-        if not is_flash_attn2_available():
+        if not is_flash_attn_2_available():
             logger.warning("FlashAttention-2 is not installed.")
             return
 

src/llamafactory/model/model_utils/checkpointing.py

@@ -1,3 +1,21 @@
+# Copyright 2024 HuggingFace Inc. and the LlamaFactory team.
+#
+# This code is inspired by the HuggingFace's Transformers and PEFT library.
+# https://github.com/huggingface/transformers/blob/v4.40.0/src/transformers/modeling_utils.py
+# https://github.com/huggingface/peft/blob/v0.10.0/src/peft/utils/other.py
+#
+# 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 inspect
 from functools import partial
 from types import MethodType
@@ -60,15 +78,12 @@ def _fp32_forward_post_hook(
     return output.to(torch.float32)
 
 
-def prepare_model_for_training(
-    model: "PreTrainedModel", model_args: "ModelArguments", output_layer_name: str = "lm_head"
-) -> None:
+def prepare_model_for_training(model: "PreTrainedModel", model_args: "ModelArguments") -> None:
     r"""
     Includes:
         (1) cast the layernorm in fp32
         (2) make output embedding layer require grads
         (3) add the upcasting of the lm_head in fp32
-    Inspired by: https://github.com/huggingface/peft/blob/v0.7.1/src/peft/utils/other.py#L72
     """
     if model_args.upcast_layernorm:
         logger.info("Upcasting layernorm weights in float32.")
@@ -87,8 +102,8 @@ def prepare_model_for_training(
             setattr(model.config, "use_cache", False)  # turn off when gradient checkpointing is enabled
             logger.info("Gradient checkpointing enabled.")
 
-    if hasattr(model, output_layer_name) and model_args.upcast_lmhead_output:
-        logger.info("Upcasting lm_head outputs in float32.")
-        output_layer = getattr(model, output_layer_name)
+    if model_args.upcast_lmhead_output:
+        output_layer = model.get_output_embeddings()
         if isinstance(output_layer, torch.nn.Linear) and output_layer.weight.dtype != torch.float32:
+            logger.info("Upcasting lm_head outputs in float32.")
             output_layer.register_forward_hook(_fp32_forward_post_hook)

src/llamafactory/model/model_utils/embedding.py

@@ -1,3 +1,17 @@
+# Copyright 2024 the LlamaFactory 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 math
 from contextlib import nullcontext
 from typing import TYPE_CHECKING

src/llamafactory/model/model_utils/longlora.py

@@ -1,3 +1,22 @@
+# Copyright 2024 EleutherAI, HuggingFace Inc., Yukang Chen, and the LlamaFactory team.
+#
+# This code is based on the EleutherAI's GPT-NeoX and the HuggingFace's Transformers libraries.
+# https://github.com/huggingface/transformers/blob/v4.40.0/src/transformers/models/llama/modeling_llama.py
+# This code is also inspired by the original LongLoRA implementation.
+# https://github.com/dvlab-research/LongLoRA/blob/main/llama_attn_replace.py
+#
+# 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 math
 from typing import TYPE_CHECKING, Optional, Tuple
 
@@ -96,7 +115,8 @@ def llama_attention_forward(
             (
                 attn_output[:, :, : self.num_heads // 2],
                 attn_output[:, :, self.num_heads // 2 :].roll(groupsz // 2, dims=1),
-            )
+            ),
+            dim=2,
         )
 
     attn_output = attn_output.reshape(bsz, q_len, self.hidden_size)
@@ -181,11 +201,9 @@ def llama_flash_attention_2_forward(
         query_states, key_states, value_states = shift(query_states), shift(key_states), shift(value_states)
         if attention_mask is not None:
             attention_mask = attention_mask[:, :groupsz].repeat(num_groups, 1)
-    else:
-        groupsz = q_len
 
     attn_output: torch.Tensor = self._flash_attention_forward(
-        query_states, key_states, value_states, attention_mask, groupsz, dropout=dropout_rate
+        query_states, key_states, value_states, attention_mask, query_states.size(1), dropout=dropout_rate
     )
 
     if getattr(self.config, "group_size_ratio", None) and self.training:  # shift back
@@ -194,7 +212,8 @@ def llama_flash_attention_2_forward(
             (
                 attn_output[:, :, : self.num_heads // 2],
                 attn_output[:, :, self.num_heads // 2 :].roll(groupsz // 2, dims=1),
-            )
+            ),
+            dim=2,
         )
 
     attn_output = attn_output.reshape(bsz, q_len, self.hidden_size).contiguous()
@@ -293,7 +312,8 @@ def llama_sdpa_attention_forward(
             (
                 attn_output[:, :, : self.num_heads // 2],
                 attn_output[:, :, self.num_heads // 2 :].roll(groupsz // 2, dims=1),
-            )
+            ),
+            dim=2,
         )
 
     attn_output = attn_output.reshape(bsz, q_len, self.hidden_size)
@@ -303,7 +323,7 @@ def llama_sdpa_attention_forward(
 
 
 def _apply_llama_patch() -> None:
-    require_version("transformers==4.40.2", "To fix: pip install transformers==4.40.2")
+    require_version("transformers==4.41.2", "To fix: pip install transformers==4.41.2")
     LlamaAttention.forward = llama_attention_forward
     LlamaFlashAttention2.forward = llama_flash_attention_2_forward
     LlamaSdpaAttention.forward = llama_sdpa_attention_forward

src/llamafactory/model/model_utils/misc.py

@@ -1,3 +1,17 @@
+# Copyright 2024 the LlamaFactory 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.
+
 from typing import TYPE_CHECKING, List
 
 from ...extras.logging import get_logger

src/llamafactory/model/model_utils/mod.py

@@ -1,3 +1,17 @@
+# Copyright 2024 the LlamaFactory 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.
+
 from typing import TYPE_CHECKING
 
 from ...extras.constants import MOD_SUPPORTED_MODELS

src/llamafactory/model/model_utils/moe.py

@@ -1,5 +1,20 @@
-from typing import TYPE_CHECKING
+# Copyright 2024 the LlamaFactory 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.
 
+from typing import TYPE_CHECKING, Sequence
+
+import torch
 from transformers.integrations import is_deepspeed_zero3_enabled
 from transformers.utils.versions import require_version
 
@@ -10,6 +25,13 @@ if TYPE_CHECKING:
     from ...hparams import ModelArguments
 
 
+def _set_z3_leaf_modules(model: "PreTrainedModel", leaf_modules: Sequence["torch.nn.Module"]) -> None:
+    require_version("deepspeed>=0.13.0", "To fix: pip install deepspeed>=0.13.0")
+    from deepspeed.utils import set_z3_leaf_modules  # type: ignore
+
+    set_z3_leaf_modules(model, leaf_modules)
+
+
 def add_z3_leaf_module(model: "PreTrainedModel") -> None:
     r"""
     Sets module as a leaf module to skip partitioning in deepspeed zero3.
@@ -17,33 +39,30 @@ def add_z3_leaf_module(model: "PreTrainedModel") -> None:
     if not is_deepspeed_zero3_enabled():
         return
 
-    require_version("deepspeed>=0.13.0", "To fix: pip install deepspeed>=0.13.0")
-    from deepspeed.utils import set_z3_leaf_modules  # type: ignore
-
     if getattr(model.config, "model_type", None) == "dbrx":
         from transformers.models.dbrx.modeling_dbrx import DbrxFFN
 
-        set_z3_leaf_modules(model, [DbrxFFN])
+        _set_z3_leaf_modules(model, [DbrxFFN])
 
     if getattr(model.config, "model_type", None) == "jamba":
         from transformers.models.jamba.modeling_jamba import JambaSparseMoeBlock
 
-        set_z3_leaf_modules(model, [JambaSparseMoeBlock])
+        _set_z3_leaf_modules(model, [JambaSparseMoeBlock])
 
     if getattr(model.config, "model_type", None) == "jetmoe":
         from transformers.models.jetmoe.modeling_jetmoe import JetMoeMoA, JetMoeMoE
 
-        set_z3_leaf_modules(model, [JetMoeMoA, JetMoeMoE])
+        _set_z3_leaf_modules(model, [JetMoeMoA, JetMoeMoE])
 
     if getattr(model.config, "model_type", None) == "mixtral":
         from transformers.models.mixtral.modeling_mixtral import MixtralSparseMoeBlock
 
-        set_z3_leaf_modules(model, [MixtralSparseMoeBlock])
+        _set_z3_leaf_modules(model, [MixtralSparseMoeBlock])
 
     if getattr(model.config, "model_type", None) == "qwen2moe":
         from transformers.models.qwen2_moe.modeling_qwen2_moe import Qwen2MoeSparseMoeBlock
 
-        set_z3_leaf_modules(model, [Qwen2MoeSparseMoeBlock])
+        _set_z3_leaf_modules(model, [Qwen2MoeSparseMoeBlock])
 
 
 def configure_moe(config: "PretrainedConfig", model_args: "ModelArguments", is_trainable: bool) -> None:

src/llamafactory/model/model_utils/quantization.py

@@ -1,3 +1,21 @@
+# Copyright 2024 HuggingFace Inc. and the LlamaFactory team.
+#
+# This code is inspired by the HuggingFace's Transformers and Optimum library.
+# https://github.com/huggingface/transformers/blob/v4.41.0/src/transformers/utils/quantization_config.py
+# https://github.com/huggingface/optimum/blob/v1.20.0/optimum/gptq/data.py
+#
+# 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 os
 import random
 from enum import Enum, unique
@@ -5,7 +23,7 @@ from typing import TYPE_CHECKING, Any, Dict, List
 
 import torch
 from datasets import load_dataset
-from transformers import BitsAndBytesConfig, GPTQConfig
+from transformers import BitsAndBytesConfig, EetqConfig, GPTQConfig, HqqConfig
 from transformers.integrations import is_deepspeed_zero3_enabled
 from transformers.modeling_utils import is_fsdp_enabled
 from transformers.utils.versions import require_version
@@ -39,10 +57,9 @@ class QuantizationMethod(str, Enum):
     HQQ = "hqq"
 
 
-def _get_quantization_dataset(tokenizer: "PreTrainedTokenizer", model_args: "ModelArguments") -> List[str]:
+def _get_quantization_dataset(tokenizer: "PreTrainedTokenizer", model_args: "ModelArguments") -> List[Dict[str, Any]]:
     r"""
-    Inspired by: https://github.com/huggingface/optimum/blob/v1.16.0/optimum/gptq/data.py#L133
-    TODO: remove tokenizer.decode() https://github.com/huggingface/optimum/pull/1600
+    Prepares the tokenized dataset to perform AutoGPTQ. Do not use tensor output for JSON serialization.
     """
     if os.path.isfile(model_args.export_quantization_dataset):
         data_path = FILEEXT2TYPE.get(model_args.export_quantization_dataset.split(".")[-1], None)
@@ -51,20 +68,32 @@ def _get_quantization_dataset(tokenizer: "PreTrainedTokenizer", model_args: "Mod
         data_path = model_args.export_quantization_dataset
         data_files = None
 
-    dataset = load_dataset(path=data_path, data_files=data_files, split="train", cache_dir=model_args.cache_dir)
-    maxlen = model_args.export_quantization_maxlen
+    dataset = load_dataset(
+        path=data_path,
+        data_files=data_files,
+        split="train",
+        cache_dir=model_args.cache_dir,
+        token=model_args.hf_hub_token,
+    )
 
     samples = []
+    maxlen = model_args.export_quantization_maxlen
     for _ in range(model_args.export_quantization_nsamples):
+        n_try = 0
         while True:
+            if n_try > 100:
+                raise ValueError("Cannot find satisfying example, considering decrease `export_quantization_maxlen`.")
+
             sample_idx = random.randint(0, len(dataset) - 1)
-            sample: Dict[str, torch.Tensor] = tokenizer(dataset[sample_idx]["text"], return_tensors="pt")
-            if sample["input_ids"].size(1) >= maxlen:
+            sample: Dict[str, "torch.Tensor"] = tokenizer(dataset[sample_idx]["text"], return_tensors="pt")
+            n_try += 1
+            if sample["input_ids"].size(1) > maxlen:
                 break  # TODO: fix large maxlen
 
         word_idx = random.randint(0, sample["input_ids"].size(1) - maxlen - 1)
         input_ids = sample["input_ids"][:, word_idx : word_idx + maxlen]
-        samples.append(tokenizer.decode(input_ids[0].tolist(), skip_special_tokens=True))
+        attention_mask = sample["attention_mask"][:, word_idx : word_idx + maxlen]
+        samples.append({"input_ids": input_ids.tolist(), "attention_mask": attention_mask.tolist()})
 
     return samples
 
@@ -76,14 +105,14 @@ def configure_quantization(
     init_kwargs: Dict[str, Any],
 ) -> None:
     r"""
-    Priority: PTQ-quantized (training) > AutoGPTQ (export) > Bitsandbytes (training)
+    Priority: PTQ-quantized (train/infer) > AutoGPTQ (export) > On-the-fly quantization (train/infer)
     """
     if getattr(config, "quantization_config", None):  # ptq
-        if is_deepspeed_zero3_enabled():
-            raise ValueError("DeepSpeed ZeRO-3 is incompatible with quantized models.")
+        if model_args.quantization_bit is not None:
+            logger.warning("`quantization_bit` will not affect on the PTQ-quantized models.")
 
-        if model_args.quantization_device_map != "auto":
-            init_kwargs["device_map"] = {"": get_current_device()}
+        if is_deepspeed_zero3_enabled() or is_fsdp_enabled():
+            raise ValueError("DeepSpeed ZeRO-3 or FSDP is incompatible with PTQ-quantized models.")
 
         quantization_config: Dict[str, Any] = getattr(config, "quantization_config", None)
         quant_method = quantization_config.get("quant_method", "")
@@ -105,46 +134,72 @@ def configure_quantization(
         logger.info("Loading {}-bit {}-quantized model.".format(quant_bits, quant_method.upper()))
 
     elif model_args.export_quantization_bit is not None:  # auto-gptq
-        require_version("optimum>=1.16.0", "To fix: pip install optimum>=1.16.0")
+        if model_args.export_quantization_bit not in [8, 4, 3, 2]:
+            raise ValueError("AutoGPTQ only accepts 2/3/4/8-bit quantization.")
+
+        require_version("optimum>=1.17.0", "To fix: pip install optimum>=1.17.0")
         require_version("auto_gptq>=0.5.0", "To fix: pip install auto_gptq>=0.5.0")
         from accelerate.utils import get_max_memory
 
         if getattr(config, "model_type", None) == "chatglm":
-            raise ValueError("ChatGLM model is not supported.")
+            raise ValueError("ChatGLM model is not supported yet.")
 
         init_kwargs["quantization_config"] = GPTQConfig(
             bits=model_args.export_quantization_bit,
-            tokenizer=tokenizer,
             dataset=_get_quantization_dataset(tokenizer, model_args),
         )
         init_kwargs["device_map"] = "auto"
         init_kwargs["max_memory"] = get_max_memory()
-        logger.info("Quantizing model to {} bit.".format(model_args.export_quantization_bit))
+        logger.info("Quantizing model to {} bit with AutoGPTQ.".format(model_args.export_quantization_bit))
 
-    elif model_args.quantization_bit is not None:  # bnb
-        if model_args.quantization_bit == 8:
-            require_version("bitsandbytes>=0.37.0", "To fix: pip install bitsandbytes>=0.37.0")
-            init_kwargs["quantization_config"] = BitsAndBytesConfig(load_in_8bit=True)
+    elif model_args.quantization_bit is not None:  # on-the-fly
+        if model_args.quantization_method == QuantizationMethod.BITS_AND_BYTES.value:
+            if model_args.quantization_bit == 8:
+                require_version("bitsandbytes>=0.37.0", "To fix: pip install bitsandbytes>=0.37.0")
+                init_kwargs["quantization_config"] = BitsAndBytesConfig(load_in_8bit=True)
+            elif model_args.quantization_bit == 4:
+                require_version("bitsandbytes>=0.39.0", "To fix: pip install bitsandbytes>=0.39.0")
+                init_kwargs["quantization_config"] = BitsAndBytesConfig(
+                    load_in_4bit=True,
+                    bnb_4bit_compute_dtype=model_args.compute_dtype,
+                    bnb_4bit_use_double_quant=model_args.double_quantization,
+                    bnb_4bit_quant_type=model_args.quantization_type,
+                    bnb_4bit_quant_storage=model_args.compute_dtype,  # crucial for fsdp+qlora
+                )
+            else:
+                raise ValueError("Bitsandbytes only accepts 4-bit or 8-bit quantization.")
 
-        elif model_args.quantization_bit == 4:
-            require_version("bitsandbytes>=0.39.0", "To fix: pip install bitsandbytes>=0.39.0")
-            init_kwargs["quantization_config"] = BitsAndBytesConfig(
-                load_in_4bit=True,
-                bnb_4bit_compute_dtype=model_args.compute_dtype,
-                bnb_4bit_use_double_quant=model_args.double_quantization,
-                bnb_4bit_quant_type=model_args.quantization_type,
-                bnb_4bit_quant_storage=model_args.compute_dtype,  # crucial for fsdp+qlora
-            )
+            # Do not assign device map if:
+            # 1. deepspeed zero3 or fsdp (train)
+            # 2. auto quantization device map (inference)
+            if is_deepspeed_zero3_enabled() or is_fsdp_enabled() or model_args.quantization_device_map == "auto":
+                if model_args.quantization_bit != 4:
+                    raise ValueError("Only 4-bit quantized model can use fsdp+qlora or auto device map.")
 
-        if is_deepspeed_zero3_enabled() or is_fsdp_enabled() or model_args.quantization_device_map == "auto":
-            if model_args.quantization_bit != 4:
-                raise ValueError("Only 4-bit quantized model can use auto device map.")
+                require_version("bitsandbytes>=0.43.0", "To fix: pip install bitsandbytes>=0.43.0")
+            else:
+                init_kwargs["device_map"] = {"": get_current_device()}  # change auto device map for inference
 
-            require_version("transformers>=4.39.0", "To fix: pip install transformers>=4.39.0")
-            require_version("accelerate>=0.28.0", "To fix: pip install accelerate>=0.28.0")
-            require_version("bitsandbytes>=0.43.0", "To fix: pip install bitsandbytes>=0.43.0")
-            init_kwargs["torch_dtype"] = model_args.compute_dtype  # fsdp+qlora requires same dtype
-        else:
-            init_kwargs["device_map"] = {"": get_current_device()}
+            logger.info("Quantizing model to {} bit with bitsandbytes.".format(model_args.quantization_bit))
+        elif model_args.quantization_method == QuantizationMethod.HQQ.value:
+            if model_args.quantization_bit not in [8, 6, 5, 4, 3, 2, 1]:
+                raise ValueError("HQQ only accepts 1/2/3/4/5/6/8-bit quantization.")
 
-        logger.info("Quantizing model to {} bit.".format(model_args.quantization_bit))
+            if is_deepspeed_zero3_enabled() or is_fsdp_enabled():
+                raise ValueError("HQQ quantization is incompatible with DeepSpeed ZeRO-3 or FSDP.")
+
+            require_version("hqq", "To fix: pip install hqq")
+            init_kwargs["quantization_config"] = HqqConfig(
+                nbits=model_args.quantization_bit, quant_zero=False, quant_scale=False, axis=0
+            )  # use ATEN kernel (axis=0) for performance
+            logger.info("Quantizing model to {} bit with HQQ.".format(model_args.quantization_bit))
+        elif model_args.quantization_method == QuantizationMethod.EETQ.value:
+            if model_args.quantization_bit != 8:
+                raise ValueError("EETQ only accepts 8-bit quantization.")
+
+            if is_deepspeed_zero3_enabled() or is_fsdp_enabled():
+                raise ValueError("EETQ quantization is incompatible with DeepSpeed ZeRO-3 or FSDP.")
+
+            require_version("eetq", "To fix: pip install eetq")
+            init_kwargs["quantization_config"] = EetqConfig()
+            logger.info("Quantizing model to {} bit with EETQ.".format(model_args.quantization_bit))

src/llamafactory/model/model_utils/rope.py

@@ -1,3 +1,21 @@
+# Copyright 2024 LMSYS and the LlamaFactory team.
+# Copyright 2023 Rohan Taori, Ishaan Gulrajani, Tianyi Zhang, Yann Dubois, Xuechen Li
+#
+# This code is inspired by the LMSYS's FastChat library.
+# https://github.com/lm-sys/FastChat/blob/v0.2.30/fastchat/train/train.py
+#
+# 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 math
 from typing import TYPE_CHECKING
 
@@ -21,8 +39,8 @@ def configure_rope(config: "PretrainedConfig", model_args: "ModelArguments", is_
         logger.warning("Current model does not support RoPE scaling.")
         return
 
-    if is_trainable:
-        if model_args.rope_scaling == "dynamic":
+    if model_args.model_max_length is not None:
+        if is_trainable and model_args.rope_scaling == "dynamic":
             logger.warning(
                 "Dynamic NTK scaling may not work well with fine-tuning. "
                 "See: https://github.com/huggingface/transformers/pull/24653"

src/llamafactory/model/model_utils/unsloth.py

@@ -1,3 +1,17 @@
+# Copyright 2024 the LlamaFactory 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.
+
 from typing import TYPE_CHECKING, Any, Dict, Optional
 
 from ...extras.logging import get_logger

src/llamafactory/model/model_utils/valuehead.py

@@ -1,3 +1,17 @@
+# Copyright 2024 the LlamaFactory 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.
+
 from typing import TYPE_CHECKING, Dict
 
 import torch

src/llamafactory/model/model_utils/visual.py

@@ -1,3 +1,20 @@
+# Copyright 2024 HuggingFace Inc. and the LlamaFactory team.
+#
+# This code is inspired by the HuggingFace's Transformers library.
+# https://github.com/huggingface/transformers/blob/v4.40.0/src/transformers/models/llava/modeling_llava.py
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
 from typing import TYPE_CHECKING, Tuple
 
 import torch