QAT (#2590)

QAT and quantization w/torchao

Author: salmanmohammadi

_quarto.yml

@@ -43,6 +43,7 @@ quartodoc:
         - cli.vllm_serve
         - cli.cloud.base
         - cli.cloud.modal_
+        - cli.quantize
     - title: Trainers
       desc: Training implementations
       contents:
@@ -147,6 +148,7 @@ quartodoc:
         - utils.optimizers.adopt
         - utils.data.pretraining
         - utils.data.sft
+        - utils.quantization
     - title: Schemas
       desc: Pydantic data models for Axolotl config
       contents:
@@ -196,7 +198,7 @@ quartodoc:
         - utils.callbacks.lisa
         - utils.callbacks.mlflow_
         - utils.callbacks.comet_
-
+        - utils.callbacks.qat
 website:
   title: "Axolotl"
   description: "We make fine-tuning accessible, scalable, and fun"
@@ -256,6 +258,8 @@ website:
             - docs/lr_groups.qmd
             - docs/lora_optims.qmd
             - docs/dataset_loading.qmd
+            - docs/qat.qmd
+            - docs/quantize.qmd
 
         - section: "Core Concepts"
           contents:

docs/cli.qmd

@@ -209,6 +209,16 @@ axolotl delinearize-llama4 --model path/to/model_dir --output path/to/output_dir
 
 This would be necessary to use with other frameworks. If you have an adapter, merge it with the non-quantized linearized model before delinearizing.
 
+### quantize
+
+Quantizes a model using the quantization configuration specified in your YAML file.
+
+```bash
+axolotl quantize config.yml
+```
+
+See [Quantization](./quantize.qmd) for more details.
+
 
 ## Legacy CLI Usage
 

docs/config.qmd

@@ -65,6 +65,20 @@ bnb_config_kwargs:
   bnb_4bit_quant_type: nf4
   bnb_4bit_use_double_quant: true
 
+# quantization aware training
+qat:
+  activation_dtype: # Optional[str] = "int8". Fake quantization layout to use for activation quantization. Valid options are "int4" and "int8"
+  weight_dtype: # Optional[str] = "int8". Fake quantization layout to use for weight quantization. Valid options are "int4" and "int8"
+  group_size: # Optional[int] = 32. The number of elements in each group for per-group fake quantization
+  fake_quant_after_n_steps: # Optional[int] = None. The number of steps to apply fake quantization after
+
+# post-training quantization
+quantization:
+  weight_dtype: # Optional[str] = "int8". Fake quantization layout to use for weight quantization. Valid options are uintX for X in [1, 2, 3, 4, 5, 6, 7], or int4, or int8
+  activation_dtype: # Optional[str] = "int8". Fake quantization layout to use for activation quantization. Valid options are "int4" and "int8"
+  group_size: # Optional[int] = 32. The number of elements in each group for per-group fake quantization
+  quantize_embedding: # Optional[bool] = False. Whether to quantize the embedding layer.
+
 
 # Whether you are training a 4-bit GPTQ quantized model
 gptq: true

docs/qat.qmd

@@ -0,0 +1,32 @@
+---
+title: "Quantization Aware Training (QAT)"
+back-to-top-navigation: true
+toc: true
+toc-expand: 2
+toc-depth: 4
+---
+
+## Overview
+
+[Quantization Aware Training](https://pytorch.org/blog/introduction-to-quantization-on-pytorch/#quantization-aware-training) (QAT) is a technique for improving the accuracy of models which are quantized
+by applying "fake" quantizations to the model's weights (and optionally, activations) during training. This fake
+quantization allows for the model to adjust for noise introduced by the quantization, so when the model is eventually
+quantized, the accuracy loss is minimized. We use the quantization techniques implemented in [torchao](https://github.com/pytorch/ao) to provide
+support for QAT and post-training quantization (PTQ) in axolotl.
+
+We recommend reviewing the excellent QAT tutorial in the [torchtune library](https://pytorch.org/torchtune/main/tutorials/qat_finetune.html#quantizing-the-qat-model),
+and the QAT documentation in the [torchao library](https://github.com/pytorch/ao/tree/main/torchao/quantization/qat), for more details.
+
+## Configuring QAT in Axolotl
+
+To enable QAT in axolotl, add the following to your configuration file:
+
+```yaml
+qat:
+  activation_dtype: # Optional[str] = "int8". Fake quantization layout to use for activation quantization. Valid options are "int4" and "int8"
+  weight_dtype: # Optional[str] = "int8". Fake quantization layout to use for weight quantization. Valid options are "int4" and "int8"
+  group_size: # Optional[int] = 32. The number of elements in each group for per-group fake quantization
+  fake_quant_after_n_steps: # Optional[int] = None. The number of steps to apply fake quantization after
+```
+
+Once you have finished training, you must quantize your model by using the same quantization configuration which you used to train the model with. You can use the [`quantize` command](./quantize.md) to do this.

docs/quantize.qmd

@@ -0,0 +1,53 @@
+---
+title: "Quantization with torchao"
+back-to-top-navigation: true
+toc: true
+toc-expand: 2
+toc-depth: 4
+---
+
+Quantization is a technique to lower the memory footprint of your model, potentially at the cost of accuracy or model performance. We support quantizing your model using the [torchao](https://github.com/pytorch/ao) library. Quantization is supported for both post-training quantization (PTQ) and quantization-aware training (QAT).
+
+
+::: {.callout-note}
+
+We do not currently support quantization techniques such as GGUF/GPTQ,EXL2 at the moment.
+
+:::
+
+## Configuring Quantization in Axolotl
+
+Quantization is configured using the `quantization` key in your configuration file.
+
+```yaml
+base_model: # The path to the model to quantize.
+quantization:
+  weight_dtype: # Optional[str] = "int8". Fake quantization layout to use for weight quantization. Valid options are uintX for X in [1, 2, 3, 4, 5, 6, 7], or int4, or int8
+  activation_dtype: # Optional[str] = "int8". Fake quantization layout to use for activation quantization. Valid options are "int4" and "int8"
+  group_size: # Optional[int] = 32. The number of elements in each group for per-group fake quantization
+  quantize_embedding: # Optional[bool] = False. Whether to quantize the embedding layer.
+
+output_dir:  # The path to the output directory.
+```
+
+Once quantization is complete, your quantized model will be saved in the `{output_dir}/quantized` directory.
+
+You may also use the `quantize` command to quantize a model which has been trained with [QAT](./qat.md) - you can do this by using the existing QAT configuration file which
+you used to train the model:
+
+```yaml
+# qat.yml
+qat:
+  activation_dtype: int8
+  weight_dtype: int8
+  group_size: 256
+  quantize_embedding: true
+
+output_dir: # The path to the output directory used during training where the final checkpoint has been saved.
+```
+
+```bash
+axolotl quantize qat.yml
+```
+
+This ensures that an identical quantization configuration is used to quantize the model as was used to train it.

examples/llama-3/3b-qat-fsdp2.yaml

@@ -0,0 +1,79 @@
+base_model: meta-llama/Llama-3.2-3B
+# Automatically upload checkpoint and final model to HF
+# hub_model_id: username/custom_model_name
+
+load_in_8bit: false
+load_in_4bit: false
+strict: false
+
+plugins:
+  - axolotl.integrations.liger.LigerPlugin
+
+liger_rope: true
+liger_rms_norm: true
+liger_glu_activation: true
+liger_layer_norm: true
+liger_fused_linear_cross_entropy: true
+
+datasets:
+  - path: yahma/alpaca-cleaned
+    type: alpaca
+
+output_dir: ./outputs/qat_out/
+
+sample_packing: true
+pad_to_sequence_len: true
+sequence_len: 512
+
+flex_attention: true
+flex_attn_compile_kwargs:
+  dynamic: false
+  mode: max-autotune-no-cudagraphs
+
+qat:
+  activation_dtype: int8
+  weight_dtype: int4
+  group_size: 32
+
+wandb_project:
+wandb_entity:
+wandb_watch:
+wandb_name:
+wandb_log_model:
+
+gradient_accumulation_steps: 1
+micro_batch_size: 16
+num_epochs: 1
+optimizer: adamw_torch_fused
+
+cosine_constant_lr_ratio: 0
+cosine_min_lr_ratio: 1.0
+learning_rate: 2e-5
+save_only_model: true
+bf16: true
+
+resume_from_checkpoint:
+logging_steps: 1
+
+evals_per_epoch: 1
+saves_per_epoch: 1
+
+warmup_steps: 10
+weight_decay: 0.0
+fsdp:
+  - full_shard
+  - auto_wrap
+
+fsdp_config:
+  fsdp_version: 2
+  fsdp_offload_params: false
+  fsdp_cpu_ram_efficient_loading: true
+  fsdp_auto_wrap_policy: TRANSFORMER_BASED_WRAP
+  fsdp_transformer_layer_cls_to_wrap: LlamaDecoderLayer
+  fsdp_state_dict_type: FULL_STATE_DICT
+  fsdp_sharding_strategy: FULL_SHARD
+  fsdp_reshard_after_forward: true
+  fsdp_activation_checkpointing: true
+
+special_tokens:
+  pad_token: <|end_of_text|>

examples/qwen3/8b-qat-fsdp2.yml

@@ -0,0 +1,78 @@
+base_model: Qwen/Qwen3-8B
+# Automatically upload checkpoint and final model to HF
+# hub_model_id: username/custom_model_name
+
+load_in_8bit: false
+load_in_4bit: false
+strict: false
+
+plugins:
+  - axolotl.integrations.liger.LigerPlugin
+
+liger_rope: true
+liger_rms_norm: true
+liger_glu_activation: true
+liger_layer_norm: true
+liger_fused_linear_cross_entropy: true
+
+datasets:
+  - path: tatsu-lab/alpaca
+    type: alpaca
+
+output_dir: ./outputs/qat_out/
+
+sequence_len: 2048
+sample_packing: true
+flex_attention: true
+pad_to_sequence_len: true
+
+flex_attn_compile_kwargs:
+  dynamic: false
+  mode: max-autotune-no-cudagraphs
+
+qat:
+  activation_dtype: int8
+  weight_dtype: int4
+  group_size: 256
+  fake_quant_after_n_steps: 1000
+
+wandb_project:
+wandb_entity:
+wandb_watch:
+wandb_name:
+wandb_log_model:
+
+gradient_accumulation_steps: 1
+micro_batch_size: 2
+max_steps: 2000
+optimizer: adamw_torch_fused
+lr_scheduler: cosine
+learning_rate: 2e-5
+
+bf16: true
+tf32: true
+
+resume_from_checkpoint:
+logging_steps: 1
+
+evals_per_epoch: 1
+saves_per_epoch: 1
+
+warmup_steps: 10
+weight_decay: 0.0
+fsdp:
+  - full_shard
+  - auto_wrap
+
+fsdp_config:
+  fsdp_version: 2
+  fsdp_offload_params: false
+  fsdp_cpu_ram_efficient_loading: true
+  fsdp_auto_wrap_policy: TRANSFORMER_BASED_WRAP
+  fsdp_transformer_layer_cls_to_wrap: Qwen3DecoderLayer
+  fsdp_state_dict_type: FULL_STATE_DICT
+  fsdp_sharding_strategy: FULL_SHARD
+  fsdp_reshard_after_forward: true
+  fsdp_activation_checkpointing: true
+
+special_tokens:

requirements.txt

@@ -63,7 +63,7 @@ langdetect==1.0.9
 immutabledict==4.2.0
 antlr4-python3-runtime==4.13.2
 
-torchao==0.9.0
+torchao==0.10.0
 schedulefree==1.4.1
 
 axolotl-contribs-lgpl==0.0.6

src/axolotl/cli/args.py

@@ -90,6 +90,18 @@ class VllmServeCliArgs:
     )
 
 
+@dataclass
+class QuantizeCliArgs:
+    """Dataclass with CLI arguments for `axolotl quantize` command."""
+
+    base_model: Optional[str] = field(default=None)
+    weight_dtype: Optional[str] = field(default=None)
+    activation_dtype: Optional[str] = field(default=None)
+    quantize_embedding: Optional[bool] = field(default=None)
+    group_size: Optional[int] = field(default=None)
+    output_dir: Optional[str] = field(default=None)
+
+
 @dataclass
 class EvaluateCliArgs:
     """Dataclass with CLI arguments for `axolotl evaluate` command."""

src/axolotl/cli/main.py

@@ -17,6 +17,7 @@
 from axolotl.cli.args import (
     EvaluateCliArgs,
     PreprocessCliArgs,
+    QuantizeCliArgs,
     TrainerCliArgs,
     VllmServeCliArgs,
 )
@@ -333,6 +334,16 @@ def vllm_serve(config: str, **cli_args: VllmServeCliArgs):
     do_vllm_serve(config, cli_args)
 
 
+@cli.command()
+@click.argument("config", type=click.Path(exists=True, path_type=str))
+@add_options_from_dataclass(QuantizeCliArgs)
+@filter_none_kwargs
+def quantize(config: str, **cli_args: QuantizeCliArgs):
+    from axolotl.cli.quantize import do_quantize
+
+    do_quantize(config, cli_args)
+
+
 @cli.command()
 @click.argument("model", type=click.Path(exists=True, path_type=str))
 @click.argument("output", type=click.Path(exists=False, path_type=str))