Merge pull request #68 from vkuzo/20251003_compressed_tensors

scripts for torchao -> compressed_tensors checkpoint conversion

Author: vkuzo

hf_torchao_vllm/convert_torchao_checkpoint_to_compressed_tensors.py

@@ -0,0 +1,201 @@
+import copy
+import filecmp
+import json
+import pathlib
+import shutil
+import subprocess
+from typing import Dict, Any
+
+import fire
+
+import torch
+from torchao.core.config import AOBaseConfig, config_from_dict
+from torchao.quantization import Float8DynamicActivationFloat8WeightConfig, PerRow
+from torchao.quantization.quantize_.workflows.float8.float8_tensor import Float8Tensor
+
+from safetensors import safe_open
+from safetensors.torch import save_file
+
+def ao_config_to_compressed_tensors_config(aobaseconfig: AOBaseConfig) -> Dict[str, Any]:
+    # for now, allowlist of recipes we know how to convert and hand convert
+    # them here
+    # for a production version, we'll need a more scalable way to do this
+
+    assert isinstance(aobaseconfig, Float8DynamicActivationFloat8WeightConfig), "unsupported"
+    assert aobaseconfig.granularity == [PerRow(), PerRow()], "unsupported"
+
+    ct_config = {
+        "format": "float-quantized",
+        "input_activations": {
+            "dynamic": True,
+            "num_bits": 8,
+            "strategy": "token",
+            "symmetric": True,
+            "type": "float",
+        },
+        "output_activations": None,
+        "targets": ["Linear"],
+        "weights": {
+            "dynamic": False,
+            "num_bits": 8,
+            "observer": "minmax",
+            "strategy": "channel",
+            "symmetric": True,
+            "type": "float",
+        },
+    }
+    return ct_config 
+
+def run(
+    # original torchao checkpoint
+    dir_source: str = 'data/torchao/fp8-opt-125m',
+    # new compressed-tensors checkpoint
+    dir_target: str = 'data/torchao_compressed_tensors/fp8-opt-125m',
+    # existing compressed-tensors checkpoint to validate against
+    dir_validation: str = 'data/llmcompressor/fp8-opt-125m',
+    skip_conversion: bool = False,
+):
+    config_name_source = f"{dir_source}/config.json"
+    config_name_target = f"{dir_target}/config.json"
+    config_name_validation = f"{dir_validation}/config.json"
+    weights_name_source = f"{dir_source}/pytorch_model.bin" 
+    weights_name_target = f"{dir_target}/model.safetensors"
+    weights_name_validation = f"{dir_validation}/model.safetensors"
+
+    if not skip_conversion:
+        #
+        # convert config.json
+        #
+
+        with open(config_name_source, 'r') as f:
+            config_source = json.load(f)
+
+        # get torchao config format
+        # example: https://www.internalfb.com/phabricator/paste/view/P1975688376
+        # we need to translate it to compressed-tensors format
+        # example: https://www.internalfb.com/phabricator/paste/view/P1975642629
+        old_hf_quantization_config = config_source["quantization_config"]
+        fqn_to_serialized_aobaseconfig = old_hf_quantization_config["quant_type"]
+        assert len(fqn_to_serialized_aobaseconfig) == 1, "unsupported"
+
+        new_hf_quantization_config = {
+            "config_groups": {},
+            "format": "float-quantized",
+            "ignore": ["lm_head"],
+            "quant_method": "compressed-tensors",
+            "quantization_status": "compressed",
+            "sparsity_config": {},
+            "transform_config": {},
+            "version": "torchao_hack",
+        }
+
+        for fqn, serialized_aobaseconfig in fqn_to_serialized_aobaseconfig.items():
+            print(fqn, serialized_aobaseconfig)
+            aobaseconfig = config_from_dict(serialized_aobaseconfig)
+            print(aobaseconfig)
+            ct_config = ao_config_to_compressed_tensors_config(aobaseconfig)
+            print(json.dumps(ct_config, indent=2))
+
+            assert fqn == "default", "unsupported"
+            new_hf_quantization_config["config_groups"]["group_0"] = ct_config
+
+        # for now, modify config_source inplace
+        config_source["quantization_config"] = new_hf_quantization_config
+
+        # save to new location
+        with open(config_name_target, 'w') as f:
+            json.dump(config_source, f, indent=2)
+
+        #
+        # convert the checkpoint
+        #
+
+        # not sure why I still need this
+        torch.serialization.add_safe_globals([getattr])
+
+        old_state_dict = torch.load(weights_name_source, weights_only=True)
+        new_state_dict = {}
+
+        for k, v in old_state_dict.items():
+            print(k, v.shape, type(v))
+            if type(v) == torch.Tensor:
+                
+                if "lm_head" in k:
+                    # work around issues detailed in 
+                    # https://huggingface.co/docs/safetensors/torch_shared_tensors
+                    v = copy.deepcopy(v)
+
+                new_state_dict[k] = v
+            elif type(v) == Float8Tensor:
+                new_state_dict[k] = v.qdata
+                # for now, manually cast scale to bfloat16 to match currnt 
+                # llm-compressor script
+                # TODO(future): prob needs to be user controllable 
+                new_state_dict[k + '_scale'] = v.scale.bfloat16()
+            else:
+                raise AssertionError(f'unsupported type {type(v)}')
+        save_file(new_state_dict, weights_name_target)
+
+        # move all the other files over
+        for dir_and_file_path in pathlib.Path(dir_source).iterdir():
+            if not dir_and_file_path.is_file():
+                continue
+            file_path = dir_and_file_path.parts[-1]
+            if file_path in ('config.json', 'pytorch_model.bin'):
+                # these are converted in custom logic elsewhere in this script
+                continue
+            # if we got here, we just need to copy the file over without any changes
+            target_file_path = f"{dir_target}/{str(file_path)}"
+            shutil.copyfile(dir_and_file_path, target_file_path)
+
+    # validate target_dir vs validation_dir
+    for dir_and_file_path in pathlib.Path(dir_target).iterdir():
+        if not dir_and_file_path.is_file():
+            continue
+        file_path_target = dir_and_file_path.parts[-1]
+        print("\nvalidating", file_path_target)
+        dir_and_file_path_validation = f"{dir_validation}/{str(file_path_target)}"
+
+        if file_path_target == 'config.json':
+            # for now just diff and print the output to stdout
+            command = f'diff {dir_and_file_path} {dir_and_file_path_validation}'
+            try:
+                result = subprocess.run(command, capture_output=False, text=True, shell=True, check=True)
+            except subprocess.CalledProcessError as e:
+                # this will always fail, for now, as we are not perfectly matching
+                print(e.stderr) 
+
+        elif file_path_target == 'model.safetensors':
+            # TODO implement me
+            pass
+
+            with safe_open(dir_and_file_path, framework='pt') as f_target:
+                with safe_open(dir_and_file_path_validation, framework='pt') as f_validation:
+                    k_target_seen = set()
+                    for k_target in f_target.keys():
+                        v_target = f_target.get_tensor(k_target)
+                        v_validation = f_validation.get_tensor(k_target)
+
+                        # ensure metadata matches
+                        if v_target.shape != v_validation.shape:
+                            print(f"shape mismatch: {k_target=}, {v_target.shape=}, {v_validation.shape=}")
+
+                        if v_target.dtype != v_validation.dtype: 
+                            print(f"dtype mismatch: {k_target=}, {v_target.dtype=}, {v_validation.dtype=}")
+
+                        # for now, no numerical checks
+
+                        k_target_seen.add(k_target)
+
+                    for k_validation in f_validation.keys():
+                        if k_validation not in k_target_seen:
+                            print(f"key {k_validation} not present in target")
+
+        else:
+            # approx check, currently fails because modification timestamp is not the
+            # same. Since we copy these files ourselves, low-pri to make this better.
+            is_equal = filecmp.cmp(dir_and_file_path, dir_and_file_path_validation, shallow=False)
+            print('filecmp equal', is_equal)
+
+if __name__ == '__main__':
+    fire.Fire(run)

hf_torchao_vllm/inspect_llm_compressor_output.py

@@ -6,7 +6,7 @@
 import fire
 
 def run(
-    dir_name: str = 'data/llmcompressor/opt-125m-FP8-Dynamic',
+    dir_name: str = 'data/llmcompressor/fp8-opt-125m',
 ):
     json_config_name = f'{dir_name}/config.json'
     with open(json_config_name, 'r') as f:

hf_torchao_vllm/quantize_hf_model_with_llm_compressor.py

@@ -1,5 +1,7 @@
 # https://github.com/vllm-project/llm-compressor/blob/main/examples/quantization_w8a8_fp8/llama3_example.py
 
+import torch
+
 from transformers import AutoModelForCausalLM, AutoTokenizer
 
 from llmcompressor import oneshot
@@ -14,7 +16,7 @@ def run():
     MODEL_ID = "facebook/opt-125m"
 
     # Load model.
-    model = AutoModelForCausalLM.from_pretrained(MODEL_ID, torch_dtype="auto")
+    model = AutoModelForCausalLM.from_pretrained(MODEL_ID, torch_dtype=torch.bfloat16)
     tokenizer = AutoTokenizer.from_pretrained(MODEL_ID)
 
     # Configure the quantization algorithm and scheme.
@@ -39,7 +41,7 @@ def run():
     print("==========================================")
 
     # Save to disk in compressed-tensors format.
-    SAVE_DIR = "data/llmcompressor/" + MODEL_ID.rstrip("/").split("/")[-1] + "-FP8-Dynamic"
+    SAVE_DIR = "data/llmcompressor/" + "fp8-" + MODEL_ID.rstrip("/").split("/")[-1]
     model.save_pretrained(SAVE_DIR)
     tokenizer.save_pretrained(SAVE_DIR)
 

hf_torchao_vllm/run_quantized_model_in_vllm.py

@@ -54,7 +54,7 @@ def print_vllm_torchao_quant_info(model: torch.nn.Module):
 
 def main(
     # model_name: str = "Qwen/Qwen2-7B-Instruct",
-    model_name: str = "data/fp8-opt-125m",
+    model_name: str = "data/torchao/fp8-opt-125m",
     max_tokens=64,
     tp_size: int = 1,
     compile: bool = True,