getting better

src/axolotl/utils/models.py

@@ -1,16 +1,21 @@
 """Module for models and model loading"""
+import json
 import logging
 import math
 import os
 from typing import Optional, Tuple  # noqa: F401
 
+import accelerate
 import bitsandbytes as bnb
 import tensor_parallel as tp
 import torch
 import transformers
+import transformers.utils.bitsandbytes
+from huggingface_hub import hf_hub_download
 from optimum.bettertransformer import BetterTransformer
 from peft import PeftConfig, prepare_model_for_kbit_training
 from peft.tuners.lora import QuantLinear
+from safetensors.torch import load_file as load_safetensors_file
 from transformers import (  # noqa: F401
     AddedToken,
     AutoConfig,
@@ -222,7 +227,7 @@ def load_model(
             load_in_4bit=True,
             llm_int8_threshold=6.0,
             llm_int8_has_fp16_weight=False,
-            bnb_4bit_compute_dtype=cfg.torch_dtype,
+            bnb_4bit_compute_dtype=torch.float16,
             bnb_4bit_use_double_quant=True,
             bnb_4bit_quant_type="nf4",
         )
@@ -327,27 +332,15 @@ def load_model(
                 base_model,
                 trust_remote_code=cfg.trust_remote_code or False,
             )
-            # with accelerate.init_empty_weights():
-            #     model = AutoModelForCausalLM.from_config(
-            #         config=config,
-            #         trust_remote_code=cfg.trust_remote_code or False,
-            #     ).half()
-            # model = tp.TensorParallelPreTrainedModel(
-            #     model,
-            #     sharded=False,
-            # )
-            model = AutoModelForCausalLM.from_pretrained(
-                base_model,
-                config=config,
-                trust_remote_code=cfg.trust_remote_code or False,
-                low_cpu_mem_usage=True,
-                load_in_8bit=cfg.load_in_8bit and cfg.adapter is not None,
-                load_in_4bit=cfg.load_in_4bit and cfg.adapter is not None,
-                torch_dtype=cfg.torch_dtype,
-                device_map={"": "cpu"},
-                **model_kwargs,
+            with accelerate.init_empty_weights():
+                model = AutoModelForCausalLM.from_config(
+                    config=config,
+                    trust_remote_code=cfg.trust_remote_code or False,
+                ).half()
+            model = tp.TensorParallelPreTrainedModel(
+                model,
+                sharded=False,
             )
-            model = tp.tensor_parallel(model, sharded=False)
         else:
             config = AutoConfig.from_pretrained(
                 base_model,
@@ -474,6 +467,52 @@ def load_model(
 
     model, lora_config = load_adapter(model, cfg, cfg.adapter)
 
+    if cfg.tensor_parallel and cfg.adapter == "qlora":
+        device_map = tp.infer_sharded_device_map(model)
+
+        load_file = torch.load
+        try:
+            with open(
+                hf_hub_download(base_model, "pytorch_model.bin.index.json"), "r"
+            ) as index_file:
+                shard_filenames = set(json.load(index_file)["weight_map"].values())
+        except:
+            with open(
+                hf_hub_download(base_model, "model.safetensors.index.json"), "r"
+            ) as index_file:
+                shard_filenames = set(json.load(index_file)["weight_map"].values())
+            load_file = load_safetensors_file
+
+        for shard_filename in sorted(shard_filenames):
+            # Download a shard
+            shard_path = hf_hub_download(base_model, shard_filename)
+
+            # Convert model shard
+            converted_state_dict = (
+                tp.convert_state_dict(  # <- tensor_parallel helper function.
+                    load_file(
+                        shard_path
+                    ),  #    Creates a tensor_parallel checkpoint form a normal one
+                    model.tensor_parallel_config,
+                    world_size=torch.cuda.device_count(),
+                    for_pretrained=True,
+                )
+            )
+
+            # Dispatch the shard
+            for param_name, param in converted_state_dict.items():
+                module_name = param_name
+
+                while len(module_name) > 0 and module_name not in device_map:
+                    module_name = ".".join(module_name.split(".")[:-1])
+                param_device = device_map[module_name]
+
+                transformers.utils.bitsandbytes.set_module_quantized_tensor_to_device(
+                    model, param_name, param_device, value=param.to(dtype=torch.float16)
+                )
+                converted_state_dict[param_name] = None
+            del converted_state_dict
+
     if cfg.ddp and not load_in_8bit:
         model.to(f"cuda:{cfg.local_rank}")