Remove obsolete DeepSpeed inference code (2023 relic)

modules/transformers_loader.py

@@ -1,4 +1,3 @@
-import os
 import pprint
 from pathlib import Path
 
@@ -6,11 +5,7 @@ import torch
 import torch.nn.functional as F
 import transformers
 from accelerate import infer_auto_device_map, init_empty_weights
-from accelerate.utils import (
-    is_ccl_available,
-    is_npu_available,
-    is_xpu_available
-)
+from accelerate.utils import is_xpu_available
 from transformers import (
     AutoConfig,
     AutoModel,
@@ -28,31 +23,6 @@ from modules.torch_utils import get_device
 
 transformers.logging.set_verbosity_error()
 
-local_rank = None
-if shared.args.deepspeed:
-    import deepspeed
-    from transformers.integrations.deepspeed import (
-        HfDeepSpeedConfig,
-        is_deepspeed_zero3_enabled
-    )
-
-    from modules.deepspeed_parameters import generate_ds_config
-
-    # Distributed setup
-    local_rank = shared.args.local_rank if shared.args.local_rank is not None else int(os.getenv("LOCAL_RANK", "0"))
-    world_size = int(os.getenv("WORLD_SIZE", "1"))
-    if is_xpu_available() and is_ccl_available():
-        torch.xpu.set_device(local_rank)
-        deepspeed.init_distributed(backend="ccl")
-    elif is_npu_available():
-        torch.npu.set_device(local_rank)
-        deepspeed.init_distributed(dist_backend="hccl")
-    else:
-        torch.cuda.set_device(local_rank)
-        deepspeed.init_distributed()
-    ds_config = generate_ds_config(shared.args.bf16, 1 * world_size, shared.args.nvme_offload_dir)
-    dschf = HfDeepSpeedConfig(ds_config)  # Keep this object alive for the Transformers integration
-
 
 class _StopEverythingStoppingCriteria(transformers.StoppingCriteria):
     def __init__(self):
@@ -163,7 +133,6 @@ def load_model_HF(model_name):
         shared.args.load_in_8bit,
         shared.args.load_in_4bit,
         shared.args.disk,
-        shared.args.deepspeed,
         shared.args.cpu_memory is not None,
         shared.args.compress_pos_emb > 1,
         shared.args.alpha_value > 1,
@@ -183,25 +152,6 @@ def load_model_HF(model_name):
             if device:
                 model = model.to(device)
 
-    # DeepSpeed ZeRO-3
-    elif shared.args.deepspeed:
-        model = LoaderClass.from_pretrained(
-            path_to_model,
-            torch_dtype=params['torch_dtype'],
-            trust_remote_code=params.get('trust_remote_code')
-        )
-
-        model = deepspeed.initialize(
-            model=model,
-            config_params=ds_config,
-            model_parameters=None,
-            optimizer=None,
-            lr_scheduler=None
-        )[0]
-
-        model.module.eval()  # Inference
-        logger.info(f'DeepSpeed ZeRO-3 is enabled: {is_deepspeed_zero3_enabled()}')
-
     # Load with quantization and/or offloading
     else:
         if not any((shared.args.cpu, torch.cuda.is_available(), is_xpu_available(), torch.backends.mps.is_available())):