--no-cache and --gpu-memory in MiB for fine VRAM control

README.md

@@ -183,6 +183,7 @@ Optionally, you can use the following command-line flags:
 | `--disk-cache-dir DISK_CACHE_DIR` | Directory to save the disk cache to. Defaults to `cache/`. |
 |  `--gpu-memory GPU_MEMORY [GPU_MEMORY ...]` |  Maxmimum GPU memory in GiB to be allocated per GPU. Example: `--gpu-memory 10` for a single GPU, `--gpu-memory 10 5` for two GPUs. |
 | `--cpu-memory CPU_MEMORY` | Maximum CPU memory in GiB to allocate for offloaded weights. Must be an integer number. Defaults to 99.|
+| `--no-cache`     | Set `use_cache` to False while generating text. This reduces the VRAM usage a bit at a performance cost.')
 | `--flexgen`      | Enable the use of FlexGen offloading. |
 |  `--percent PERCENT [PERCENT ...]` |  FlexGen: allocation percentages. Must be 6 numbers separated by spaces (default: 0, 100, 100, 0, 100, 0). |
 |  `--compress-weight` |  FlexGen: Whether to compress weight (default: False).|

modules/models.py

@@ -1,5 +1,6 @@
 import json
 import os
+import re
 import time
 import zipfile
 from pathlib import Path
@@ -120,11 +121,12 @@ def load_model(model_name):
                 params["torch_dtype"] = torch.float16
 
             if shared.args.gpu_memory:
-                memory_map = shared.args.gpu_memory
+                memory_map = list(map(lambda x : x.strip(), shared.args.gpu_memory))
+                max_cpu_memory = shared.args.cpu_memory.strip() if shared.args.cpu_memory is not None else '99GiB'
                 max_memory = {}
                 for i in range(len(memory_map)):
-                    max_memory[i] = f'{memory_map[i]}GiB'
-                max_memory['cpu'] = f'{shared.args.cpu_memory or 99}GiB'
+                    max_memory[i] = f'{memory_map[i]}GiB' if not re.match('.*ib$', memory_map[i].lower()) else memory_map[i]
+                max_memory['cpu'] = max_cpu_memory
                 params['max_memory'] = max_memory
             elif shared.args.auto_devices:
                 total_mem = (torch.cuda.get_device_properties(0).total_memory / (1024*1024))

modules/shared.py

@@ -85,8 +85,9 @@ parser.add_argument('--bf16', action='store_true', help='Load the model with bfl
 parser.add_argument('--auto-devices', action='store_true', help='Automatically split the model across the available GPU(s) and CPU.')
 parser.add_argument('--disk', action='store_true', help='If the model is too large for your GPU(s) and CPU combined, send the remaining layers to the disk.')
 parser.add_argument('--disk-cache-dir', type=str, default="cache", help='Directory to save the disk cache to. Defaults to "cache".')
-parser.add_argument('--gpu-memory', type=int, nargs="+", help='Maxmimum GPU memory in GiB to be allocated per GPU. Example: --gpu-memory 10 for a single GPU, --gpu-memory 10 5 for two GPUs.')
-parser.add_argument('--cpu-memory', type=int, help='Maximum CPU memory in GiB to allocate for offloaded weights. Must be an integer number. Defaults to 99.')
+parser.add_argument('--gpu-memory', type=str, nargs="+", help='Maxmimum GPU memory in GiB to be allocated per GPU. Example: --gpu-memory 10 for a single GPU, --gpu-memory 10 5 for two GPUs.')
+parser.add_argument('--cpu-memory', type=str, help='Maximum CPU memory in GiB to allocate for offloaded weights. Must be an integer number. Defaults to 99.')
+parser.add_argument('--no-cache', action='store_true', help='Set use_cache to False while generating text. This reduces the VRAM usage a bit at a performance cost.')
 parser.add_argument('--flexgen', action='store_true', help='Enable the use of FlexGen offloading.')
 parser.add_argument('--percent', type=int, nargs="+", default=[0, 100, 100, 0, 100, 0], help='FlexGen: allocation percentages. Must be 6 numbers separated by spaces (default: 0, 100, 100, 0, 100, 0).')
 parser.add_argument("--compress-weight", action="store_true", help="FlexGen: activate weight compression.")

modules/text_generation.py

@@ -136,7 +136,9 @@ def generate_reply(question, max_new_tokens, do_sample, temperature, top_p, typi
         t = encode(stopping_string, 0, add_special_tokens=False)
         stopping_criteria_list.append(_SentinelTokenStoppingCriteria(sentinel_token_ids=t, starting_idx=len(input_ids[0])))
 
-    generate_params = {}
+    generate_params = {
+        'use_cache': not shared.args.no_cache,
+    }
     if not shared.args.flexgen:
         generate_params.update({
             "max_new_tokens": max_new_tokens,