persistent model storage

server/gpu/modal/reflector_llm.py

@@ -7,6 +7,7 @@ import json
 import os
 from typing import Optional
 
+import modal
 from modal import Image, Secret, Stub, asgi_app, method
 
 # LLM
@@ -15,16 +16,17 @@ LLM_LOW_CPU_MEM_USAGE: bool = True
 LLM_TORCH_DTYPE: str = "bfloat16"
 LLM_MAX_NEW_TOKENS: int = 300
 
-IMAGE_MODEL_DIR = "/model"
+IMAGE_MODEL_DIR = "/root/llm_models"
+volume = modal.NetworkFileSystem.persisted("reflector-llm-models")
 
-stub = Stub(name="reflector-llm")
+stub = Stub(name="reflector-llmtest1")
 
 
 def download_llm():
     from huggingface_hub import snapshot_download
 
     print("Downloading LLM model")
-    snapshot_download(LLM_MODEL, local_dir=IMAGE_MODEL_DIR)
+    snapshot_download(LLM_MODEL, cache_dir=IMAGE_MODEL_DIR)
     print("LLM model downloaded")
 
 
@@ -38,7 +40,7 @@ def migrate_cache_llm():
     from transformers.utils.hub import move_cache
 
     print("Moving LLM cache")
-    move_cache(cache_dir=IMAGE_MODEL_DIR)
+    move_cache()
     print("LLM cache moved")
 
 
@@ -58,7 +60,6 @@ llm_image = (
     )
     .env({"HF_HUB_ENABLE_HF_TRANSFER": "1"})
     .run_function(download_llm)
-    .run_function(migrate_cache_llm)
 )
 
 
@@ -68,6 +69,7 @@ llm_image = (
     container_idle_timeout=60 * 5,
     concurrency_limit=2,
     image=llm_image,
+    network_file_systems={IMAGE_MODEL_DIR: volume},
 )
 class LLM:
     def __enter__(self):
@@ -77,9 +79,10 @@ class LLM:
 
         print("Instance llm model")
         model = AutoModelForCausalLM.from_pretrained(
-            IMAGE_MODEL_DIR,
+            LLM_MODEL,
             torch_dtype=getattr(torch, LLM_TORCH_DTYPE),
             low_cpu_mem_usage=LLM_LOW_CPU_MEM_USAGE,
+            cache_dir=IMAGE_MODEL_DIR
         )
 
         # generation configuration
@@ -91,7 +94,10 @@ class LLM:
 
         # load tokenizer
         print("Instance llm tokenizer")
-        tokenizer = AutoTokenizer.from_pretrained(LLM_MODEL)
+        tokenizer = AutoTokenizer.from_pretrained(
+            LLM_MODEL,
+            cache_dir=IMAGE_MODEL_DIR
+        )
 
         # move model to gpu
         print("Move llm model to GPU")

server/gpu/modal/reflector_transcriber.py

@@ -6,6 +6,7 @@ Reflector GPU backend - transcriber
 import os
 import tempfile
 
+import modal
 from modal import Image, Secret, Stub, asgi_app, method
 from pydantic import BaseModel
 
@@ -17,12 +18,13 @@ WHISPER_NUM_WORKERS: int = 1
 # Translation Model
 TRANSLATION_MODEL = "facebook/m2m100_418M"
 
-MODEL_DIR = "model"
+IMAGE_MODEL_DIR = "/root/transcription_models"
+volume = modal.NetworkFileSystem.persisted("reflector-transcribe-models")
 
-stub = Stub(name="reflector-transtest")
+stub = Stub(name="reflector-transtest1")
 
 
-def download_whisper(cache_dir: str = None):
+def download_whisper(cache_dir: str | None = None):
     from faster_whisper.utils import download_model
 
     print("Downloading Whisper model")
@@ -30,32 +32,24 @@ def download_whisper(cache_dir: str = None):
     print("Whisper model downloaded")
 
 
-def download_translation_model(cache_dir: str = None):
+def download_translation_model(cache_dir: str | None = None):
     from huggingface_hub import snapshot_download
 
     print("Downloading Translation model")
     ignore_patterns = ["*.ot"]
-    snapshot_download(TRANSLATION_MODEL, cache_dir=cache_dir, ignore_patterns=ignore_patterns)
+    snapshot_download(
+        TRANSLATION_MODEL,
+        cache_dir=cache_dir,
+        ignore_patterns=ignore_patterns
+    )
     print("Translation model downloaded")
 
 
 def download_models():
-    download_whisper(cache_dir=MODEL_DIR)
+    print(f"Downloading models to {IMAGE_MODEL_DIR=}")
-    download_translation_model(cache_dir=MODEL_DIR)
+    download_whisper(cache_dir=IMAGE_MODEL_DIR)
-
+    download_translation_model(cache_dir=IMAGE_MODEL_DIR)
-
+    print(f"Model downloads complete.")
-def migrate_cache_llm():
-    """
-    XXX The cache for model files in Transformers v4.22.0 has been updated.
-    Migrating your old cache. This is a one-time only operation. You can
-    interrupt this and resume the migration later on by calling
-    `transformers.utils.move_cache()`.
-    """
-    from transformers.utils.hub import move_cache
-
-    print("Moving LLM cache")
-    move_cache()
-    print("LLM cache moved")
 
 
 whisper_image = (
@@ -71,7 +65,6 @@ whisper_image = (
         "huggingface_hub==0.16.4",
     )
     .run_function(download_models)
-    .run_function(migrate_cache_llm)
     .env(
         {
             "LD_LIBRARY_PATH": (
@@ -87,6 +80,7 @@ whisper_image = (
     gpu="A10G",
     container_idle_timeout=60,
     image=whisper_image,
+    network_file_systems={IMAGE_MODEL_DIR: volume},
 )
 class Whisper:
     def __enter__(self):
@@ -101,14 +95,15 @@ class Whisper:
             device=self.device,
             compute_type=WHISPER_COMPUTE_TYPE,
             num_workers=WHISPER_NUM_WORKERS,
+            download_root=IMAGE_MODEL_DIR
         )
         self.translation_model = M2M100ForConditionalGeneration.from_pretrained(
             TRANSLATION_MODEL,
-            cache_dir=MODEL_DIR
+            cache_dir=IMAGE_MODEL_DIR
         ).to(self.device)
         self.translation_tokenizer = M2M100Tokenizer.from_pretrained(
             TRANSLATION_MODEL,
-            cache_dir=MODEL_DIR
+            cache_dir=IMAGE_MODEL_DIR
         )
 
     @method()