change docker

Dockerfile

@@ -1,20 +1,37 @@
-FROM witszhang/songgeneration_vllm:v2
+FROM vllm/vllm-openai:v0.8.5
+
+# The official vllm/vllm-openai image hard-codes
+#   ENTRYPOINT ["python3", "-m", "vllm.entrypoints.openai.api_server"]
+# Clearing it makes our CMD ["python3", "app.py"] actually run.
+ENTRYPOINT []
 
 USER root
 
+# git-lfs       : pulling model weights from HF
+# ffmpeg / libsndfile1 : librosa needs them to decode mp3/m4a/ogg uploads
 RUN apt-get update && \
-    apt-get install -y git-lfs && \
+    apt-get install -y --no-install-recommends git-lfs ffmpeg libsndfile1 && \
     git lfs install && \
     rm -rf /var/lib/apt/lists/*
 
 RUN useradd -m -u 1000 user
 USER user
 ENV PATH="/home/user/.local/bin:$PATH"
+# Force Python stdout/stderr to be unbuffered so that progress bars (tqdm in
+# snapshot_download), vLLM INFO logs and our own prints show up in HF Space
+# logs immediately. Without this the container looks "stuck" at the startup
+# banner because everything sits in a 4-8 KB block buffer until the process
+# exits or the buffer fills up.
+ENV PYTHONUNBUFFERED=1
+ENV PYTHONIOENCODING=utf-8
 
 WORKDIR /app
 
 COPY --chown=user ./requirements.txt requirements.txt
-RUN pip install --no-cache-dir --upgrade -r requirements.txt
+# --prefer-binary : never try to build any wheel from source (flash_attn
+#                   would OOM the build container). The base image already
+#                   provides torch / vllm / flash_attn / transformers.
+RUN pip install --no-cache-dir --upgrade --prefer-binary -r requirements.txt
 
 COPY --chown=user . /app
 CMD ["python3", "app.py"]

app.py

@@ -6,21 +6,6 @@ import json
 import tempfile
 from datetime import datetime
 
-# ---------------------------------------------------------------------------
-# vLLM env tweaks – set BEFORE `import vllm` so they actually take effect.
-#
-# Why these are here:
-#   * VLLM_USE_V1=0  -> stay on the single-process v0 engine. The default v1
-#     engine forks an EngineCore subprocess which roughly doubles host RAM
-#     because the 7B weights get loaded twice. That is the most common cause
-#     of `exit 137 / OOMKilled` on hosts with limited RAM (HF Spaces etc.),
-#     even when the GPU has plenty of free VRAM (L40s 48GB here).
-#   * VLLM_WORKER_MULTIPROC_METHOD=fork  -> if v1 is ever re-enabled, prefer
-#     fork over spawn so weights are CoW-shared instead of reloaded.
-# ---------------------------------------------------------------------------
-os.environ.setdefault("VLLM_USE_V1", "0")
-os.environ.setdefault("VLLM_WORKER_MULTIPROC_METHOD", "fork")
-
 import gradio as gr
 import torch
 import torchaudio
@@ -67,23 +52,11 @@ class VllmInf:
         self.tokenizer = build_tokenizer(args)
         self.text_offset = len(self.tokenizer.tokenizer.get_vocab())
         self.max_tokens = 8192
-        # NOTE on memory:
-        #   - swap_space=0       : disable the default 4 GiB CPU KV swap pool
-        #                          (saves ~4 GB host RAM; GPU has plenty of
-        #                          VRAM so we never need to swap).
-        #   - enforce_eager=True : skip CUDA-graph capture, which otherwise
-        #                          allocates several GB of pinned/host
-        #                          workspace at startup.
-        #   - gpu_memory_utilization stays at default (0.9) — we have a 48GB
-        #     L40s, the OOM is on host RAM, not VRAM.
         self.llm = LLM(
             model=model_path,
             trust_remote_code=True,
             max_model_len=self.max_tokens,
             dtype="bfloat16",
-            swap_space=0,
-            enforce_eager=True,
-            disable_log_stats=True,
         )
 
     def run(self, audios):

requirements.txt

@@ -1,10 +1,10 @@
 # Notes:
-#   The base image `witszhang/songgeneration_vllm:v2` already ships:
-#     - torch / torchaudio / torchvision (CUDA build)
+#   The base image `vllm/vllm-openai:v0.8.5` already ships:
+#     - torch 2.6 / torchaudio / torchvision (CUDA 12.4 build)
 #     - vllm 0.8.5
-#     - transformers
-#     - flash_attn (used by vllm backend)
-#   Re-installing them here triggers a full source build of flash_attn
+#     - transformers (compatible with vllm 0.8.5)
+#     - flash_attn 2.7.x (used by vllm's flash-attn backend)
+#   Re-installing any of them here triggers a full source build of flash_attn
 #   (no PyPI wheel available), which OOM-kills the HF Space *build*
 #   container. Keep this list to lightweight, app-only dependencies.
 librosa>=0.11.0