Remove the Chat tab and Gemma: this is a graphics studio

README.md

@@ -8,14 +8,10 @@ app_file: app.py
 pinned: false
 license: apache-2.0
 short_description: FLUX image, LTX video, MusicGen music, Kokoro+Qwen narration
-hf_oauth: true
-hf_oauth_scopes:
-  - inference-api
 tags:
   - text-to-image
   - image-to-video
   - text-to-video
-  - text-generation
   - text-to-speech
   - text-to-audio
   - audio-text-to-text
@@ -46,7 +42,6 @@ models:
   - hexgrad/Kokoro-82M
   - openai/whisper-large-v3-turbo
   - laion/clap-htsat-unfused
-  - google/gemma-4-31B-it
   - florence-community/Florence-2-large
   - HuggingFaceTB/SmolVLM2-2.2B-Instruct
   - Qwen/Qwen3-4B
@@ -62,8 +57,6 @@ models:
 
 **Narration tab (text to speech, speech generation, machine speech):** Kokoro (hexgrad) reads your text aloud in a choice of eight voices.
 
-**Chat tab:** Gemma 4 31B (google) answers questions and writes text. It runs in the Space in 8-bit, so no sign-in is needed; a reply uses your own ZeroGPU quota.
-
 **Describe tab:** three ways to describe what you have.
 
 - **Image to text (image captioning, image recognition):** Florence-2 large (microsoft, via florence-community) writes a detailed caption of an uploaded picture.
@@ -80,6 +73,6 @@ models:
 
 **How it stays light on ZeroGPU quota:** each model is kept in its own pack, and a button copies only the models it needs onto the GPU, one after another where a job uses several. An image copies only the image model, and a video loads FLUX, then LTX-Video, then MusicGen in turn, so every job fits the default half-card slice.
 
-**Licences:** LTX-Video, Kokoro, Whisper, CLAP, Qwen3 4B, Gemma 4 31B, Florence-2 large and SmolVLM2 2.2B are openly licensed. **FLUX.2 klein 9B is under the FLUX Non-Commercial License** and **MusicGen medium is CC-BY-NC 4.0**, so images, videos and music made here are for non-commercial use.
+**Licences:** LTX-Video, Kokoro, Whisper, CLAP, Qwen3 4B, Florence-2 large and SmolVLM2 2.2B are openly licensed. **FLUX.2 klein 9B is under the FLUX Non-Commercial License** and **MusicGen medium is CC-BY-NC 4.0**, so images, videos and music made here are for non-commercial use.
 
 Runs on ZeroGPU: set the Space hardware to **ZeroGPU** in Settings.

app.py

@@ -5,7 +5,6 @@ Music tab (text to music, text to audio): MusicGen medium turns a description in
 Transcribe tab (speech to text, speech recognition, automatic speech recognition): Whisper large v3 turbo
 writes down speech from an uploaded file or the microphone, in its own language or translated to English.
 Narration tab (text to speech, machine speech, speech generation): Kokoro reads your text aloud.
-Chat tab: Gemma 4 31B answers questions, running here in 8-bit (hosted Gemma with sign-in as fallback).
 Describe tab: three sub-tabs for the describing jobs the Space does.
   Image to text (image captioning, image recognition): Florence-2 large writes a detailed caption for
     an uploaded picture.
@@ -23,8 +22,6 @@ Each model has its own ZeroGPU pack (see zero_packs.py), and each button copies
 needs onto the GPU, one after another where a job uses several big models. That keeps every job
 inside the default half-card slice, which is not billed double. If zero_packs is switched off or
 cannot work, everything is copied on every request as ZeroGPU normally does, on the full card.
-The Chat tab uses hosted Gemma 4 31B (billed to each visitor's own HF inference credits) and no
-ZeroGPU time at all.
 """
 import os
 
@@ -46,7 +43,6 @@ _store = _pick_model_store()
 if _store:
     os.environ["HF_HOME"] = _store          # Must be set before the Hugging Face libraries are imported
 
-import gc
 import json
 import random
 import re
@@ -67,7 +63,6 @@ import zero_packs as packs
 from diffusers import FluxPipeline, LTXConditionPipeline
 from diffusers.pipelines.ltx.pipeline_ltx_condition import LTXVideoCondition
 from diffusers.utils import export_to_video
-from huggingface_hub import InferenceClient
 from PIL import Image
 from transformers import (AutoModelForCausalLM, AutoModelForImageTextToText, AutoProcessor,
                           AutoTokenizer, MusicgenForConditionalGeneration, pipeline)
@@ -85,7 +80,6 @@ IMAGE_CAPTION_MODEL = "florence-community/Florence-2-large"      # Picture to ca
 VIDEO_CAPTION_MODEL = "HuggingFaceTB/SmolVLM2-2.2B-Instruct"     # Video frames to description
 LISTEN_MODEL = "laion/clap-htsat-unfused"      # Recognises genre, mood, instruments and tempo
 SPEECH_MODEL = "openai/whisper-large-v3-turbo" # Speech to text
-CHAT_MODEL = "google/gemma-4-31B-it"           # Runs here in 8-bit; hosted inference is the fallback
 VOICE = "af_heart"                             # Kokoro voice
 VOICE_RATE = 24000
 VOICES = {"Heart (woman)": "af_heart", "Bella (woman)": "af_bella", "Nicole (woman, soft)": "af_nicole",
@@ -112,34 +106,6 @@ SPEED_SUFFIX = ", moving at natural real-time speed"
 # Pure per-call .to('cuda') from ordinary RAM was tried on 2026-10-07: the first FLUX move took 79 s.
 # The packs are read straight from disk at about 17 GB a second, like ZeroGPU's own single pack.
 # Once a model is sealed, never call .to("cuda") on it again here: its weights are on disk.
-# Chat: Gemma 4 31B with its weights squeezed to 8-bit by bitsandbytes (about 32 GB, so it fits the
-# half card on its own). The squeezing is done here on the CPU at start-up, which takes a few minutes.
-# If anything goes wrong the Space still starts, and the Chat tab uses hosted Gemma with sign-in.
-# It is loaded first, while the disk is empty: on 2026-10-09 downloading Gemma (62 GB) after the other
-# models went over the Space's 150 GB disk limit and the Space was evicted. Every model's downloaded
-# files are now deleted as soon as its pack is sealed. Set the Space variable LOCAL_CHAT to "0" to skip it.
-try:        # Which disks there are and how full: the packs' disk and the download cache have separate limits
-    print("[disk]", subprocess.run(["df", "-h"], capture_output=True, text=True).stdout.replace("\n", " | "), flush=True)
-except Exception as e:
-    print(f"[disk] df failed: {e}", flush=True)
-chat_model = chat_processor = None
-try:
-    if not packs.enabled or os.environ.get("LOCAL_CHAT", "0") != "1":
-        raise RuntimeError("switched off")
-    print("[load] Gemma 4 31B (8-bit)...", flush=True)
-    from transformers import BitsAndBytesConfig
-    _t0 = time.time()
-    chat_processor = AutoProcessor.from_pretrained(CHAT_MODEL)
-    chat_model = AutoModelForImageTextToText.from_pretrained(
-        CHAT_MODEL, quantization_config=BitsAndBytesConfig(load_in_8bit=True),
-        torch_dtype=torch.bfloat16, device_map={"": "cuda"})
-    print(f"[load] Gemma ready in {time.time() - _t0:.0f}s", flush=True)
-    packs.seal("gemma", CHAT_MODEL)
-except Exception as e:
-    print(f"[load] no local Gemma, Chat will use hosted Gemma: {type(e).__name__}: {str(e)[:300]}", flush=True)
-    chat_model = None
-    gc.collect()                            # Drop any half-loaded weights, so they are not packed with the next model
-    packs.free_downloads(CHAT_MODEL)        # And its downloaded files, so the disk limit is not hit later
 print("[load] image model...", flush=True)
 try:
     from diffusers import Flux2KleinPipeline
@@ -203,15 +169,6 @@ speech_pipe = pipeline("automatic-speech-recognition", model=SPEECH_MODEL, torch
                        device="cuda", chunk_length_s=30)
 packs.seal("whisper", SPEECH_MODEL)
 print("[load] done.", flush=True)
-try:        # Where the downloads and packs really are, and what counts towards the disk limits
-    from huggingface_hub import constants as _hub
-    _script = ("du -xsh /root/.cache/* /root/.cache/huggingface/* /root/.zerogpu /data-nvme/* /tmp /home/* 2>/dev/null"
-               " | sort -h | tail -25; uname -r; free -g | head -2")
-    print("[where]", f"HF_HUB_CACHE={_hub.HF_HUB_CACHE} HOME={os.path.expanduser('~')} offload={packs.PACK_DIR}", "|",
-          subprocess.run(["sh", "-c", _script], capture_output=True, text=True, timeout=120).stdout.replace("\n", " | "),
-          flush=True)
-except Exception as e:
-    print(f"[where] failed: {e}", flush=True)
 
 # With a pack per model every job fits the default half-card slice ("large"), which is billed once.
 # Without them ZeroGPU copies all the models on every call, and that needs the whole card.
@@ -529,71 +486,6 @@ def video_frames(path, count=6):
     return frames, length
 
 
-# ---------------------------------------------------------------- Chat (hosted Gemma; no GPU)
-
-def as_text(content):
-    if isinstance(content, str):
-        return content
-    if isinstance(content, dict):
-        return str(content.get("text", ""))
-    if isinstance(content, (list, tuple)):
-        return " ".join(as_text(part) for part in content)
-    return str(content or "")
-
-
-def chat_seconds(messages):
-    return gpu_seconds(60)
-
-
-@spaces.GPU(duration=chat_seconds, size=GPU_SIZE)
-def chat_gpu(messages):
-    """Gemma 4 31B (8-bit) answers on the GPU."""
-    with timed("chat"):
-        packs.load("gemma")
-        turns = [{"role": m["role"], "content": [{"type": "text", "text": m["content"]}]} for m in messages]
-        inputs = chat_processor.apply_chat_template(
-            turns, add_generation_prompt=True, tokenize=True, return_dict=True, return_tensors="pt").to("cuda")
-        start = time.time()
-        out = chat_model.generate(**inputs, max_new_tokens=300, do_sample=True, temperature=0.7, top_p=0.9)
-        new = out[0][inputs["input_ids"].shape[1]:]
-        print(f"[chat] {len(new)} tokens at {len(new) / (time.time() - start):.1f} a second", flush=True)
-        return chat_processor.decode(new, skip_special_tokens=True).strip()
-
-
-def chat(message, history, oauth_token: gr.OAuthToken | None = None):
-    """Gemma 4 31B. Runs here in 8-bit; if that could not be loaded, hosted Gemma with the visitor's sign-in."""
-    if chat_model is not None:
-        messages = [{"role": "system", "content": "You are a helpful, friendly assistant. Answer clearly and briefly."}]
-        for turn in history[-8:]:
-            if isinstance(turn, dict):
-                messages.append({"role": turn.get("role", "user"), "content": as_text(turn.get("content"))})
-        messages.append({"role": "user", "content": as_text(message)})
-        return chat_gpu(messages)
-    if oauth_token is None:
-        return ("Please press **Sign in with Hugging Face** above. Each visitor's chat reply is billed to their "
-                "own Hugging Face inference credits, not the owner's — so sign-in is required for this tab.")
-    messages = [{"role": "system",
-                 "content": "You are a helpful, friendly assistant. Answer clearly and briefly."}]
-    for turn in history[-12:]:
-        if isinstance(turn, dict):
-            messages.append({"role": turn.get("role", "user"), "content": as_text(turn.get("content"))})
-        else:
-            messages.append({"role": "user", "content": as_text(turn[0])})
-            if turn[1]:
-                messages.append({"role": "assistant", "content": as_text(turn[1])})
-    messages.append({"role": "user", "content": as_text(message)})
-    try:
-        client = InferenceClient(token=oauth_token.token)
-        out = client.chat_completion(model=CHAT_MODEL, messages=messages, max_tokens=700)
-        return out.choices[0].message.content.strip()
-    except Exception as e:
-        name = type(e).__name__
-        msg = str(e)
-        if "credits" in msg.lower() or "quota" in msg.lower() or "402" in msg:
-            return "Your Hugging Face inference credits are used up for now. Please try again later."
-        return f"The chat service could not be reached ({name}). Please try again in a moment."
-
-
 # ---------------------------------------------------------------- The buttons
 
 @spaces.GPU(duration=gpu_seconds(7), size=GPU_SIZE)
@@ -829,7 +721,7 @@ def on_open():
 
 with gr.Blocks(title="Graphics Studio") as demo:
     gr.Markdown("# Graphics Studio\n**FLUX.2 klein** images, **LTX-Video** animation, **MusicGen** music, "
-                "**Qwen3** + **Kokoro** narration, **Whisper** transcription, **Gemma 4 31B** chat, "
+                "**Qwen3** + **Kokoro** narration, **Whisper** transcription, "
                 "**Florence-2** image captions, **SmolVLM2** video descriptions, **CLAP** music classification.")
     counter = gr.Markdown(counts_text())
 
@@ -882,16 +774,6 @@ with gr.Blocks(title="Graphics Studio") as demo:
         read_button = gr.Button("Read aloud", variant="primary")
         read_out = gr.Audio(label="Narration", type="filepath", autoplay=False)
 
-    with gr.Tab("Chat"):
-        if chat_model is not None:
-            gr.Markdown("Chat with **Gemma 4 31B** (google), running here in 8-bit. No sign-in needed. "
-                        "A reply takes a while and uses your own ZeroGPU quota.")
-        else:
-            gr.Markdown("Chat with hosted **Gemma 4 31B** (google). Sign in with Hugging Face first — each visitor's "
-                        "reply is billed to their own Hugging Face inference credits, not the owner's.")
-            gr.LoginButton()
-        gr.ChatInterface(chat)
-
     with gr.Tab("Describe"):
         gr.Markdown("Three ways to describe what you have: an image (image captioning), a short video "
                     "(video captioning), or a piece of music (music captioning, music classification). "

requirements.txt

@@ -15,4 +15,3 @@ huggingface_hub>=0.24.0
 timm
 einops
 num2words
-bitsandbytes>=0.46.0