Update app.py

app.py

@@ -18,15 +18,14 @@ def convert_and_upload(token, source_repo, target_repo, precision, target_compon
         yield "❌ Error: Please select at least one component to quantize."
         return
 
-    # Map precision string to PyTorch dtype
-    if precision == "FP8":
-        target_dtype = torch.float8_e4m3fn
-    elif precision == "FP16":
-        target_dtype = torch.float16
-    elif precision == "BF16":
-        target_dtype = torch.bfloat16
-    else:
-        target_dtype = None
+    # Map precision string to PyTorch dtype or INT8 flag
+    target_dtype = None
+    is_int8 = False
+    
+    if precision == "FP8": target_dtype = torch.float8_e4m3fn
+    elif precision == "FP16": target_dtype = torch.float16
+    elif precision == "BF16": target_dtype = torch.bfloat16
+    elif precision == "INT8": is_int8 = True
 
     api = HfApi(token=token)
     yield f"🔄 Connecting to Hugging Face and verifying target repo: {target_repo}..."
@@ -44,21 +43,21 @@ def convert_and_upload(token, source_repo, target_repo, precision, target_compon
         yield f"❌ Error fetching files: {str(e)}"
         return
 
-    # Create a unique cache directory for this specific run to prevent collisions
+    # Unique cache directory to prevent collisions
     cache_dir = f"./hf_cache_{uuid.uuid4().hex[:8]}"
-    
     success_count = 0
     error_count = 0
 
+    # Exclusion list for INT8 (highly sensitive layers)
+    exclude_prefixes = ["norm", "ln_", "embed", "time_in", "vector_in", "guidance_in", "txt_in", "img_in"]
+
     for file in files:
-        # AUTO-DELETE/SKIP LOGIC: Detect large .safetensors files at the root level
         is_root_safetensor = "/" not in file and file.endswith(".safetensors")
         
         if is_root_safetensor:
             yield f"🗑️ Auto-skipping massive root model: {file}..."
             try:
-                api.delete_file(path_in_repo=file, repo_id=target_repo, token=token, commit_message=f"Auto-deleted massive root file {file}")
-                yield f"✅ Ensured {file} is removed from target repository."
+                api.delete_file(path_in_repo=file, repo_id=target_repo, token=token, commit_message=f"Auto-deleted {file}")
             except Exception:
                 pass 
             continue
@@ -68,7 +67,6 @@ def convert_and_upload(token, source_repo, target_repo, precision, target_compon
         try:
             os.makedirs(cache_dir, exist_ok=True)
 
-            # CRITICAL FIX: Added token=token here so gated FLUX models don't block the download
             local_path = hf_hub_download(
                 repo_id=source_repo,
                 filename=file,
@@ -79,21 +77,46 @@ def convert_and_upload(token, source_repo, target_repo, precision, target_compon
             in_target_component = any(f"{comp}/" in file for comp in target_components)
 
             if file.endswith(".safetensors") and in_target_component:
-                yield f"🧠 Quantizing {file} to {precision} (This will take a few minutes)..."
+                yield f"🧠 Quantizing {file} to {precision}..."
                 
                 tensors = load_file(local_path)
-
-                if target_dtype:
-                    keys = list(tensors.keys())
-                    for k in keys:
-                        if tensors[k].is_floating_point():
-                            tensors[k] = tensors[k].to(target_dtype)
+                new_tensors = {}
+
+                for k, v in tensors.items():
+                    # --- BRANCH 1: INT8 Symmetric Quantization ---
+                    if is_int8:
+                        is_2d_weight = "weight" in k and len(v.shape) == 2
+                        is_excluded = any(ex in k for ex in exclude_prefixes)
+
+                        if is_2d_weight and not is_excluded:
+                            # Upcast to BF16 for math
+                            if v.dtype == torch.float8_e4m3fn:
+                                v = v.to(torch.bfloat16)
+
+                            scale = v.abs().max(dim=1, keepdim=True)[0] / 127.0
+                            scale = scale.clamp(min=1e-8)
+                            weight_int8 = torch.round(v / scale).clamp(-127, 127).to(torch.int8)
+                            
+                            base_name = k.rsplit(".", 1)[0]
+                            new_tensors[f"{base_name}.weight_int8"] = weight_int8
+                            new_tensors[f"{base_name}.weight_scale"] = scale.to(torch.bfloat16)
+                        else:
+                            # Excluded layers / 1D tensors pass through as BF16
+                            new_tensors[k] = v.to(torch.bfloat16) if v.is_floating_point() else v
+
+                    # --- BRANCH 2: Standard Floating Point Casting ---
+                    else:
+                        if v.is_floating_point():
+                            new_tensors[k] = v.to(target_dtype)
+                        else:
+                            new_tensors[k] = v
 
                 converted_path = "converted.safetensors"
-                save_file(tensors, converted_path)
+                save_file(new_tensors, converted_path)
 
                 # Aggressive memory flush
                 del tensors
+                del new_tensors
                 gc.collect()
 
                 yield f"☁️ Uploading {precision} version of {file}..."
@@ -117,92 +140,73 @@ def convert_and_upload(token, source_repo, target_repo, precision, target_compon
 
             success_count += 1
 
-            # EXTREME DISK CLEANUP: Nuke the cache directory after every file to prevent the 50GB Space Crash
+            # Disk cleanup
             if os.path.exists(cache_dir):
                 shutil.rmtree(cache_dir)
-            
             gc.collect()
 
         except Exception as e:
             error_count += 1
-            yield f"⚠️ Error processing {file}: {str(e)}\nSkipping to next file..."
+            yield f"⚠️ Error processing {file}: {str(e)}\nSkipping..."
 
-    # Final cleanup sweep
     if os.path.exists(cache_dir):
         shutil.rmtree(cache_dir)
 
     yield f"✅ Finished! Successfully processed {success_count} files. Errors encountered: {error_count}."
 
-# Dynamic UI Update for Target Repo Name
+# UI Updates
 def update_target_repo(username, source, precision):
     user_prefix = username.strip() if username.strip() else "your-username"
     model_name = source.split("/")[-1] if "/" in source else source
     return f"{user_prefix}/{model_name}-{precision}"
 
-# Build the Gradio UI
+def update_warnings(precision):
+    if precision == "INT8":
+        return gr.update(value="⚠️ **INT8 Warning:** Modifies layer keys (`weight_int8`, `weight_scale`). Requires custom inference code to run.", visible=True)
+    else:
+        return gr.update(visible=False)
+
 with gr.Blocks(theme=gr.themes.Soft()) as demo:
-    gr.Markdown("# 🚀 FLUX.2-klein Dedicated Quantizer")
+    gr.Markdown("# 🚀 Universal FLUX.2-klein Quantizer")
     gr.Markdown(
-        "Convert sharded **FLUX.2-klein** models (4B and 9B) to lower precisions (FP8, FP16, BF16).\n\n"
-        "**Auto-Delete & Disk Protection:** This tool actively purges Hugging Face's download cache after every single shard. "
-        "This ensures the 9B model won't crash the free Space by filling up the 50GB hard drive limit."
+        "Convert sharded models to floating-point precisions (FP8/FP16/BF16) or dynamically trigger symmetric integer quantization (INT8)."
     )
 
     with gr.Row():
         with gr.Column(scale=2):
-            hf_token = gr.Textbox(
-                label="Hugging Face Token (Write Access Required)", 
-                type="password",
-                placeholder="hf_..."
-            )
-            hf_username = gr.Textbox(
-                label="Your Hugging Face Username", 
-                placeholder="e.g., rootlocalghost"
-            )
+            hf_token = gr.Textbox(label="Hugging Face Token (Write Access)", type="password", placeholder="hf_...")
+            hf_username = gr.Textbox(label="Hugging Face Username", placeholder="e.g., rootlocalghost")
             source_repo = gr.Dropdown(
-                choices=[
-                    "black-forest-labs/FLUX.2-klein-9B", 
-                    "black-forest-labs/FLUX.2-klein-4B"
-                ], 
-                value="black-forest-labs/FLUX.2-klein-9B", 
-                label="Source Repository",
-                allow_custom_value=False
+                choices=["black-forest-labs/FLUX.2-klein-9B", "black-forest-labs/FLUX.2-klein-4B"], 
+                value="black-forest-labs/FLUX.2-klein-9B", label="Source Repository", allow_custom_value=True
             )
             
             target_components = gr.CheckboxGroup(
                 choices=["text_encoder", "transformer", "vae"],
-                value=["text_encoder", "transformer"],
-                label="Components to Quantize",
-                info="Select which folders should be cast to the new precision. Unselected folders will be copied as-is."
+                value=["transformer"],
+                label="Components to Quantize"
             )
             
             precision = gr.Dropdown(
-                choices=["FP8", "FP16", "BF16"], 
-                value="FP8", 
+                choices=["FP8", "FP16", "BF16", "INT8"], 
+                value="INT8", 
                 label="Target Precision"
             )
-            target_repo = gr.Textbox(
-                label="Target Repository (Auto-generated)", 
-                value="your-username/FLUX.2-klein-9B-FP8",
-                interactive=True
-            )
+            
+            int8_warning = gr.Markdown(visible=True, value="⚠️ **INT8 Warning:** Modifies layer keys (`weight_int8`, `weight_scale`). Requires custom inference code to run.")
+            
+            target_repo = gr.Textbox(label="Target Repository (Auto-generated)", value="your-username/FLUX.2-klein-9B-INT8", interactive=True)
             start_btn = gr.Button("Start Quantization & Upload", variant="primary")
         
         with gr.Column(scale=3):
-            output_log = gr.Textbox(
-                label="Operation Logs", 
-                lines=20, 
-                interactive=False,
-                max_lines=25
-            )
+            output_log = gr.Textbox(label="Operation Logs", lines=20, interactive=False, max_lines=25)
 
+    # Dynamic UI wiring
     inputs_to_watch = [hf_username, source_repo, precision]
     for inp in inputs_to_watch:
-        inp.change(
-            fn=update_target_repo, 
-            inputs=inputs_to_watch, 
-            outputs=[target_repo]
-        )
+        inp.change(fn=update_target_repo, inputs=inputs_to_watch, outputs=[target_repo])
+        
+    precision.change(fn=update_warnings, inputs=[precision], outputs=[int8_warning])
 
     start_btn.click(
         fn=convert_and_upload,