Upload 129 files

README.md

@@ -6,6 +6,7 @@ colorTo: gray
 sdk: gradio
 sdk_version: "5.29.0"
 app_file: app.py
+persistent_storage: true
 pinned: true
 license: agpl-3.0
 tags:

app.py

@@ -106,6 +106,9 @@ _session_models: dict[str, dict] = {}
 # Legacy alias — some internal code may still reference _bench_configs
 _bench_configs = _session_models
 
+# Label of the most recently obliterated model (for auto-selecting in Chat tab dropdown)
+_last_obliterated_label: str = ""
+
 # Counter for unique obliteration save directories
 _obliterate_counter: int = 0
 
@@ -1564,9 +1567,11 @@ def obliterate(model_choice: str, method_choice: str, hub_repo: str,
             pass  # Telemetry is best-effort
 
         # ── Session cache: register this obliteration for Chat tab switching ──
+        global _last_obliterated_label
         _ts = datetime.now().strftime("%H:%M")
         _short_model = model_id.split("/")[-1] if "/" in model_id else model_id
         _cache_label = f"{method} on {_short_model} ({_ts})"
+        _last_obliterated_label = _cache_label
         _session_models[_cache_label] = {
             "model_id": model_id,
             "model_choice": model_choice,
@@ -1792,6 +1797,16 @@ def chat_respond(message: str, history: list[dict], system_prompt: str,
         yield "No model loaded yet. Go to the **Obliterate** tab first and liberate a model."
         return
 
+    # ZeroGPU safety: ensure model is on GPU if available.
+    # Between GPU allocations, ZeroGPU may have moved the model to CPU/meta.
+    # The @spaces.GPU decorator guarantees a GPU is available here, so move if needed.
+    try:
+        dev = next(model.parameters()).device
+        if torch.cuda.is_available() and dev.type != "cuda":
+            model.to("cuda")
+    except (StopIteration, RuntimeError):
+        pass
+
     # Sanitize inputs to prevent resource exhaustion
     system_prompt = (system_prompt or "")[:4096]
     message = (message or "")[:8192]
@@ -1916,8 +1931,8 @@ def load_bench_into_chat(choice: str, progress=gr.Progress()):
 
     On ZeroGPU, uses the visitor's GPU quota.
     """
-    if choice not in _bench_configs:
-        yield "**Error:** No benchmark result selected.", ""
+    if not choice or choice not in _bench_configs:
+        yield "**Error:** No benchmark result selected. Pick a model from the dropdown first.", ""
         return
 
     cfg = _bench_configs[choice]
@@ -1925,6 +1940,18 @@ def load_bench_into_chat(choice: str, progress=gr.Progress()):
     method_key = cfg["method"]
     checkpoint_dir = cfg.get("output_dir")
 
+    # If this model is already the active one, skip the destructive reload
+    with _lock:
+        if (_state["status"] == "ready"
+                and _state["model"] is not None
+                and _state["model_name"] == cfg.get("model_choice", "")
+                and _state["method"] == method_key):
+            yield (
+                f"**Already loaded!** `{choice}` is ready — just type in the chat below.",
+                get_chat_header(),
+            )
+            return
+
     with _lock:
         if _state["status"] == "obliterating":
             yield "**Error:** An obliteration is already in progress.", ""
@@ -2105,6 +2132,14 @@ def ab_chat_respond(message: str, history_left: list[dict], history_right: list[
                "#### Abliterated")
         return
 
+    # ZeroGPU safety: ensure model is on GPU if available
+    try:
+        dev = next(abliterated_model.parameters()).device
+        if torch.cuda.is_available() and dev.type != "cuda":
+            abliterated_model.to("cuda")
+    except (StopIteration, RuntimeError):
+        pass
+
     # Build header strings showing model name on each side
     header_left = f"#### Original (Pre-Abliteration)\n`{model_name}`"
     header_right = f"#### Abliterated\n`{model_name}`"
@@ -2254,9 +2289,12 @@ def ab_chat_respond(message: str, history_left: list[dict], history_right: list[
     except Exception as e:
         original_response = f"*Could not load original model for comparison: {e}*"
 
-    # Restore abliterated model to GPU for subsequent chat/operations
+    # Restore abliterated model to GPU for subsequent chat/operations.
+    # Use torch.device("cuda") rather than the captured abl_device, since
+    # on ZeroGPU the original device reference may point to a stale context.
     try:
-        abliterated_model.to(abl_device)
+        restore_device = torch.device("cuda") if torch.cuda.is_available() else abl_device
+        abliterated_model.to(restore_device)
     except Exception:
         pass  # If GPU restore fails, model stays on CPU (still usable)
 
@@ -2747,7 +2785,7 @@ button.tab-nav.selected {
 
 /* ---- CARD-STYLE BLOCKS ---- */
 .gr-panel, .gr-box, .gr-form, .gr-group,
-div.block { position: relative; }
+div.block { position: relative; padding-left: 10px !important; }
 div.block::before {
     content: '';
     position: absolute;
@@ -3903,7 +3941,10 @@ Built on the shoulders of:
                 custom_harmful_tb, custom_harmless_tb] + _adv_controls,
         outputs=[status_md, log_box, chat_status],
     ).then(
-        fn=lambda: (gr.update(choices=_get_session_model_choices()), _get_vram_html()),
+        fn=lambda: (
+            gr.update(choices=_get_session_model_choices(), value=_last_obliterated_label or None),
+            _get_vram_html(),
+        ),
         outputs=[session_model_dd, vram_display],
     )
 

hf-spaces/README.md

@@ -7,6 +7,7 @@ sdk: gradio
 sdk_version: "5.29.0"
 app_file: app.py
 hardware: zero-a10g
+persistent_storage: true
 pinned: true
 license: agpl-3.0
 tags:

obliteratus/cli.py

@@ -448,12 +448,36 @@ def _cmd_abliterate(args):
             live.update(make_display())
             raise
 
+    # ── Telemetry: send pipeline report to community leaderboard ──
+    try:
+        from obliteratus.telemetry import maybe_send_pipeline_report
+        maybe_send_pipeline_report(pipeline)
+    except Exception:
+        pass  # Telemetry is best-effort
+
+    # ── Community contribution (--contribute flag) ──
+    contrib_path = None
+    if getattr(args, "contribute", False):
+        try:
+            from obliteratus.community import save_contribution
+            contrib_path = save_contribution(
+                pipeline,
+                model_name=model_name,
+                notes=getattr(args, "contribute_notes", ""),
+            )
+        except Exception as e:
+            console.print(f"[yellow]Could not save contribution: {e}[/yellow]")
+
     console.print()
+    contrib_line = ""
+    if contrib_path:
+        contrib_line = f"\n  Contribution: [cyan]{contrib_path}[/]"
     console.print(
         Panel(
             f"[bold green]Abliteration complete![/]\n\n"
             f"  Model saved to: [cyan]{result_path}[/]\n"
-            f"  Metadata: [cyan]{result_path}/abliteration_metadata.json[/]\n\n"
+            f"  Metadata: [cyan]{result_path}/abliteration_metadata.json[/]"
+            f"{contrib_line}\n\n"
             f"  [dim]Load with:[/] AutoModelForCausalLM.from_pretrained('{result_path}')",
             border_style="green",
             title="[bold green]✓ REBIRTH COMPLETE[/]",