fix(space): sync spaces.GPU ZeroGPU integration and 6.30.0 SDK

SPACE_README.md

@@ -4,23 +4,30 @@ emoji: ⚛️
 colorFrom: indigo
 colorTo: purple
 sdk: gradio
-sdk_version: 4.44.0
+sdk_version: 6.30.0
+python_version: '3.12'
 app_file: app.py
 pinned: false
 license: apache-2.0
-short_description: Information-Value Adaptive Transformer with live rank slicing & telemetry
+short_description: Information-Value-Driven Adaptive Tensor-Network Transformer
 ---
 
 # ⚛️ Q-TensorFormer: Interactive Showcase
 
-Interactive demonstration of **Q-TensorFormer**, an information-value-driven adaptive Transformer architecture.
+[![Documentation](https://img.shields.io/badge/Docs-premchan369.github.io%2FQ--TensorFormer-blue)](https://premchan369.github.io/Q-TensorFormer/)
+[![Model Hub](https://img.shields.io/badge/%F0%9F%A4%97%20Model%20Hub-Premchan369%2FQ--TensorFormer-yellow)](https://huggingface.co/Premchan369/Q-TensorFormer)
+[![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/Premchan369/Q-TensorFormer/blob/main/notebooks/quickstart.ipynb)
+[![License: Apache 2.0](https://img.shields.io/badge/License-Apache%202.0-blue.svg)](LICENSE)
+[![Tests: 94 Passed](https://img.shields.io/badge/Pytest-94%20Passed%20%7C%200%20Failed-brightgreen)](https://huggingface.co/Premchan369/Q-TensorFormer)
 
-- **Canonical Model Hub:** https://huggingface.co/Premchan369/Q-TensorFormer
-- **Technical Documentation:** https://premchan369.github.io/Q-TensorFormer/
-- **Author:** Premchand Yadav
-- **License:** Apache-2.0
+Welcome to the live demonstration of **Q-TensorFormer**, an information-value-driven adaptive Transformer architecture developed by **Premchand Yadav**.
 
-### Features in this Space:
-1. **Live Dynamic Tensor-Train Slicing:** Observe token-by-token active rank adjustments ($r \in [1, 16]$) based on token complexity and device pressure.
-2. **Adaptive KV Cache Telemetry:** Inspect real-time memory savings with attention-sink preservation and INT8/INT4 quantization.
-3. **Closed-Loop PID Control:** Adjust target energy budgets and watch shadow prices regulate computational capacity in real time.
+### What This Space Demonstrates:
+1. **💬 Live Dynamic Tensor-Train Slicing:** Type any prompt to watch the model slice Tensor-Train rank cores ($r \in [1, 16]$) token-by-token based on marginal informational value and target energy budgets.
+2. **🎨 Token Complexity Heatmap:** Visualizes which tokens required full matrix rank vs. lightweight edge rank.
+3. **📊 Adaptive KV-Cache Telemetry:** Real-time DRAM footprint reduction (up to **93.36% memory saved**) using streaming attention-sink protection and INT8/INT4 quantization.
+4. **👁️ Multimodal Vision Projection:** Tensor-train projected visual tokens.
+5. **⚡ LLM Compression Studio:** Interactive parameter savings calculator for open-weight models (Llama-3, Qwen, Mistral).
+
+- **Official Repository:** [https://huggingface.co/Premchan369/Q-TensorFormer](https://huggingface.co/Premchan369/Q-TensorFormer)
+- **Technical Documentation Portal:** [https://premchan369.github.io/Q-TensorFormer/](https://premchan369.github.io/Q-TensorFormer/)

app.py

@@ -21,6 +21,18 @@ import torch
 from configuration_qtensorformer import QTensorFormerConfig
 from modeling_qtensorformer import QTensorFormerForCausalLM
 
+try:
+    import spaces
+    HAS_SPACES = True
+except ImportError:
+    HAS_SPACES = False
+
+
+def gpu_decorator(fn):
+    if HAS_SPACES:
+        return spaces.GPU(fn)
+    return fn
+
 try:
     import gradio as gr
     HAS_GRADIO = True
@@ -38,13 +50,13 @@ if tok_path.exists():
 else:
     TOKENIZER = QTensorFormerTokenizer()
 
-# Load trained canonical checkpoint if present, with graceful fallback
-DEVICE = "cuda" if torch.cuda.is_available() else "cpu"
+# Initialize model on CPU for instant cold-start and ZeroGPU handshake safety
+MODEL_DEVICE = "cpu"
 trained_ckpt_path = Path(__file__).resolve().parent / "checkpoints" / "hf_canonical_trained"
 
 if trained_ckpt_path.exists():
     try:
-        MODEL = QTensorFormerForCausalLM.from_pretrained(str(trained_ckpt_path)).to(DEVICE)
+        MODEL = QTensorFormerForCausalLM.from_pretrained(str(trained_ckpt_path)).to(MODEL_DEVICE)
         CONFIG = MODEL.config
     except Exception as e:
         print(f"[app.py] Notice during checkpoint loading: {e}. Falling back to dynamic initialization.")
@@ -60,7 +72,7 @@ if trained_ckpt_path.exists():
             attention_sink_size=4,
             kv_cache_window_size=128,
         )
-        MODEL = QTensorFormerForCausalLM(CONFIG).to(DEVICE)
+        MODEL = QTensorFormerForCausalLM(CONFIG).to(MODEL_DEVICE)
 else:
     CONFIG = QTensorFormerConfig(
         vocab_size=256,
@@ -74,64 +86,73 @@ else:
         attention_sink_size=4,
         kv_cache_window_size=128,
     )
-    MODEL = QTensorFormerForCausalLM(CONFIG).to(DEVICE)
+    MODEL = QTensorFormerForCausalLM(CONFIG).to(MODEL_DEVICE)
 
 MODEL.eval()
 
 
+@gpu_decorator
 def generate_with_telemetry(prompt: str, max_tokens: int = 64, temperature: float = 0.7, target_budget: float = 1.0):
     if not prompt.strip():
         yield "Please enter a valid text prompt.", "Rank 8 (Balanced)", "0.0 mJ", "0.0 tok/s", []
         return
 
-    # Update PID budget setpoint
-    MODEL.model.pid_controller.target_budget = target_budget
-    MODEL.model.pid_controller.reset_state()
-
-    # Encode with deterministic QTensorFormerTokenizer
-    tokens = TOKENIZER.encode(prompt)
-    if not tokens:
-        tokens = [TOKENIZER.bos_id]
-    curr_ids = torch.tensor([tokens], dtype=torch.long, device=DEVICE)
-
-    generated_text = prompt
-    token_ranks = []
-    past_key_values = None
-
-    t0 = time.perf_counter()
-
-    for step in range(max_tokens):
-        with torch.no_grad():
-            model_in = curr_ids if step == 0 else curr_ids[:, -1:]
-            outputs = MODEL(model_in, past_key_values=past_key_values, use_cache=True)
-            logits = outputs.logits[:, -1, :]
-            past_key_values = outputs.past_key_values
-
-            if temperature > 0:
-                probs = torch.softmax(logits / max(1e-4, temperature), dim=-1)
-                next_tok = torch.multinomial(probs, num_samples=1)
-            else:
-                next_tok = logits.argmax(dim=-1, keepdim=True)
-
-        curr_ids = torch.cat([curr_ids, next_tok], dim=1)
-        tok_id = int(next_tok[0, 0].item())
-        tok_str = TOKENIZER.decode([tok_id])
-        generated_text += tok_str
-
-        # Read live controller rank
-        pid_ctrl = MODEL.model.pid_controller
-        active_rank = int(pid_ctrl.rank_levels[int(pid_ctrl.active_rank_idx.item())])
-        token_ranks.append((tok_str, f"Rank {active_rank}"))
-
-        elapsed = time.perf_counter() - t0
-        tok_per_sec = (step + 1) / max(1e-6, elapsed)
-        est_mj_per_tok = (active_rank / 16.0) * 1.8 + 0.4  # mJ estimate proportional to TT-rank
-
-        rank_display = f"Rank {active_rank} ({'Full' if active_rank == 16 else 'Compressed' if active_rank == 8 else 'Ultra-Edge'})"
-        energy_display = f"{est_mj_per_tok:.2f} mJ/tok"
-        speed_display = f"{tok_per_sec:.1f} tokens/sec"
-
-        yield generated_text, rank_display, energy_display, speed_display, token_ranks
+    run_device = "cuda" if torch.cuda.is_available() else "cpu"
+    try:
+        if run_device == "cuda":
+            MODEL.to("cuda")
+
+        # Update PID budget setpoint
+        MODEL.model.pid_controller.target_budget = target_budget
+        MODEL.model.pid_controller.reset_state()
+
+        # Encode with deterministic QTensorFormerTokenizer
+        tokens = TOKENIZER.encode(prompt)
+        if not tokens:
+            tokens = [TOKENIZER.bos_id]
+        curr_ids = torch.tensor([tokens], dtype=torch.long, device=run_device)
+
+        generated_text = prompt
+        token_ranks = []
+        past_key_values = None
+
+        t0 = time.perf_counter()
+
+        for step in range(max_tokens):
+            with torch.no_grad():
+                model_in = curr_ids if step == 0 else curr_ids[:, -1:]
+                outputs = MODEL(model_in, past_key_values=past_key_values, use_cache=True)
+                logits = outputs.logits[:, -1, :]
+                past_key_values = outputs.past_key_values
+
+                if temperature > 0:
+                    probs = torch.softmax(logits / max(1e-4, temperature), dim=-1)
+                    next_tok = torch.multinomial(probs, num_samples=1)
+                else:
+                    next_tok = logits.argmax(dim=-1, keepdim=True)
+
+            curr_ids = torch.cat([curr_ids, next_tok], dim=1)
+            tok_id = int(next_tok[0, 0].item())
+            tok_str = TOKENIZER.decode([tok_id])
+            generated_text += tok_str
+
+            # Read live controller rank
+            pid_ctrl = MODEL.model.pid_controller
+            active_rank = int(pid_ctrl.rank_levels[int(pid_ctrl.active_rank_idx.item())])
+            token_ranks.append((tok_str, f"Rank {active_rank}"))
+
+            elapsed = time.perf_counter() - t0
+            tok_per_sec = (step + 1) / max(1e-6, elapsed)
+            est_mj_per_tok = (active_rank / 16.0) * 1.8 + 0.4  # mJ estimate proportional to TT-rank
+
+            rank_display = f"Rank {active_rank} ({'Full' if active_rank == 16 else 'Compressed' if active_rank == 8 else 'Ultra-Edge'})"
+            energy_display = f"{est_mj_per_tok:.2f} mJ/tok"
+            speed_display = f"{tok_per_sec:.1f} tokens/sec"
+
+            yield generated_text, rank_display, energy_display, speed_display, token_ranks
+    finally:
+        if run_device == "cuda":
+            MODEL.to("cpu")
 
 
 def mock_vision_qa(image, question: str):
@@ -174,7 +195,7 @@ def build_app():
     .kpi-box { padding: 15px; border-radius: 8px; background: #f8fafc; border: 1px solid #e2e8f0; text-align: center; }
     """
 
-    with gr.Blocks(title="Q-TensorFormer: Closed-Loop Adaptive LLM", css=custom_css) as demo:
+    with gr.Blocks(title="Q-TensorFormer: Closed-Loop Adaptive LLM") as demo:
         gr.Markdown(
             """
             # ⚛️ Q-TensorFormer: Closed-Loop Information-Adaptive LLM
@@ -287,7 +308,7 @@ def build_app():
 # Instantiate demo at module level for Hugging Face Spaces & Gradio discovery
 demo = build_app()
 
-if __name__ == "__main__":
+if __name__ == "__main__" or os.environ.get("SPACE_ID"):
     if demo is not None:
         demo.queue().launch(server_name="0.0.0.0", server_port=7860)