Publish riemann-prerelease-demo

README.md

@@ -1,13 +1,32 @@
----
-title: My Name Is Riemann Prerelease Basic Demo
-emoji: 📚
-colorFrom: pink
-colorTo: gray
-sdk: gradio
-sdk_version: 6.30.0
-python_version: '3.12'
-app_file: app.py
-pinned: false
----
-
-Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference
+---
+title: My Name is Riemann Prerelease basic demo
+emoji: 🪜
+colorFrom: indigo
+colorTo: gray
+sdk: gradio
+sdk_version: 5.49.1
+app_file: app.py
+pinned: false
+license: other
+license_name: opnl-2-and-mit
+license_link: https://huggingface.co/OpenPeerAI/My-Name-is-Riemann-Prerelease/blob/main/LICENSE
+models:
+  - OpenPeerAI/My-Name-is-Riemann-Prerelease
+  - OpenPeerAI/OpenPeerLLM
+short_description: Chat with the 13.1B My Name is Riemann Prerelease on ZeroGPU
+---
+
+# My Name is Riemann Prerelease basic demo
+
+Interactive text generation with the stage 6 checkpoint of
+[My Name is Riemann](https://huggingface.co/OpenPeerAI/My-Name-is-Riemann-Prerelease),
+a 13.1 billion parameter model grown from OpenPeerLLM's published weights and
+fine-tuned with the OpenPeer-AI-NTK-Trainer gate controller. The toggle in the
+sidebar attaches or removes the controller so the two can be compared on the
+same prompt.
+
+The model loads in bf16 and runs on a ZeroGPU slot. Expect a cold start of a
+minute or two the first time a GPU is assigned.
+
+Developed by OpenPeer AI, Riemann Computing Inc., and Andrew Magdy Kamal.
+Released under the Open Innovation License v2 and MIT.

app.py

@@ -0,0 +1,174 @@
+from __future__ import annotations
+import os
+import threading
+import torch
+import gradio as gr
+import spaces
+from huggingface_hub import hf_hub_download
+from transformers import AutoTokenizer, AutoModelForCausalLM, TextIteratorStreamer
+
+MODEL_REPO = os.environ.get("MODEL_REPO", "OpenPeerAI/My-Name-is-Riemann-Prerelease")
+CONTROLLER_PATH_IN_REPO = "ntk_controller/stage6_12.85b.pt"
+HF_TOKEN = os.environ.get("HF_TOKEN")
+
+tokenizer = AutoTokenizer.from_pretrained(MODEL_REPO, token=HF_TOKEN)
+if tokenizer.pad_token is None:
+    tokenizer.pad_token = tokenizer.eos_token
+model = AutoModelForCausalLM.from_pretrained(
+    MODEL_REPO,
+    trust_remote_code=True,
+    torch_dtype=torch.bfloat16,
+    token=HF_TOKEN,
+    low_cpu_mem_usage=True,
+)
+model.eval()
+
+_controller = {"tuner": None, "attached": False, "error": None}
+
+
+def _ensure_controller():
+    if _controller["tuner"] is not None or _controller["error"]:
+        return
+    try:
+        from ntkmirror import ForwardFineTuner
+
+        path = hf_hub_download(MODEL_REPO, CONTROLLER_PATH_IN_REPO, token=HF_TOKEN)
+        tuner = ForwardFineTuner(model, tokenizer)
+        tuner.load(path, allow_model_mismatch=True, map_location="cpu")
+        _controller["tuner"] = tuner
+    except Exception as exc:
+        _controller["error"] = str(exc)
+
+
+def _set_controller(enabled: bool):
+    _ensure_controller()
+    tuner = _controller["tuner"]
+    if tuner is None:
+        return
+    if enabled and not _controller["attached"]:
+        tuner.controller.to(model.device)
+        tuner.controller.attach()
+        _controller["attached"] = True
+    elif not enabled and _controller["attached"]:
+        tuner.controller.remove()
+        _controller["attached"] = False
+
+
+def _format_prompt(message: str, history: list[dict], system: str) -> str:
+    parts = []
+    if system.strip():
+        parts.append(f"system: {system.strip()}")
+    for turn in history:
+        role = turn.get("role", "user")
+        content = turn.get("content", "")
+        if content:
+            parts.append(f"{role}: {content}")
+    parts.append(f"user: {message.strip()}")
+    parts.append("assistant:")
+    return "\n".join(parts)
+
+
+@spaces.GPU(duration=120)
+def generate(
+    message,
+    history,
+    system,
+    use_controller,
+    max_new_tokens,
+    temperature,
+    top_p,
+    repetition_penalty,
+):
+    model.to("cuda")
+    _set_controller(bool(use_controller))
+    prompt = _format_prompt(message, history, system)
+    inputs = tokenizer(
+        prompt,
+        return_tensors="pt",
+        truncation=True,
+        max_length=2048 - int(max_new_tokens),
+    ).to("cuda")
+    streamer = TextIteratorStreamer(
+        tokenizer, skip_prompt=True, skip_special_tokens=True
+    )
+    kwargs = dict(
+        **inputs,
+        streamer=streamer,
+        max_new_tokens=int(max_new_tokens),
+        do_sample=temperature > 0,
+        temperature=max(float(temperature), 1e-3),
+        top_p=float(top_p),
+        repetition_penalty=float(repetition_penalty),
+        pad_token_id=tokenizer.eos_token_id,
+        eos_token_id=tokenizer.eos_token_id,
+    )
+    thread = threading.Thread(target=model.generate, kwargs=kwargs)
+    thread.start()
+    out = ""
+    for piece in streamer:
+        out += piece
+        if "\nuser:" in out:
+            out = out.split("\nuser:")[0]
+            break
+        yield out
+    thread.join()
+    yield out.strip()
+
+
+DESCRIPTION = f"""
+# My Name is Riemann Prerelease
+
+A 13.1 billion parameter decoder-only model grown rung by rung from
+[OpenPeerAI/OpenPeerLLM](https://huggingface.co/OpenPeerAI/OpenPeerLLM) and
+fine-tuned with the [OpenPeer-AI-NTK-Trainer](https://huggingface.co/OpenPeerAI/OpenPeer-AI-NTK-Trainer)
+gate controller instead of LoRA. Weights load in bf16 on ZeroGPU.
+
+This is a prerelease trained on about 2.2 billion tokens. It is not an aligned assistant;
+treat output as a draft. The NTK toggle attaches the 6,732-gate controller fitted on the frozen weights.
+
+Model card and weights: [{MODEL_REPO}](https://huggingface.co/{MODEL_REPO})
+"""
+
+with gr.Blocks(title="My Name is Riemann Prerelease") as demo:
+    gr.Markdown(DESCRIPTION)
+    with gr.Row():
+        with gr.Column(scale=3):
+            chat = gr.ChatInterface(
+                fn=generate,
+                type="messages",
+                additional_inputs=[
+                    gr.Textbox(
+                        value="You are My Name is Riemann, a careful coding and reasoning assistant. Think step by step and show your work.",
+                        label="System prompt",
+                        lines=2,
+                    ),
+                    gr.Checkbox(value=True, label="Attach NTK gate controller"),
+                    gr.Slider(16, 512, value=256, step=16, label="Max new tokens"),
+                    gr.Slider(
+                        0.0, 1.5, value=0.7, step=0.05, label="Temperature (0 = greedy)"
+                    ),
+                    gr.Slider(0.1, 1.0, value=0.9, step=0.05, label="Top-p"),
+                    gr.Slider(
+                        1.0, 1.5, value=1.1, step=0.05, label="Repetition penalty"
+                    ),
+                ],
+                examples=[
+                    [
+                        "Write a Python function that returns the n-th prime, with a short explanation."
+                    ],
+                    [
+                        "Explain the difference between a mutex and a semaphore in two paragraphs."
+                    ],
+                    ["اشرح نظرية فيثاغورس بإيجاز."],
+                    [
+                        "Refactor this to be idempotent:\n\ndef add_user(db, name):\n    db.append({'name': name})"
+                    ],
+                ],
+                cache_examples=False,
+            )
+    gr.Markdown(
+        "Licensed under the Open Innovation License v2 (OPNL-2) and MIT. "
+        "Developed by OpenPeer AI, Riemann Computing Inc., and Andrew Magdy Kamal."
+    )
+
+demo.queue(max_size=16).launch()

requirements.txt

@@ -0,0 +1,8 @@
+torch>=2.4
+transformers>=4.56,<5
+safetensors>=0.4
+accelerate>=0.33
+gradio>=5.0
+spaces
+huggingface_hub>=0.24
+ntkmirror @ git+https://github.com/leochlon/ntkmirror.git