| """ |
| app.py — SimQuantum Tuning Lab |
| ================================= |
| AMD Developer Hackathon 2026. |
| |
| Before running, set env vars (on MI300X): |
| |
| export QDOT_LLM_BASE_URL=http://localhost:8000/v1 |
| export QDOT_LLM_MODEL=Qwen/Qwen3-8B |
| streamlit run app.py --server.port 8501 --server.address 0.0.0.0 |
| |
| Or locally (LLM offline, physics sim still runs): |
| streamlit run app.py |
| """ |
|
|
| from __future__ import annotations |
| import os, sys, threading, time, re |
| from pathlib import Path |
| import numpy as np |
|
|
| sys.path.insert(0, str(Path(__file__).parent)) |
| import streamlit as st |
|
|
| st.set_page_config( |
| page_title="SimQuantum Tuning Lab", |
| page_icon="⚛", |
| layout="wide", |
| initial_sidebar_state="expanded", |
| ) |
|
|
| |
| |
| |
| st.markdown(""" |
| <style> |
| @import url('https://fonts.googleapis.com/css2?family=Libre+Franklin:wght@300;400;600;700&family=JetBrains+Mono:wght@400;500&display=swap'); |
| |
| html,body,.stApp{background:#F2F0EB!important;font-family:'Libre Franklin',sans-serif;color:#1C2333;} |
| #MainMenu,footer,header{visibility:hidden;} |
| .block-container{padding:1.4rem 1.8rem 1rem!important;max-width:1440px;} |
| |
| .topbar{display:flex;align-items:flex-start;justify-content:space-between; |
| padding-bottom:14px;border-bottom:2px solid #1C2333;margin-bottom:14px;} |
| .topbar-title{font-size:21px;font-weight:700;letter-spacing:-.5px;color:#1C2333;} |
| .topbar-sub{font-size:11px;color:#8A9AB0;font-family:'JetBrains Mono',monospace;margin-top:3px;} |
| |
| .badge{font-family:'JetBrains Mono',monospace;font-size:10px;padding:3px 8px; |
| border-radius:3px;font-weight:500;display:inline-block;margin-left:4px;} |
| .badge-live{background:#E6F4F1;color:#00897B;border:1px solid #B2DFDB;} |
| .badge-idle{background:#FFF3E0;color:#E65100;border:1px solid #FFCC80;} |
| .badge-mi300x{background:#EDE7F6;color:#5E35B1;border:1px solid #D1C4E9;} |
| .badge-warn{background:#FFF3E0;color:#B85000;border:1px solid #FFD09A;} |
| |
| .timeline{display:flex;align-items:center;margin:12px 0 14px;} |
| .tn{display:flex;flex-direction:column;align-items:center; |
| font-family:'JetBrains Mono',monospace;font-size:9px;min-width:88px;} |
| .tn-c{width:26px;height:26px;border-radius:50%;display:flex;align-items:center; |
| justify-content:center;font-size:11px;font-weight:700;margin-bottom:4px;border:2px solid;} |
| .tn-l{color:#8A9AB0;text-align:center;line-height:1.3;} |
| .tn-done .tn-c{background:#E6F4F1;border-color:#00897B;color:#00897B;} |
| .tn-active .tn-c{background:#1C2333;border-color:#1C2333;color:#F2F0EB; |
| box-shadow:0 0 0 4px rgba(28,35,51,.10);} |
| .tn-pending .tn-c{background:#F2F0EB;border-color:#D0CBB8;color:#C0B8A8;} |
| .tn-phase3 .tn-c{background:#F8F7F4;border-color:#E0D8C8;color:#C8C0B0;font-size:8px;} |
| .tn-done .tn-l{color:#00897B;} .tn-active .tn-l{color:#1C2333;font-weight:600;} |
| .tn-phase3 .tn-l{color:#C0B8A8;font-style:italic;} |
| .tline{flex:1;height:2px;background:#D0CBB8;margin-top:-18px;} |
| .tline-done{background:#00897B;} |
| |
| .kpi-row{display:grid;grid-template-columns:repeat(4,1fr);gap:8px;margin-bottom:12px;} |
| .kpi{background:#FFF;border:1px solid #DDD9D0;border-radius:7px;padding:10px 12px;text-align:center;} |
| .kpi-v{font-family:'JetBrains Mono',monospace;font-size:21px;font-weight:500;color:#1C2333;} |
| .kpi-u{font-size:10px;color:#8A9AB0;} |
| .kpi-l{font-size:9px;letter-spacing:1px;text-transform:uppercase;color:#A8B0BC;margin-top:2px;} |
| |
| .card{background:#FFF;border-radius:8px;border:1px solid #DDD9D0;padding:14px 16px;margin-bottom:10px;} |
| .card-title{font-size:10px;font-weight:700;letter-spacing:1.5px;text-transform:uppercase; |
| color:#8A9AB0;margin-bottom:10px;font-family:'JetBrains Mono',monospace;} |
| |
| .chat-outer{background:#FFF;border:1px solid #DDD9D0;border-radius:8px; |
| display:flex;flex-direction:column;height:460px;} |
| .chat-head{padding:10px 14px;border-bottom:1px solid #EDE8DF;flex-shrink:0; |
| font-size:10px;font-weight:700;letter-spacing:1.5px;text-transform:uppercase; |
| color:#8A9AB0;font-family:'JetBrains Mono',monospace; |
| display:flex;align-items:center;justify-content:space-between;} |
| .chat-body{flex:1;overflow-y:auto;padding:12px 14px;display:flex;flex-direction:column;gap:9px;} |
| .msg{display:flex;flex-direction:column;max-width:92%;} |
| .msg-u{align-self:flex-end;} .msg-a{align-self:flex-start;} |
| .bubble{padding:8px 12px;border-radius:10px;font-size:12.5px;line-height:1.65;} |
| .msg-u .bubble{background:#1C2333;color:#F2F0EB;border-radius:10px 10px 2px 10px;} |
| .msg-a .bubble{background:#F2F0EB;color:#1C2333;border:1px solid #DDD9D0; |
| border-radius:10px 10px 10px 2px;font-family:'JetBrains Mono',monospace;font-size:11.5px;} |
| .msg-ev .bubble{background:#FFF8E8;border:1px solid #FFD070;color:#7A5000; |
| font-family:'JetBrains Mono',monospace;font-size:11px;} |
| .mlabel{font-size:9px;letter-spacing:.5px;color:#A8B0BC;margin-bottom:2px; |
| font-family:'JetBrains Mono',monospace;} |
| .msg-u .mlabel{text-align:right;} |
| |
| .hitl-card{background:#FFFAED;border:2px solid #E8A020;border-radius:8px; |
| padding:14px 16px;margin-bottom:10px;} |
| .hitl-title{font-size:12px;font-weight:700;color:#B85000;margin-bottom:6px; |
| font-family:'JetBrains Mono',monospace;} |
| .hitl-body{font-size:12px;color:#5A4000;line-height:1.5;} |
| |
| .spy-grid{display:grid;grid-template-columns:repeat(5,1fr);gap:6px;} |
| .spy{background:#F8F6F2;border:1px solid #DDD9D0;border-radius:6px; |
| padding:9px 5px;text-align:center;transition:all .3s;} |
| .spy-on{background:#FFF;border-color:#00897B;box-shadow:0 0 0 2px rgba(0,137,123,.12);} |
| .spy-em{font-size:19px;margin-bottom:3px;} |
| .spy-name{font-size:10px;font-weight:600;color:#1C2333;} |
| .spy-role{font-size:9px;color:#A8B0BC;} |
| .spy-dot{width:7px;height:7px;border-radius:50%;margin:5px auto 0;background:#D0CBB8;} |
| .spy-dot-on{background:#00897B;} |
| |
| div[data-testid="stProgressBar"]>div{background:#E6F4F1;} |
| div[data-testid="stProgressBar"]>div>div{background:#00897B!important;} |
| .stButton>button{font-family:'Libre Franklin',sans-serif!important; |
| font-size:13px!important;font-weight:600!important;} |
| section[data-testid="stSidebar"]{background:#1C2333!important;} |
| section[data-testid="stSidebar"] label{color:#8A9AB0!important;} |
| ::-webkit-scrollbar{width:4px;} |
| ::-webkit-scrollbar-thumb{background:#D0CBB8;border-radius:2px;} |
| </style> |
| """, unsafe_allow_html=True) |
|
|
|
|
| |
| |
| |
| STAGES = [ |
| ("BOOTSTRAPPING", "⚡","Integrity check", False), |
| ("COARSE_SURVEY", "◈", "Voltage survey", False), |
| ("HYPERSURFACE_SEARCH", "◎","Find boundary", False), |
| ("CHARGE_ID", "◇","Classify charge", False), |
| ("NAVIGATION", "→","Navigate to (1,1)", True), |
| ("VERIFICATION", "✓","Verify stability", True), |
| ] |
| SPY_AGENTS = [ |
| ("perception","🔬","Perception","Quality Inspector",{"BOOTSTRAPPING","COARSE_SURVEY","CHARGE_ID"}), |
| ("executive", "🏛","Executive", "Mission Conductor",set(s[0] for s in STAGES)), |
| ("planning", "📐","Planning", "Navigator", {"HYPERSURFACE_SEARCH","NAVIGATION"}), |
| ("safety", "🛡","Safety", "Hardware Marshal", {"NAVIGATION","VERIFICATION"}), |
| ("hitl", "🛑","HITL", "Human Governor", set()), |
| ] |
| START_KWS = {"start","begin","run","tune","go","launch","init","initialize","initialise","proceed"} |
|
|
| STAGE_DESC = { |
| "BOOTSTRAPPING": ("64-pt line scan. Verifies gate response and sensor signal.","~64 pts"), |
| "COARSE_SURVEY": ("32×32 sweep across full voltage bounds.", "~1024 pts"), |
| "HYPERSURFACE_SEARCH": ("16×16 local scan around survey peak. Confirms boundary.", "~256 pts"), |
| "CHARGE_ID": ("32×32 scan. 5-model CNN ensemble classifies charge state.", "~1024 pts"), |
| "NAVIGATION": ("Bayesian BO proposes voltage moves toward (1,1).", "variable"), |
| "VERIFICATION": ("3× repeated 16×16 scans confirming (1,1) stability.", "~768 pts"), |
| } |
|
|
| STABILITY_CS = [ |
| [0.00,"#07070A"],[0.30,"#1A0E40"],[0.55,"#7A1800"], |
| [0.75,"#D84000"],[0.90,"#FF9000"],[1.00,"#FFE040"], |
| ] |
| PLOT_LAYOUT = dict( |
| paper_bgcolor="rgba(0,0,0,0)", plot_bgcolor="#FAFAF8", |
| font=dict(color="#8A9AB0",size=10,family="JetBrains Mono"), |
| margin=dict(l=10,r=10,t=30,b=10), |
| ) |
|
|
| |
| DR_Q_SYSTEM = """\ |
| You are Dr. Q, the AI co-pilot for SimQuantum — an autonomous quantum dot tuning system |
| running on AMD MI300X hardware. You are Qwen3-8B. |
| |
| Your role: have a real conversation about this experiment with whoever is asking. |
| Adapt your depth to the person — a curious student gets clear analogies, an expert |
| physicist like Natalia Ares gets precise technical detail. Read the question, match it. |
| |
| THE EXPERIMENT: |
| A 6-stage POMDP agent autonomously navigates gate voltage space (Vg1, Vg2) to tune |
| a double quantum dot to the (1,1) charge state — one electron per dot — required |
| for spin qubit operation. |
| |
| Stages: BOOTSTRAPPING → COARSE_SURVEY → HYPERSURFACE_SEARCH → CHARGE_ID → NAVIGATION → VERIFICATION |
| |
| Physics simulator: Capacitive Interaction Model (CIM) — real semiconductor physics, |
| not a toy model. Parameters: charging energy E_c, lever arm, tunnel coupling t_c. |
| |
| CNN: 5-model TinyCNN ensemble trained on 51,000 simulated stability diagrams. |
| Val accuracy 91.4%. OOD detection via Mahalanobis distance on penultimate features. |
| |
| The agent uses Bayesian optimisation (MultiResBO) for navigation and a particle filter |
| belief state over charge configurations. |
| |
| IMPORTANT HONESTY: Navigation (Phase 3) is unsolved. The Bayesian optimiser has no |
| converging reward signal in intermediate voltage space — the CNN sees "misc" everywhere |
| except at charge boundaries. The agent stops at CHARGE_ID reliably; NAVIGATION wanders. |
| Be honest about this if asked. |
| |
| CURRENT RUN STATE: |
| {state_block} |
| |
| CONVERSATION STYLE: |
| - Talk like a physicist who is also a good teacher |
| - 2-4 sentences usually, more only if genuinely needed |
| - Reference actual numbers from the state when relevant |
| - Be honest about what works and what doesn't |
| - Never say "Great news!" or "Certainly!" or "As an AI" |
| - If something is unclear, ask for clarification |
| """ |
|
|
|
|
| |
| |
| |
| def _init(): |
| d = dict( |
| agent=None, exp_state=None, narrator=None, hitl_manager=None, |
| done_event=None, thread=None, running=False, run_count=0, |
| chat=[], |
| llm_url="", |
| llm_api_key="", |
| llm_model="accounts/fireworks/models/qwen3-8b", |
| use_cnn=True, |
| meas_budget=8096, |
| max_steps=140, |
| ) |
| for k,v in d.items(): |
| if k not in st.session_state: |
| st.session_state[k] = v |
| _init() |
|
|
| |
| if not st.session_state.llm_url: |
| env_url = os.environ.get("QDOT_LLM_BASE_URL", "") |
| if env_url: |
| st.session_state.llm_url = env_url |
|
|
| if not st.session_state.llm_api_key: |
| env_key = os.environ.get("QDOT_LLM_API_KEY", "") |
| if env_key: |
| st.session_state.llm_api_key = env_key |
|
|
| if st.session_state.llm_model == "accounts/fireworks/models/qwen3-8b": |
| env_model = os.environ.get("QDOT_LLM_MODEL", "") |
| if env_model: |
| st.session_state.llm_model = env_model |
|
|
|
|
| |
| |
| |
| def _llm_available() -> bool: |
| return bool(st.session_state.llm_url.strip()) |
|
|
|
|
| def _build_state_block() -> str: |
| """Snapshot of current agent state for the system prompt.""" |
| exp = st.session_state.exp_state |
| agnt = st.session_state.agent |
| if exp is None: |
| return "No run active. Agent is idle." |
| budget = agnt.measurement_budget if agnt else st.session_state.meas_budget |
| meas = exp.total_measurements |
| stage = exp.stage.name |
| vg1 = exp.current_voltage.vg1 |
| vg2 = exp.current_voltage.vg2 |
| snr = f"{exp.last_dqc.snr_db:.1f}dB" if exp.last_dqc else "unknown" |
| dqc = exp.last_dqc.quality.value if exp.last_dqc else "unknown" |
| bt = exp.total_backtracks |
| done = st.session_state.done_event and st.session_state.done_event.is_set() |
|
|
| cls_str = "none yet" |
| if exp.last_classification: |
| c = exp.last_classification |
| ood = "OOD" if exp.is_ood else "in-dist" |
| cls_str = f"{c.label.value} ({c.confidence:.0%} conf, {ood})" |
|
|
| p11 = exp.belief.charge_probs.get((1,1), 0.0) |
| ml = exp.belief.most_likely_state() |
|
|
| return ( |
| f"Stage: {stage}\n" |
| f"Measurements used: {meas}/{budget} ({100*meas//max(budget,1)}%)\n" |
| f"Current voltage: Vg1={vg1:+.3f}V, Vg2={vg2:+.3f}V\n" |
| f"Signal quality: SNR={snr}, DQC={dqc}\n" |
| f"CNN classification: {cls_str}\n" |
| f"Belief P(1,1)={p11:.3f}, most likely: {ml}\n" |
| f"Backtracks: {bt}\n" |
| f"Run finished: {'yes' if done else 'no'}" |
| ) |
|
|
|
|
| def _call_qwen(user_msg: str, image_bytes: bytes | None = None) -> str: |
| """ |
| Call Dr. Q. System prompt + conversation history + user message → response. |
| Supports Fireworks AI (https://api.fireworks.ai/inference/v1) and local vLLM. |
| If image_bytes provided, sends the image alongside the message (VL models). |
| No fallbacks, no scripts. If the LLM is down, say so clearly. |
| """ |
| url = st.session_state.llm_url.strip().rstrip("/") |
| model = st.session_state.llm_model.strip() or "accounts/fireworks/models/qwen3-8b" |
| api_key = st.session_state.llm_api_key.strip() or os.environ.get("QDOT_LLM_API_KEY", "EMPTY") |
|
|
| |
| system = DR_Q_SYSTEM.format(state_block=_build_state_block()) |
| messages = [{"role": "system", "content": system}] |
| for msg in st.session_state.chat[-16:]: |
| role = "user" if msg["role"] == "user" else "assistant" |
| messages.append({"role": role, "content": msg["content"]}) |
|
|
| |
| if image_bytes and messages and messages[-1]["role"] == "user": |
| import base64 |
| b64 = base64.b64encode(image_bytes).decode() |
| messages[-1]["content"] = [ |
| {"type": "image_url", "image_url": {"url": f"data:image/png;base64,{b64}"}}, |
| {"type": "text", "text": messages[-1]["content"]}, |
| ] |
|
|
| try: |
| import openai |
| |
| if "fireworks.ai" in url: |
| api_base = url |
| else: |
| api_base = url if url.endswith("/v1") else url + "/v1" |
|
|
| client = openai.OpenAI(base_url=api_base, api_key=api_key) |
| resp = client.chat.completions.create( |
| model=model, |
| messages=messages, |
| max_tokens=1024, |
| temperature=0.6, |
| ) |
| raw = resp.choices[0].message.content or "" |
| |
| import re as _re |
| think_match = _re.search(r"<think>(.*?)</think>", raw, flags=_re.DOTALL) |
| think_block = think_match.group(1).strip() if think_match else "" |
| clean = _re.sub(r"<think>.*?</think>", "", raw, flags=_re.DOTALL).strip() |
| |
| st.session_state["_last_think"] = think_block |
| return clean |
|
|
| except ImportError: |
| return ("openai package not installed. Run: pip install openai\n" |
| "Then restart the app.") |
| except Exception as exc: |
| return (f"Cannot reach Dr. Q at {url}.\n\nError: {exc}\n\n" |
| f"Check your API key and endpoint URL in the sidebar.") |
|
|
|
|
| def _add_msg(role: str, content: str, kind: str = "n", think: str = ""): |
| st.session_state.chat.append({"role": role, "content": content, "kind": kind, "think": think}) |
|
|
|
|
| def _handle_chat(user_msg: str): |
| """Add user message, start agent if requested, get Dr. Q response.""" |
| _add_msg("user", user_msg) |
|
|
| is_start = any(kw in user_msg.lower() for kw in START_KWS) |
|
|
| if is_start and not st.session_state.running and st.session_state.agent is None: |
| |
| try: |
| agent, exp_state, narrator, hitl_mgr = _make_agent( |
| st.session_state.use_cnn, |
| st.session_state.meas_budget, |
| st.session_state.max_steps, |
| ) |
| done_event = threading.Event() |
| thread = threading.Thread( |
| target=_run_thread, args=(agent, done_event), daemon=True) |
| st.session_state.update( |
| agent=agent, exp_state=exp_state, narrator=narrator, |
| hitl_manager=hitl_mgr, done_event=done_event, thread=thread, |
| running=True, run_count=st.session_state.run_count+1, |
| ) |
| thread.start() |
| |
| if _llm_available(): |
| reply = _call_qwen(user_msg) |
| else: |
| reply = ( |
| "Tuning sequence started.\n\n" |
| "To enable Qwen responses, set the vLLM endpoint in the sidebar " |
| "to your MI300X address (e.g. http://localhost:8000/v1).\n\n" |
| "The physics simulation is running. Ask me anything." |
| ) |
| except Exception as exc: |
| reply = f"Failed to start agent: {exc}" |
| else: |
| |
| if _llm_available(): |
| reply = _call_qwen(user_msg) |
| else: |
| reply = ( |
| "Qwen is not connected. Set the vLLM endpoint in the sidebar.\n\n" |
| "If you're on the MI300X, make sure vLLM is running:\n" |
| " docker exec -it rocm /bin/bash\n" |
| " vllm serve Qwen/Qwen3-8B --host 0.0.0.0 --port 8000\n\n" |
| "Then set the URL to http://localhost:8000 in the sidebar." |
| ) |
|
|
| _add_msg("assistant", reply, think=st.session_state.pop("_last_think", "")) |
|
|
|
|
| |
| |
| |
| def _make_agent(use_cnn, meas_budget, max_steps): |
| from qdot.core.state import ExperimentState |
| from qdot.core.types import ChargeLabel |
| from qdot.core.governance import GovernanceLogger |
| from qdot.core.hitl import HITLManager |
| from qdot.hardware.safety import SafetyCritic |
| from qdot.perception.dqc import DQCGatekeeper |
| from qdot.simulator.cim import CIMSimulatorAdapter |
| from qdot.agent.executive import ExecutiveAgent |
|
|
| rng = np.random.default_rng() |
| E_c = float(rng.uniform(2.1, 2.9)) |
| adapter = CIMSimulatorAdapter(device_id="demo_qdot", params={ |
| "E_c1": E_c, "E_c2": E_c*float(rng.uniform(0.93,1.07)), |
| "t_c": 0.05, "T": 0.015, |
| "lever_arm": float(rng.uniform(0.68,0.82)), "noise_level": 0.02, |
| }) |
| state = ExperimentState.new(device_id="demo_qdot", target_label=ChargeLabel.DOUBLE_DOT) |
|
|
| inspection = None |
| if use_cnn: |
| try: |
| from qdot.perception.inspector import InspectionAgent |
| inspection = InspectionAgent() |
| except Exception as e: |
| st.sidebar.warning(f"CNN unavailable: {e}") |
|
|
| run_dir = Path("results/demo") / state.run_id |
| run_dir.mkdir(parents=True, exist_ok=True) |
|
|
| hitl_mgr = HITLManager(queue_dir=str(run_dir/"hitl")) |
| hitl_mgr.set_test_mode() |
|
|
| agent = ExecutiveAgent( |
| state=state, adapter=adapter, inspection_agent=inspection, |
| dqc=DQCGatekeeper(), |
| safety_critic=SafetyCritic(voltage_bounds=state.voltage_bounds, l1_max=0.10), |
| hitl_manager=hitl_mgr, |
| governance_logger=GovernanceLogger( |
| run_id=state.run_id, log_dir=str(run_dir/"governance")), |
| max_steps=max_steps, measurement_budget=meas_budget, |
| ) |
| return agent, state, agent.narrator, hitl_mgr |
|
|
|
|
| def _run_thread(agent, done_event): |
| try: agent.run() |
| except Exception: pass |
| finally: done_event.set() |
|
|
|
|
| |
| |
| |
| import plotly.graph_objects as go |
|
|
| def _best_scan(state): |
| if state.last_classification: |
| m = state.measurements.get(state.last_classification.measurement_id) |
| if m and m.is_2d and m.array is not None: return m |
| best, best_area = None, 0.0 |
| for m in state.measurements.values(): |
| if not m.is_2d or m.array is None: continue |
| if m.v1_range and m.v2_range: |
| area = (m.v1_range[1]-m.v1_range[0])*(m.v2_range[1]-m.v2_range[0]) |
| if area > best_area: best_area,best = area,m |
| return best |
|
|
| def _fig_stability(state): |
| m = _best_scan(state) |
| fig = go.Figure() |
| if m is None: |
| fig.add_annotation(text="Awaiting first 2D scan…",x=0.5,y=0.5,showarrow=False, |
| font=dict(color="#8A9AB0",size=12,family="JetBrains Mono")) |
| else: |
| arr = np.asarray(m.array, dtype=np.float32) |
| p2,p98 = np.percentile(arr,[2,98]) |
| arrn = np.clip((arr-p2)/max(p98-p2,1e-8),0,1) |
| v1lo,v1hi = m.v1_range or (-8,8) |
| v2lo,v2hi = m.v2_range or (-8,8) |
| fig.add_trace(go.Heatmap( |
| z=arrn, x=np.linspace(v1lo,v1hi,arrn.shape[1]), |
| y=np.linspace(v2lo,v2hi,arrn.shape[0]), |
| colorscale=STABILITY_CS, showscale=True, |
| colorbar=dict(thickness=8,tickvals=[0,.5,1], |
| ticktext=["Blockade","—","Peak"], |
| tickfont=dict(size=8,family="JetBrains Mono"), |
| title=dict(text="G",font=dict(size=9))), |
| )) |
| vg1,vg2 = state.current_voltage.vg1,state.current_voltage.vg2 |
| fig.add_shape(type="line",x0=vg1,x1=vg1,y0=v2lo,y1=v2hi, |
| line=dict(color="rgba(255,255,255,0.55)",width=1,dash="dot")) |
| fig.add_shape(type="line",x0=v1lo,x1=v1hi,y0=vg2,y1=vg2, |
| line=dict(color="rgba(255,255,255,0.55)",width=1,dash="dot")) |
| fig.add_trace(go.Scatter(x=[vg1],y=[vg2],mode="markers", |
| marker=dict(size=8,color="#FF4040",symbol="cross-thin", |
| line=dict(width=2.5,color="#FF4040")),showlegend=False)) |
| fig.update_layout( |
| title=dict(text="Charge Stability Diagram",font=dict(size=11,color="#5A6478")), |
| xaxis=dict(title="Vg₁ (V)",gridcolor="#E8E4DC",zeroline=False), |
| yaxis=dict(title="Vg₂ (V)",gridcolor="#E8E4DC",zeroline=False), |
| height=265, **PLOT_LAYOUT) |
| return fig |
|
|
| def _fig_belief(state): |
| probs = state.belief.charge_probs |
| z = np.zeros((3,3)) |
| for (n1,n2),p in probs.items(): |
| if 0<=n1<=2 and 0<=n2<=2: z[n2][n1]=float(p) |
| fig = go.Figure(data=go.Heatmap( |
| z=z, x=["N₁=0","N₁=1","N₁=2"], y=["N₂=0","N₂=1","N₂=2"], |
| colorscale=[[0,"#F2F0EB"],[0.5,"#B2DFDB"],[1,"#00897B"]], |
| showscale=False, zmin=0, zmax=1, |
| text=[[f"{z[j][i]:.2f}" for i in range(3)] for j in range(3)], |
| texttemplate="%{text}", |
| textfont=dict(size=12,color="#1C2333",family="JetBrains Mono"), |
| )) |
| fig.add_shape(type="rect",x0=.5,x1=1.5,y0=.5,y1=1.5, |
| line=dict(color="#00897B",width=2.5)) |
| fig.add_annotation(x=1,y=1.5,text="TARGET",showarrow=False, |
| yshift=14,font=dict(color="#00897B",size=9,family="JetBrains Mono")) |
| fig.update_layout( |
| title=dict(text="Belief P(N₁,N₂|obs)",font=dict(size=11,color="#5A6478")), |
| height=210, **PLOT_LAYOUT) |
| return fig |
|
|
| def _fig_traj(state): |
| if len(state.trajectory)<2: return None |
| xs=[v.vg1 for v in state.trajectory]; ys=[v.vg2 for v in state.trajectory] |
| n=len(xs) |
| fig=go.Figure() |
| fig.add_trace(go.Scatter(x=xs,y=ys,mode="lines+markers", |
| line=dict(color="#B2DFDB",width=1.5), |
| marker=dict(size=5,color=list(range(n)), |
| colorscale=[[0,"#E0F4F1"],[1,"#00897B"]],showscale=False), |
| showlegend=False)) |
| fig.add_trace(go.Scatter(x=[xs[-1]],y=[ys[-1]],mode="markers", |
| marker=dict(size=9,color="#E85000",symbol="x-thin", |
| line=dict(width=2.5,color="#E85000")),showlegend=False)) |
| fig.update_layout( |
| title=dict(text="Voltage Trajectory",font=dict(size=11,color="#5A6478")), |
| xaxis=dict(title="Vg₁",gridcolor="#E8E4DC"), |
| yaxis=dict(title="Vg₂",gridcolor="#E8E4DC"), |
| height=185, **PLOT_LAYOUT) |
| return fig |
|
|
|
|
| |
| |
| |
| def _timeline(current, done_event): |
| is_done = done_event and done_event.is_set() |
| try: ci = [s[0] for s in STAGES].index(current) |
| except ValueError: ci = -1 |
| html = '<div class="timeline">' |
| for i,(sname,icon,desc,p3) in enumerate(STAGES): |
| if p3: css="tn tn-phase3" |
| elif i<ci or (is_done and i<=ci): css="tn tn-done" |
| elif i==ci: css="tn tn-active" |
| else: css="tn tn-pending" |
| chk = "✓" if "done" in css and not p3 else icon |
| html += (f'<div class="{css}"><div class="tn-c">{chk}</div>' |
| f'<div class="tn-l">{desc}</div></div>') |
| if i<len(STAGES)-1: |
| lc = "tline tline-done" if i<ci else "tline" |
| html += f'<div class="{lc}"></div>' |
| html += '</div>' |
| st.markdown(html, unsafe_allow_html=True) |
|
|
| def _kpi(state, agent): |
| b=state.total_measurements; t=agent.measurement_budget |
| vg1=state.current_voltage.vg1; vg2=state.current_voltage.vg2 |
| snr=state.last_dqc.snr_db if state.last_dqc else 0.0 |
| st.markdown(f""" |
| <div class="kpi-row"> |
| <div class="kpi"><div class="kpi-v">{b}<span class="kpi-u">/{t}</span></div> |
| <div class="kpi-l">Measurements</div></div> |
| <div class="kpi"><div class="kpi-v">{vg1:+.2f}<span class="kpi-u"> V</span></div> |
| <div class="kpi-l">Vg₁</div></div> |
| <div class="kpi"><div class="kpi-v">{vg2:+.2f}<span class="kpi-u"> V</span></div> |
| <div class="kpi-l">Vg₂</div></div> |
| <div class="kpi"><div class="kpi-v">{snr:.1f}<span class="kpi-u"> dB</span></div> |
| <div class="kpi-l">SNR</div></div> |
| </div>""", unsafe_allow_html=True) |
|
|
| def _spy(current, hitl_active): |
| html = '<div class="spy-grid">' |
| for key,em,name,role,active in SPY_AGENTS: |
| on = (key=="hitl" and hitl_active) or (key!="hitl" and current in active) |
| html += (f'<div class="{"spy spy-on" if on else "spy"}">' |
| f'<div class="spy-em">{em}</div>' |
| f'<div class="spy-name">{name}</div><div class="spy-role">{role}</div>' |
| f'<div class="{"spy-dot spy-dot-on" if on else "spy-dot"}"></div></div>') |
| html += '</div>' |
| st.markdown(html, unsafe_allow_html=True) |
|
|
| def _render_chat(): |
| """Read-only HTML chat panel. The actual input widget is at page bottom.""" |
| |
| narrator = st.session_state.narrator |
| if narrator: |
| for ev in narrator.event_log(): |
| tag = f"_ev_{ev.timestamp:.4f}" |
| if tag not in st.session_state: |
| st.session_state[tag] = True |
| if ev.kind == "exception" and ev.response: |
| |
| if _llm_available(): |
| contextualised = _call_qwen( |
| f"Agent anomaly detected: {ev.description}. " |
| f"Briefly explain what this means for the experiment.") |
| _add_msg("assistant", contextualised, kind="ev") |
| else: |
| _add_msg("assistant", f"⚠ {ev.description}", kind="ev") |
|
|
| llm_on = _llm_available() |
| badge = ( |
| '<span class="badge badge-mi300x">⬡ Qwen3-8B · AMD MI300X</span>' |
| if llm_on else |
| '<span class="badge badge-warn">LLM offline — set URL in sidebar</span>' |
| ) |
|
|
| msgs_html = "" |
| for msg in st.session_state.chat: |
| c = (msg["content"] |
| .replace("&","&").replace("<","<").replace(">",">") |
| .replace("\n","<br>")) |
| c = re.sub(r'\*\*(.+?)\*\*', r'<b>\1</b>', c) |
| if msg["role"] == "user": |
| msgs_html += (f'<div class="msg msg-u"><div class="mlabel">You</div>' |
| f'<div class="bubble">{c}</div></div>') |
| else: |
| css = "msg msg-a msg-ev" if msg.get("kind")=="ev" else "msg msg-a" |
| lbl = "Dr. Q — anomaly" if msg.get("kind")=="ev" else "Dr. Q" |
| think = msg.get("think", "") |
| think_html = "" |
| if think: |
| t = (think.replace("&","&").replace("<","<").replace(">",">") |
| .replace("\n","<br>")) |
| think_html = ( |
| f'<details style="margin-bottom:6px;cursor:pointer">' |
| f'<summary style="font-size:10px;color:#8A9AB0;font-family:JetBrains Mono,monospace;' |
| f'letter-spacing:0.5px;list-style:none;display:flex;align-items:center;gap:6px">' |
| f'<span style="color:#00897B">▶</span> Dr. Q\'s reasoning</summary>' |
| f'<div style="margin-top:6px;padding:10px 12px;background:#0D1117;border-radius:6px;' |
| f'font-size:11px;color:#5A6478;font-family:JetBrains Mono,monospace;line-height:1.6">' |
| f'{t}</div></details>' |
| ) |
| msgs_html += (f'<div class="{css}"><div class="mlabel">{lbl}</div>' |
| f'<div class="bubble">{think_html}{c}</div></div>') |
|
|
| st.markdown( |
| f'<div class="chat-outer">' |
| f' <div class="chat-head"><span>DR. Q — AI CO-PILOT</span>{badge}</div>' |
| f' <div class="chat-body" id="cq-body">{msgs_html}</div>' |
| f'</div>' |
| f'<script>(function(){{var e=document.getElementById("cq-body");' |
| f'if(e)e.scrollTop=e.scrollHeight;}})();</script>', |
| unsafe_allow_html=True, |
| ) |
|
|
|
|
| |
| |
| |
| running = st.session_state.running |
| exp_state = st.session_state.exp_state |
| done_event = st.session_state.done_event |
|
|
| with st.sidebar: |
| st.markdown( |
| '<div style="color:#F2F0EB;font-size:16px;font-weight:700;' |
| 'padding:8px 0 16px;border-bottom:1px solid #2C3545;margin-bottom:16px">' |
| '⚛ SimQuantum</div>', |
| unsafe_allow_html=True, |
| ) |
|
|
| |
| st.markdown( |
| '<div style="color:#8A9AB0;font-size:10px;font-family:JetBrains Mono,monospace;' |
| 'letter-spacing:1px;margin-bottom:6px">DR. Q — ENDPOINT</div>', |
| unsafe_allow_html=True, |
| ) |
| new_url = st.text_input( |
| "Endpoint URL", |
| value=st.session_state.llm_url, |
| placeholder="https://api.fireworks.ai/inference/v1", |
| label_visibility="collapsed", |
| ) |
| if new_url != st.session_state.llm_url: |
| st.session_state.llm_url = new_url |
|
|
| new_key = st.text_input( |
| "API Key", |
| value=st.session_state.llm_api_key, |
| placeholder="fw_xxxxxxxxxxxxxxxx", |
| type="password", |
| label_visibility="collapsed", |
| ) |
| if new_key != st.session_state.llm_api_key: |
| st.session_state.llm_api_key = new_key |
|
|
| if st.session_state.llm_url: |
| st.markdown( |
| f'<div style="font-size:10px;color:#00897B;font-family:JetBrains Mono,' |
| f'monospace;margin-bottom:12px">● connected to {st.session_state.llm_url}</div>', |
| unsafe_allow_html=True, |
| ) |
| |
| if st.button("Test connection", use_container_width=True): |
| try: |
| import openai |
| url = st.session_state.llm_url.strip().rstrip("/") |
| api_key = st.session_state.llm_api_key.strip() or os.environ.get("QDOT_LLM_API_KEY", "EMPTY") |
| c = openai.OpenAI(base_url=url, api_key=api_key) |
| models = c.models.list() |
| names = [m.id for m in models.data][:3] |
| st.success(f"Connected ✓") |
| except Exception as e: |
| st.error(f"Failed: {e}") |
| else: |
| st.markdown( |
| '<div style="font-size:10px;color:#E65100;font-family:JetBrains Mono,' |
| 'monospace;margin-bottom:12px">● offline — paste Fireworks URL + key above</div>', |
| unsafe_allow_html=True, |
| ) |
|
|
| st.markdown('<div style="height:8px"></div>', unsafe_allow_html=True) |
| new_model = st.text_input( |
| "Model name", |
| value=st.session_state.llm_model, |
| label_visibility="visible", |
| disabled=running, |
| ) |
| if new_model != st.session_state.llm_model: |
| st.session_state.llm_model = new_model |
|
|
| st.divider() |
|
|
| st.markdown( |
| '<div style="color:#8A9AB0;font-size:10px;font-family:JetBrains Mono,monospace;' |
| 'letter-spacing:1px;margin-bottom:8px">RUN CONFIGURATION</div>', |
| unsafe_allow_html=True, |
| ) |
| st.session_state.use_cnn = st.toggle( |
| "CNN Charge Classifier", value=st.session_state.use_cnn, disabled=running) |
| st.session_state.meas_budget = st.slider( |
| "Measurement Budget", 512, 8192, st.session_state.meas_budget, 256, |
| disabled=running) |
| st.session_state.max_steps = st.slider( |
| "Max Steps", 10, 300, st.session_state.max_steps, 10, disabled=running) |
|
|
| st.divider() |
|
|
| if st.button("Reset session", use_container_width=True, disabled=running): |
| for k in ["agent","exp_state","narrator","hitl_manager","done_event","thread","chat"]: |
| st.session_state[k] = [] if k=="chat" else None |
| st.session_state.running = False |
| st.rerun() |
|
|
| st.markdown( |
| '<div style="font-size:9px;color:#3A4A5A;font-family:JetBrains Mono,monospace;' |
| 'margin-top:16px;line-height:1.6">' |
| 'Physics: CIM simulator (CPU)<br>' |
| 'CNN: 5-model ensemble, 91.4% val acc<br>' |
| 'LLM: Qwen3-8B on AMD MI300X<br>' |
| 'BOOTSTRAPPING→CHARGE_ID: ✓ working<br>' |
| 'NAVIGATION: Phase 3, in development' |
| '</div>', |
| unsafe_allow_html=True, |
| ) |
|
|
|
|
| |
| |
| |
| if done_event and done_event.is_set() and running: |
| st.session_state.running = False |
| running = False |
| if exp_state and exp_state.stage.name == "COMPLETE": |
| st.balloons() |
| |
| already_summarised = any(m.get("kind")=="summary" for m in st.session_state.chat) |
| if not already_summarised and exp_state: |
| stage = exp_state.stage.name |
| meas = exp_state.total_measurements |
| budget = st.session_state.agent.measurement_budget if st.session_state.agent else 8096 |
| bt = exp_state.total_backtracks |
| _add_msg("user", |
| f"The run just finished — it stopped at {stage}. " |
| f"{meas}/{budget} measurements used, {bt} backtracks. " |
| f"What happened and what should I know for the next run?") |
| if _llm_available(): |
| reply = _call_qwen( |
| f"Post-run: stopped at {stage}, {meas}/{budget} measurements, {bt} backtracks.") |
| else: |
| reply = (f"Run stopped at {stage}. {meas}/{budget} measurements used, " |
| f"{bt} backtracks. Connect Qwen for a detailed analysis.") |
| _add_msg("assistant", reply, kind="summary") |
|
|
|
|
| |
| |
| |
| tl,tr = st.columns([3,1]) |
| with tl: |
| st.markdown( |
| '<div class="topbar"><div>' |
| '<div class="topbar-title">⚛ SimQuantum Tuning Lab</div>' |
| '<div class="topbar-sub">Autonomous quantum dot tuning · AMD Developer Hackathon 2026 · Kudzai Musarandega</div>' |
| '</div></div>', unsafe_allow_html=True) |
| with tr: |
| b = ('<span class="badge badge-live">● LIVE</span>' if running |
| else '<span class="badge badge-idle">● IDLE</span>') |
| if _llm_available(): |
| b += ' <span class="badge badge-mi300x">⬡ MI300X</span>' |
| st.markdown(f'<div style="text-align:right;padding-top:6px">{b}</div>', |
| unsafe_allow_html=True) |
|
|
|
|
| |
| |
| |
| if exp_state is None: |
| if not st.session_state.chat: |
| if _llm_available(): |
| |
| _add_msg("user", "Hello, introduce yourself briefly.") |
| intro = _call_qwen("Hello, introduce yourself briefly.") |
| _add_msg("assistant", intro) |
| else: |
| _add_msg("assistant", |
| "Ready. Set the vLLM endpoint in the sidebar to enable Qwen.\n\n" |
| "Type **start** to begin a tuning run, or ask me about the experiment.") |
|
|
| sl,sr = st.columns([3,2], gap="large") |
| with sl: |
| st.markdown(""" |
| <div class="card" style="padding:22px 24px"> |
| <div class="card-title">What this system does</div> |
| <p style="font-size:13px;color:#5A6478;line-height:1.75;margin:0"> |
| SimQuantum autonomously tunes a double quantum dot device to the |
| <strong>(1,1) charge state</strong> — one electron per dot — required |
| for spin qubit operation. 6-stage POMDP planner, 5-model CNN ensemble |
| (91.4% val acc), Bayesian optimisation. Qwen3-8B on AMD MI300X |
| acts as Dr. Q — ask it anything, in any register. |
| </p> |
| </div> |
| <div class="card"> |
| <div class="card-title">Reading a stability diagram</div> |
| <p style="font-size:13px;color:#5A6478;line-height:1.75;margin:0"> |
| Conductance G vs gate voltages (Vg₁, Vg₂). |
| Bright lines = Coulomb peaks (charge transitions). |
| Dark regions = Coulomb blockade (fixed electron number). |
| Intersections form a honeycomb: (0,0), (1,0), (0,1), <strong>(1,1)</strong>… |
| The agent navigates to the (1,1) diamond. |
| </p> |
| </div> |
| """, unsafe_allow_html=True) |
| img_path = Path(__file__).parent/"assets"/"simquantum.png" |
| if img_path.exists(): |
| st.image(str(img_path), use_container_width=True) |
| with sr: |
| _render_chat() |
|
|
| |
| |
| |
| else: |
| agent = st.session_state.agent |
| hitl_mgr = st.session_state.hitl_manager |
| current_stg = exp_state.stage.name |
|
|
| _timeline(current_stg, done_event) |
| _kpi(exp_state, agent) |
| pct = min(100, int(100*exp_state.total_measurements/agent.measurement_budget)) |
| st.progress(pct/100, text=f"Measurement budget {pct}%") |
|
|
| pending = hitl_mgr.get_pending() if hitl_mgr else [] |
| if pending: |
| req = pending[0] |
| st.markdown( |
| f'<div class="hitl-card">' |
| f'<div class="hitl-title">⚠ HITL GATE — Human approval required</div>' |
| f'<div class="hitl-body">Step {req["step"]} · Stage {req["stage"]} · ' |
| f'Risk {req["risk_score"]:.2f}<br><strong>{req["trigger_reason"]}</strong></div>' |
| f'</div>', unsafe_allow_html=True) |
| hc1,hc2,_ = st.columns([1,1,5]) |
| with hc1: |
| if st.button("✓ Approve",type="primary",key=f"appr_{req['id']}"): |
| hitl_mgr.approve(req["id"],deciding_human="operator") |
| _add_msg("user","I approved the HITL gate.") |
| if _llm_available(): |
| _add_msg("assistant", _call_qwen("The operator just approved the HITL gate. Acknowledge briefly.")) |
| else: |
| _add_msg("assistant","Approved. Agent continues.") |
| st.rerun() |
| with hc2: |
| if st.button("✗ Reject",key=f"rej_{req['id']}"): |
| hitl_mgr.reject(req["id"],deciding_human="operator") |
| _add_msg("user","I rejected the HITL gate.") |
| if _llm_available(): |
| _add_msg("assistant", _call_qwen("The operator rejected the HITL gate. Acknowledge and explain what happens next.")) |
| else: |
| _add_msg("assistant","Rejected. Agent backtracks.") |
| st.rerun() |
|
|
| st.markdown("<div style='height:6px'></div>",unsafe_allow_html=True) |
| left,right = st.columns([3,2],gap="large") |
|
|
| with left: |
| st.markdown('<div class="card-title">Charge Stability Diagram</div>', |
| unsafe_allow_html=True) |
| st.plotly_chart(_fig_stability(exp_state),use_container_width=True, |
| config={"displayModeBar":False},key=f"s_{time.monotonic_ns()}") |
|
|
| if exp_state.last_classification: |
| cls = exp_state.last_classification |
| ood_col = "#C84B00" if exp_state.is_ood else "#00897B" |
| ood_txt = "OOD warning" if exp_state.is_ood else "in-distribution" |
| st.markdown( |
| f'<div class="card"><span class="card-title">CNN Classification</span>' |
| f'<span style="float:right;font-size:10px;color:{ood_col};' |
| f'font-family:JetBrains Mono,monospace">{ood_txt}</span><br>' |
| f'<span style="font-family:JetBrains Mono,monospace;font-size:22px;' |
| f'font-weight:500;color:#00897B">{cls.label.value.upper()}</span>' |
| f'<span style="color:#8A9AB0;font-size:12px;margin-left:10px">' |
| f'conf {cls.confidence:.1%}</span></div>', |
| unsafe_allow_html=True) |
|
|
| bc1,bc2 = st.columns(2) |
| with bc1: |
| st.plotly_chart(_fig_belief(exp_state),use_container_width=True, |
| config={"displayModeBar":False},key=f"b_{time.monotonic_ns()}") |
| with bc2: |
| ft = _fig_traj(exp_state) |
| if ft: |
| st.plotly_chart(ft,use_container_width=True, |
| config={"displayModeBar":False},key=f"t_{time.monotonic_ns()}") |
|
|
| st.markdown('<div class="card-title" style="margin-top:4px">Agent Activity</div>', |
| unsafe_allow_html=True) |
| _spy(current_stg, bool(pending)) |
|
|
| img_path = Path(__file__).parent/"assets"/"simquantum.png" |
| if img_path.exists(): |
| st.image(str(img_path),use_container_width=True) |
|
|
| with right: |
| _render_chat() |
|
|
| if current_stg in STAGE_DESC: |
| desc,cost = STAGE_DESC[current_stg] |
| st.markdown( |
| f'<div class="card" style="margin-top:8px">' |
| f'<div class="card-title">Current stage · {current_stg}</div>' |
| f'<div style="font-size:12px;color:#5A6478;line-height:1.6">{desc}</div>' |
| f'<div style="font-size:10px;color:#A8B0BC;margin-top:5px;' |
| f'font-family:JetBrains Mono,monospace">Budget: {cost}</div>' |
| f'</div>', unsafe_allow_html=True) |
|
|
| if done_event and done_event.is_set(): |
| ok = current_stg=="COMPLETE" |
| col = "#00897B" if ok else "#C84B00" |
| txt = "MISSION COMPLETE" if ok else f"STOPPED — {current_stg}" |
| red = 1.0-(exp_state.total_measurements/max(64*64,1)) |
| st.markdown( |
| f'<div class="card" style="border-color:{col};margin-top:8px">' |
| f'<div style="font-size:14px;font-weight:700;color:{col};' |
| f'font-family:JetBrains Mono,monospace;margin-bottom:8px">{txt}</div>' |
| f'<div style="font-size:12px;color:#5A6478;display:grid;' |
| f'grid-template-columns:1fr 1fr;gap:5px">' |
| f'<span>Measurements: <b>{exp_state.total_measurements}</b></span>' |
| f'<span>Steps: <b>{agent.control_steps}</b></span>' |
| f'<span>Backtracks: <b>{exp_state.total_backtracks}</b></span>' |
| f'<span>Reduction: <b>{red:.0%}</b></span>' |
| f'</div></div>', unsafe_allow_html=True) |
| if st.button("🔄 New Run",use_container_width=True): |
| for k in ["agent","exp_state","narrator","hitl_manager","done_event","thread"]: |
| st.session_state[k] = None |
| st.session_state.running = False |
| st.rerun() |
|
|
|
|
| |
| |
| |
| |
| if prompt := st.chat_input("Ask Dr. Q anything, or type 'start' to begin…"): |
| _handle_chat(prompt) |
| st.rerun() |
|
|
|
|
| |
| |
| |
| if running and done_event and not done_event.is_set(): |
| time.sleep(0.8) |
| st.rerun() |
