⚡ DJINN 1.005B Dual-GEMM Concurrent Decision Engine

Blackwell RTX PRO 6000 Server Edition (102 GB GDDR7 VRAM, 188 SMs)

DJINN is a high-speed, non-autoregressive System One Decision Engine with 1,010,085,888 pure dense parameters active on every token ($P=1$, strictly zero Mixture of Experts, zero sequence pooling, zero token dropping).

DJINN is trained via Reinforcement Learning from Compiler Decisions (RLCD) to output instant, calibrated posterior probability distributions across:

• Choice Decisions: Multi-alternative classification with calibrated probabilities across up to 256 options.
• Score Decisions: Expected value estimation on continuous or ordinal scales (10-bin calibrated distribution).
• Noul Decisions: Calibrated binary hypothesis testing with epistemic bounds ($[P(\text{false}), P(\text{true})]$).
• 64k Long-Context Retention: Non-decaying associative memory across 64,000 raw tokens using KDA Householder overwriting and Decoupled Decision Readout.

⚡ Measured Hardware Latency on NVIDIA RTX PRO 6000 Blackwell

Dual-GEMM Concurrent Tri-Track execution with Born-in-FP8 Tensor Cores, Stride-Safe Triton SwiGLU, and Fused RMSNorm:

💡 Batched serving ($B=128, T=1024$) achieves 8.7 Million Tokens/sec (8.7 MTok/s). For interactive single queries ($B=1$), DJINN executes a complete 1.01B-parameter decision in just 1.6 to 3.5 ms (over 280 to 625 decisions per second).

🚀 Quickstart: Lightning-Fast Inference

Install dependencies:

pip install torch triton fla-linear-attention huggingface_hub

Option A: Using DJINNClient (TypeSafe System One API)

from djinn import DJINNClient, Choice, Score, Noul

client = DJINNClient()

response = client.system_one(
    state={
        "cluster": "node-blackwell-07",
        "gpu_temp_c": 104,
        "ecc_correctable": 3,
        "active_vram_gb": 93.6
    },
    questions={
        "severity": Score(
            instructions="Rate incident urgency on 1-10 scale:",
            levels=["P4 - Low", "P3 - Moderate", "P2 - High", "P1 - Critical"]
        ),
        "triage_action": Choice(
            instructions="Select immediate mitigation action:",
            criteria={
                "throttle": "Throttle GEMM power target to 400W",
                "fan_override": "Set fans to emergency 100% boost",
                "drain_node": "Drain node and migrate workloads",
                "no_action": "Continue monitoring"
            }
        ),
        "service_down": Noul(instructions="Does this require immediate cluster shutdown?")
    }
)

print("Triage Action:", response.answers["triage_action"].choice)
print("Severity Score:", response.answers["severity"].score)
print(f"End-to-End Latency: {response.latency_ms:.2f} ms")

Option B: Using DJINNInferenceEngine Directly

from djinn.infer import DJINNInferenceEngine

# Automatically loads phase_4_final.pt from cache or Hub
engine = DJINNInferenceEngine()

# 1. Multi-Alternative Choice Decision
res = engine.predict_choice(
    state="Autonomous compiler optimization pass analyzing hot loop 4. Register pressure 38/64.",
    question="Which loop unroll factor minimizes execution stalls?",
    options=["Unroll factor 2", "Unroll factor 4", "Unroll factor 8", "Do not unroll"]
)

print(f"Selected Decision: {res['best_option']}")
print(f"Confidence:        {res['confidence'] * 100:.1f}%")
print(f"Hardware Latency:  {res['latency_ms']:.2f} ms | Bucket: T={res['bucket_len']}")
print(f"Probabilities:     {res['probabilities']}")

# 2. Ordinal Score Estimation (1-10 Scale)
score_res = engine.predict_score(
    state="Emergency thermal spike detected on GPU cluster node 42.",
    question="Rate incident severity on 1-10 scale:"
)
print(f"Severity Score:    {score_res['expected_score']:.2f} / 10.0 (Latency: {score_res['latency_ms']:.2f} ms)")

# 3. Binary Hypothesis Testing (Noul)
noul_res = engine.predict_noul(
    state="Database query execution plan shows missing index on foreign key column.",
    question="Is this query vulnerable to table scan degradation?"
)
print(f"Hypothesis Valid:  {noul_res['decision']} (P(True): {noul_res['p_true']*100:.1f}%)")

CLI Inference & Benchmark

# Run interactive CLI decision
python infer.py --state "Database queue full" --question "Select scaling action:" --options "Add worker" "Throttle rate" "Drop queue"

# Run full hardware benchmark across T=128..64k
python infer.py --benchmark

📊 4-Phase Curriculum Training Scorecard

📦 Checkpoints & Model Files

• 🏆 phase_4_final.pt (2.96 GB): Master production weights with calibrated decision heads.
• 📦 phase_3_final.pt (2.96 GB): 64k long-context needle retention checkpoint.
• 📦 phase_2_final.pt (2.96 GB): Context-scaled checkpoint ($T=4096$).
• 📦 phase_1_final.pt (2.96 GB): Foundation pre-trained checkpoint.
• ⚙️ config.json: Official DJINN 1.005B Dual-GEMM parameter and hardware specifications.
• ⚡ infer.py: Lightning-fast production inference CLI & Python API.
• ⚡ client.py: High-level TypeSafe client API.
• ⚡ serving.py: High-performance serving engine.
• 📊 Metrics: metrics_phase_1.jsonl | metrics_phase_2.jsonl | metrics_phase_3.jsonl | metrics_phase_4.jsonl