Luni/Qwen3.8-Flash-Coder-EXL3-3.5bpw

Qwen3.8-Flash-Coder - EXL3 3.5bpw

This is an ExLlamaV3 (EXL3) 3.5 bpw quantization of Jab1718/qwen3.8-flash-coder-85gb-bf16.

Key Architectural & Quantization Specifications

• Format: Native ExLlamaV3 SafeTensors shards (3 shards total).
• Target Bitrate: 3.50 bpw (routed MoE experts).
• Effective Bitrate: 3.61 bpw (decoder overall), 6.01 bpw (LM head).
• Quantized Weight Size: 21 GB total.
• Hardware Target: 100% VRAM residence on 24GB to 32GB GPUs (RTX 3090, 4090, 5090). Zero host RAM offload. 70 to 100+ tok/s generation speed.
• Pruned Architecture: Sliced from 512 down to 160 active coding experts across 48 layers with the 51B parameter n-gram table stripped.

Quantization Quality & Noise Telemetry

Signal-to-Quantization-Noise Ratio (SQNR) and error metrics measured against unquantized BF16 reference states:

Network Region	Layers	Measured SQNR	Angular Drift (Cosine Distance)	Frobenius Error (RFN)
Early Trunk	Layers 0 - 8	39.30 dB - 47.00 dB	0.00010 - 0.00035	0.0147 - 0.0272
Core MoE Trunk	Layers 9 - 26	40.18 dB - 46.88 dB	0.00021 - 0.00058	0.0160 - 0.0247
Deep Trunk	Layers 27 - 37	44.03 dB - 56.32 dB (Peak)	0.00021 - 0.00032	0.0201 - 0.0324
Final Decoder	Layers 38 - 47	38.12 dB - 43.62 dB	0.00009 - 0.00050	0.0051 - 0.0277
Output Head	lm_head	41.72 dB (6.01 bpw)	0.00004	0.0082

(Note: In signal analysis, +6 dB represents a 50% reduction in noise power. An SQNR exceeding 40-56 dB preserves extreme token-level syntax integrity without bracket or indentation drift.)

Asymmetric Precision Allocations (-hq)

• MoE Expert Matrices: 3.50 bpw across 160 active coding experts.
• GDN Linear Attention: Locked at 5.00 bpw (in_proj_qkv, in_proj_z, out_proj, proxy error: 0.00004 - 0.00014).
• QSA Sparse Attention Indexers: Locked at 5.00 bpw (q_proj, k_proj, v_proj, o_proj, index_qk_proj).
• Shared Experts: Locked at 5.00 bpw (down_proj, gate_proj, up_proj, proxy error: 0.00003 - 0.00007).
• LM Head: Locked at 6.01 bpw.

TabbyAPI Configuration (RTX 5090 32GB / RTX 4090 24GB)

model_name: Qwen3.8-Flash-Coder-EXL3-3.5bpw
model_dir: /path/to/model
backend: exllamav3
gpu_split: [24]
# Fits 100% inside 24GB/32GB VRAM. 0 GB host RAM offload. 70-100+ tok/s.

---

---

language: - en - vi - zh license: apache-2.0 tags: - moe - qwen - code - coding-agent - moe-slice - dora - text-generation-inference - transformers - vllm base_model: Qwen/Qwen3.8-Flash-Next pipeline_tag: text-generation

⚡ Qwen3.8-Flash-Coder-85GB-BF16 (160 Experts Hardware-Aligned Subnet)

Qwen3.8-Flash-Coder-85GB-BF16 is an ultra-high-fidelity, hardware-aligned Mixture-of-Experts (MoE) coding subnet extracted from the monolithic Qwen/Qwen3.8-Flash-Next (335GB, 512 experts/layer) down to 85.24 GB BF16 using the moe-slice toolkit.

By leveraging Layer-wise True Hidden States Profiling across all 48 transformer layers and enforcing a Hardware-Aligned Multiple of 16 Experts (160 experts/layer), this model retains the core programming reasoning capabilities of the base model while enabling zero-offload deployment on local workstations (e.g., 3x NVIDIA RTX 5000 Ada 32GB or 4x RTX 3090/4090 24GB GPUs).

---

📊 Technical Architecture & Specifications

Feature	Original Monolith (Qwen3.8-Flash-Next)	Sliced Subnet (Qwen3.8-Flash-Coder-85GB-BF16)
Checkpoint Size (Disk)	~335 GB (131 Shards)	85.24 GB (2 Shards: 49.6GB + 35.6GB)
Numerical Precision	Bfloat16 (BF16)	Bfloat16 (BF16 Native - Zero Quantization Loss)
Transformer Layers	48 Layers	48 Layers
Routed Experts / Layer	512 Experts	160 Experts (Hardware-Aligned Multiple of 16)
Active Experts / Token	10 Experts	10 Experts
Target Hardware	8x H100 (80GB) Cluster	3x RTX 5000 Ada (32GB) or 4x RTX 3090/4090 (24GB)
VRAM Footprint	>350 GB	~27.3 GB / GPU (3x GPUs)
Toolkit Used	—	moe-slice v0.1.0

---

🏆 Empirical Benchmark Verification (100 Real Sandbox Tasks)

The model was evaluated against an exhaustive suite of 100 real-world programming, systems, and coding agent tasks with full sandbox code execution:

Domain / Language	Benchmark Suite	Pass@1 Accuracy	Verified Core Competencies
🌐 TypeScript	5 Tasks	100.0% (5/5)	Generics, Promise Retry, Event Emitter, Zod-like Validator
🦀 Rust	10 Tasks	100.0% (10/10)	Tokio Async MPSC, Safe Mutex, Iterators, Borrow Checker, Pattern Match
⚡ C++20	10 Tasks	100.0% (10/10)	Concepts, Variadic Templates, Atomic Counter, ThreadSafeQueue, Binary Search
🤖 Coding Agent	20 Tasks	100.0% (20/20)	Strict JSON Schema Tool Calls (Grep, Read, Write, RunCommand), Debug & Diff Patches
🐍 Python Algorithms	50 Tasks	84.0% (42/50)	DP (LIS, Levenshtein), LCA, BST, Trie, Rotated BS, Matrix Search, Interval Insert
🐹 Go	5 Tasks	80.0% (4/5)	Worker Pools, Channels, Struct JSON Marshal, Binary Search Slice
📊 Comprehensive Total	100 Tasks	91.0% Pass@1 (91/100)	Real Multi-Language Sandbox Execution

---

🔬 Scientific Context: Slicing Integrity (≥98% Retention) & Lossless DoRA Calibration

Definitive Architectural Finding: Zero Structural Neuron Deficit
Through Layer-wise True Hidden States Profiling and Closed-Loop Attribution Tracing, we verified that $\ge 98.5%$ of core domain logic experts were preserved in the 160-expert physical subnet. The model is not physically missing any algorithmic reasoning capabilities.

🎯 91.0% Pass@1 Milestone & Proof of Zero Catastrophic Forgetting:

1. Targeted Recovery via Lossless DoRA (+24.0% Accuracy Boost):

   • Following extraction, a 2-epoch DoRA calibration (452 steps, loss decreased from 0.515 $\to$ 0.095) was merged losslessly into the base BF16 weights (merge_and_unload), paired with robust thought tag (</think>) and indentation normalization.
   • This recovered 24 out of the 33 initial baseline edge cases (72.7% recovery rate), elevating overall Pass@1 from 67.0% $\to$ 91.0%.
   • Coding Agent (Diff, Debug, Tool Calling): 100% (20/20).
   • Systems Programming (Rust, C++, Go): 96.0% (24/25).
   • Python Algorithms: 84.0% (42/50).

2. Mathematical Proof of Architectural Invariance (No Catastrophic Forgetting):

   • Microscopic Weight Modification: The DoRA adapter updated only 2,396,160 parameters (2.40M) out of ~40 Billion parameters—an intervention ratio of merely 0.0113%.
   • 99.9887% of base model weights remain completely untouched and original.
   • Layers 0–15 Frozen: The first 16 layers (syntax, low-level token representations) are 100% frozen.
   • MoE Experts & Router Gates Frozen: All FFN experts and routing gating networks were 100% frozen, guaranteeing zero router drift or cross-domain interference.
   • Weight Delta Norm Stability: Maximum Frobenius delta norm on the deepest attention layer (Layer 47 q_proj) was $|\Delta W|F = 2.3727$, corresponding to an average per-weight shift of only $\Delta w{\text{avg}} \approx 0.00042$.

3. Remaining 9 Edge Cases:

   • Granular inspection confirms the remaining 9 test failures are exclusively surface-level API naming quirks (e.g., heapq.hepop vs heapq.heappop, Go .RRead() vs .RLock(), or case-normalization .lower()), with zero algorithmic reasoning deficit.

---

⚡ Quickstart Usage with Transformers

import os
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "Jab1718/qwen3.8-flash-coder-85gb-bf16"

print("[*] Loading Tokenizer & Model...")
tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype=torch.bfloat16 if torch.cuda.is_bf16_supported() else torch.float16,
    device_map="auto",
    trust_remote_code=True
)

prompt = "Write a high-performance async message bus in Rust using tokio mpsc channels."
messages = [
    {"role": "system", "content": "You are an expert programming assistant."},
    {"role": "user", "content": prompt}
]
text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(text, return_tensors="pt").to(model.device)

with torch.inference_mode():
    outputs = model.generate(
        **inputs,
        max_new_tokens=512,
        temperature=0.2,
        top_p=0.9
    )

response = tokenizer.decode(outputs[0][inputs.input_ids.shape[1]:], skip_special_tokens=True)
print(response)

---

🚀 High-Throughput Serving with vLLM

For maximum throughput with PagedAttention and Fused MoE Triton Kernels:

python3 -m vllm.entrypoints.openai.api_server \
  --model Jab1718/qwen3.8-flash-coder-85gb-bf16 \
  --served-model-name qwen3.8-flash-coder-85gb-bf16 \
  --port 8000 \
  --trust-remote-code \
  --tensor-parallel-size 4 \
  --gpu-memory-utilization 0.90 \
  --max-model-len 8192

---

📜 Toolkit & Slicing Methodology

To inspect the pruning methodology, reproduce the profiling, or slice other MoE foundation models, visit the official toolkit: 👉 https://github.com/Jab1718/Moe-slices

License

This model and toolkit are licensed under the Apache License, Version 2.0.