Highlights
- Single-Engine Dual Serving (
warp-34 + PEFT): Serve standard causal chat completions (lora_id = -1) and non-causal /v1/systemone structured decision queries (lora_id = "glide") from a single loaded Qwen3.5-9B backbone.
- Shared Fused Input Projections (
lora_A Deduplication): Fused sibling projections (q_proj/k_proj/v_proj, gate_proj/up_proj, in_proj_qkv/in_proj_z, in_proj_b/in_proj_a) share an identical lora_A.weight computed via Joint Teacher-ASVD (128 calibration prompts from SargeDev/jev-distill-corpus-v3 with GLiDE-9B teacher activations), reducing fused LoRA A projection FLOPs and VRAM by 2x-3x when deduplicated by the inference engine.
- Exact Offline
GemmaRMSNorm + RMSNormGated Folding: Pre-norm scale shifts (1 + g_glide) / (1 + g_base) (input_layernorm, post_attention_layernorm), linear_attn.norm (g_glide / g_base into linear_attn.out_proj), and final model.language_model.norm (folded into readout.safetensors) are absorbed into the adapter and readout weights with zero runtime overhead.
JevBench Public Eval (231 Cases: 48 Easy / 72 Original / 111 Hard)
Evaluated on kshitijthakkar/nirnaya-jevbench-public-eval-v1 using the native GLiDE quick-empty-think-medium-v1 prompt template and noncausal_full_attention:
Table with columns: Model, Overall Acc, Easy (48), Original (72), Hard (111), NLL (↓), Brier (↓)| Model | Overall Acc | Easy (48) | Original (72) | Hard (111) | NLL (↓) | Brier (↓) |
|---|
Base Qwen/Qwen3.5-9B (passes=1) | 194 / 231 (83.98%) | 48 / 48 (100.0%) | 70 / 72 (97.22%) | 76 / 111 (68.47%) | 0.4190 | 0.2212 |
Base Qwen/Qwen3.5-9B (passes=2 midloop) | 195 / 231 (84.42%) | 48 / 48 (100.0%) | 70 / 72 (97.22%) | 77 / 111 (69.37%) | 0.4188 | 0.2214 |
Qwen3.5-9B + GLiDE-LoRA (r=128, passes=1 single-pass) | 205 / 231 (88.74%) | 48 / 48 (100.0%) | 72 / 72 (100.0%) | 85 / 111 (76.58%) | 0.2819 | 0.1592 |
Qwen3.5-9B + GLiDE-LoRA (r=128, passes=2 midloop) | 207 / 231 (89.61%) | 48 / 48 (100.0%) | 72 / 72 (100.0%) | 87 / 111 (78.38%) | 0.2554 | 0.1410 |
Full fastino/GLiDE-9B (passes=1 single-pass) | 211 / 231 (91.34%) | 48 / 48 (100.0%) | 72 / 72 (100.0%) | 91 / 111 (81.98%) | 0.2068 | 0.1173 |
Full fastino/GLiDE-9B (passes=2 midloop) | 214 / 231 (92.64%) | 48 / 48 (100.0%) | 72 / 72 (100.0%) | 94 / 111 (84.68%) | 0.2049 | 0.1103 |
Repository Contents
adapter_config.json & adapter_model.safetensors: Standard PEFT LoRA weights (r=128, lora_alpha=128, rank_pattern & alpha_pattern = 64 on linear_attn.in_proj_{a,b}, bfloat16).
decision_config.json: GLiDE-9B decision runtime metadata (attention_mode: "noncausal_full_attention", prompt_format: "quick-empty-think-medium-v1", midloop: {"blocks": [[12, 17], [18, 23]], "passes": 2}, 255 option codes A..GC).
readout.safetensors: 255-way CodeReadout head ([255, 4096] in bfloat16, with final GemmaRMSNorm ratio folded in).