satgeze/Qwen3.5-0.8B-DSpark
vLLM verified at llama.cpp parity; correct the cross-stack section: the 3.7x gap was our config rope bug, not runtime semantics
Fix rope_parameters: partial_rotary_factor 1.0 explicit (qwen3_5_text class default re-injects 0.25 if key deleted). Restores vLLM acceptance to llama.cpp parity (0.155 vs 0.158 per-drafted-token)
remove write-probe file
Upload training_and_scaling.png with huggingface_hub
Upload training_v02_v03.png with huggingface_hub
Upload v01_v02_v03_ab.png with huggingface_hub
Upload config_v03.json with huggingface_hub
Upload model_v03.safetensors with huggingface_hub
Upload head_q08_v03.gguf with huggingface_hub
v0.3: 27.6k self-distilled samples, monotone acceptance gains, D1 shown at 744B
probe: verify Grok seat can write to hub after settings allow
link the published self-distillation dataset
Correction: both versions were self-distilled; v0.2s real variable is ~9x supervised tokens (full-length responses), not distribution matching
Correction: v0.1/v0.2 differed in global batch (256 vs 128); control caveat added, training chart embedded
training-loss comparison chart, epoch-normalized
embed v0.1 vs v0.2 chart in the experiment section
v0.1 vs v0.2 comparison chart
v0.2 release: distribution-matching experiment section, recommended head updated
v0.2 head: own-regen training data, +49%/+30% relative acceptance at constant volume
v0.2 head: own-regen training data, +49%/+30% relative acceptance at constant volume
v0.2 head: own-regen training data, +49%/+30% relative acceptance at constant volume
Controlled cross-stack A/B: same head, llama.cpp 15.8% vs vLLM 4.3% per-drafted-token under matched conditions; spec-type silent-fallback repro note
Loop story complete: DRY mitigation table (rescues >=2.9bpw, not IQ1_S), depth + raw-vs-chat scoping, think-field loop warning
Add measured loop-rate ladder: this 0.8B target degenerates on long greedy generation at every precision including bf16 (not a quantization effect); DRY mitigation measured
Add draft-width note (default 3 undersells a block-7 head)
Correction: refute the quant-tax claim; rewrite consequences with measured evidence; label uncalibrated Q2_K rung
Link the GGUF capture-training how-to guide
vLLM smoke PASSED on 0.25.1 (loads + speculates); Blackwell setup notes; version floor
A/B landed: capture-trained head reproduces HF pipeline (validation); no matched-head advantage at Q8 (expected)
Add measured SPEED-Bench section (Metal, spec vs base, diverse-prompt acceptance); clear from pending list
Card status note: pending rows listed, landing within a day
SGLang section: point at tracking issue #30344
Fix vLLM section: DSpark merged in mainline; direct serve command + honest untested marker
Card raised to 27B standard: own quant-tax ladder (5 rungs incl own-imatrix IQ1_S), economics ladder chart, How-to-run commands, training narrative
Add Apple Silicon (M3 Max Metal) measured rows: acceptance identical to CUDA, platform economics
Replace extrapolated IQ1_S row with measured value (0.074, imatrix-quantized target)
Add quant-tax finding: acceptance vs target precision (Q8 0.365 / Q2_K 0.22 / ~1-bit ~0.06); match heads to the deployed artifact
Link DeepSpec fork branch (qwen35-lane) used for training
Add official DeepSpec eval numbers (9 datasets) + confidence calibration
v0.1 experimental: from-scratch 12K-sample DSpark head + GGUF + honest card
initial commit