Joe lafrite

joelafrite's activity

• All
• Models
• Datasets
• Spaces
• Buckets
• Papers
• Collections
• Community
• Posts
• Upvotes
• Likes
• Articles

New activity in drowzeys/keys-DeepSeek-V4.1-Flash-1017-Abliterated 12 days ago

2x DGX Spark (Native Zig 1M TP=2) benchmark receipt: 0% refusal and zero coding regression vs stock LibertAI 2.9bpw

#7 opened 12 days ago by

joelafrite

New activity in unsloth/Qwen3.8-27B-NVFP4 12 days ago

This checkpoint ships calibrated fp8 KV-cache scales, but vLLM never loads them

👍 2

#28 opened 12 days ago by

joelafrite

New activity in LibertAIDAI/GLM-5.3-Flash-NVFP4 24 days ago

Datapoint: fp4 vs fp8 GLM-5.3-Flash endpoint parity on GSM8K + HumanEval

2

#8 opened 2 months ago by

joelafrite

New activity in joelafrite/Qwen3.8-27B-fp8-KV-scale-calibration 25 days ago

How to generate the calibrated KV Scales?

1

#1 opened 25 days ago by

Goghor

New activity in LibertAIDAI/DeepSeek-V4.1-Flash-1017-EXL3-2.9bpw 25 days ago

Thank you for the quick upload

➕ 1

3

#1 opened 26 days ago by

joelafrite

New activity in dealignai/DeepSeek-V4.1-Flash-1017-EXL3-2.9bpw-UNCENSORED 25 days ago

All-EXL3 variant? `wo_b` / `shared_experts.w2` are BF16 in layers 10–36

❤️ 1

#1 opened 25 days ago by

joelafrite

New activity in RadixArk/Qwen3.8-Flash-Next-NVFP4 about 1 month ago

Degeneration Loop when calling tool (PDF/Image)

🔥 1

4

#3 opened 3 months ago by

dangnsh

New activity in gratex/Qwen3.8-27B-DFlash2-W4A16-g128-sym-GPTQ 2 months ago

apply_dense_kv_fix.py: silent corruption risk with NVFP4 drafters (+ working variant)

3

#1 opened 2 months ago by

joelafrite

New activity in QUASAR-QAT/Qwen3.8-27B-QUASAR-NVFP4 2 months ago

Heads-up: config declares an MTP head but no mtp.* tensors ship (silent 0%-acceptance trap) — and a working graft recipe

3

#3 opened 2 months ago by

joelafrite

New activity in peculiar-ragdoll/Qwen-Sharp-Chat-Templates 2 months ago

Production adoption report (vLLM + MTP): terseness confirmed, quality up at n=4 — and the MTP think-leak did NOT reproduce

❤️ 1

1

#11 opened 2 months ago by

joelafrite

New activity in orcarouter/Qwen3.8-27B-Uncensored-NVFP4 2 months ago

vLLM boot failure: quantization target literal `lm_head` never matches — one-line fix

1

#3 opened 2 months ago by

joelafrite

New activity in RadixArk/Qwen3.8-27B-DSpark 2 months ago

v2 validated on vLLM + NVFP4 target (RTX 5090): AL ~3.22 confirms your claim — with a 32GB context warning

#8 opened 2 months ago by

joelafrite

New activity in maurienne-ai/Qwen3.8-27B-DFlash2-NVFP4-RTNcal 2 months ago

config.json `producer` string breaks vLLM draft-model loading (+ numbers on vLLM nightly)

1

#1 opened 2 months ago by

joelafrite

New activity in natfii/Qwen3.6-27B-VLM-Cascade-NVFP4-MTP 4 months ago

Great model

❤️ 1

#1 opened 4 months ago by

joelafrite

New activity in sakamakismile/Qwen3.6-27B-NVFP4 7 months ago

MTP not responding

9

#7 opened 7 months ago by

joelafrite

MTP not responding

9

#7 opened 7 months ago by

joelafrite