ege-konic's picture
Upload folder using huggingface_hub
6ee5b26 verified
|
Raw
History Blame Contribute Delete
2.36 kB
metadata
license: other
license_name: lfm-open-license-v1.0
license_link: https://huggingface.co/LiquidAI/LFM2.5-VL-3B/blob/main/LICENSE
tags:
  - vision-language-model
  - lfm2
  - compression
  - quantization
  - gptq
  - compressed-tensors
  - vllm
  - konic

LFM2.5-VL-3B-ffn8192-int4

The most compressed tier of our LFM2.5-VL-3B surgical-compression series: FFN pruned 10752 → 8192 + distillation recovery, then mixed-precision quantization — native INT4/INT8 weights in the standard compressed-tensors format (vLLM-ready).

Size vs original
Original BF16 6.25 GB
Pruned+recovered BF16 5.30 GB −15%
This checkpoint (INT4/INT8) 2.22 GB −69%

Quantization allocation (hand-rolled GPTQ-class)

Module group Scheme
FFN w1/w2/w3 (90) W4A16 GPTQ, per-(row, 128-group) asymmetric, zero-point
LM attention + vision tower linears (195) W8 per-channel
conv projections / projector / patch-embedding / embeddings BF16 (vLLM 0.26 LFM2 limitation — plain linears)
Norms / conv1d / biases BF16

A nibble-packed variant of this model (1.93 GB, custom format + loader) is the official compression deliverable — see the pipeline repo.

Quality vs base (quick suite, our harness)

Axis Value
text PPL ratio (e^ΔNLL) 0.83 (≤1.05 gate: PASS)
mm PPL ratio 0.79 (≤1.10 gate: PASS)
text kld_t64 0.192 (base = 0)
mm kld_t64 0.229
generation chrF (text/mm) 0.54 / 0.51
integrity index (record-only) 42.5 / 100

Serve with vLLM (native INT4 — no dequant)

from vllm import LLM
llm = LLM(model="konic-labs/LFM2.5-VL-3B-ffn8192-int4",
          quantization="compressed-tensors")

Live-verified in vLLM 0.26 (L4, greedy): same answers as the base on 7 image/text prompts, and tool-calling intact (5 tools × 10 rounds: right tool + right args everywhere; correct abstention on no-tool prompts).

License & provenance