Views
No views yet
CYFRAGOVPL/Llama-PLLuM-8B-chat-2512 for single-GPU vLLM deployment on consumer/prosumer hardware. Validated on 1× AMD Radeon AI PRO R9700 (gfx1201, TP=1); the compressed-tensors format is expected to be portable to NVIDIA via awq_marlin — not independently validated in this release.⚠ Note on Hugging Face parameter display: Hugging Face may show this quantized repository with reduced parameter count because the safetensors index counts packed 4-bit tensors. The logical base architecture is Llama-PLLuM-8B (8.03B parameters); this repository contains AWQ-compressed derivative weights of the 8B model. Verify with the base model card:CYFRAGOVPL/Llama-PLLuM-8B-chat-2512.
compressed-tensors) modelu Llama-PLLuM-8B-chat-2512 wytworzonego przez konsorcjum PLLuM (SpeakLeash, OPI-PIB, NASK, Politechnika Wrocławska), publikowanego na HuggingFace przez Ministerstwo Cyfryzacji RP (organizacja CYFRAGOVPL). Celem kwantyzacji jest deployment na pojedynczej karcie konsumenckiej — 5.53 GiB footprintu pozwala uruchomić 8B parametrów na 1× AMD Radeon AI PRO R9700 (32 GB VRAM) z dużym headroomem dla KV cache (22.22 GiB available, max_concurrency 88.89× przy max_seq_len=2048).compressed-tensors) tego wariantu Llama-PLLuM-8B (stan na 2026-05-26). Dla użytkowników llama.cpp/ollama dostępne są GGUF quantizacje społeczności; tutejsze wagi są zoptymalizowane pod inferencję serwerową w vLLM.compressed-tensors format) of Llama-PLLuM-8B-chat-2512 developed by the PLLuM consortium (SpeakLeash, OPI-PIB, NASK, Wrocław University of Science and Technology), published on HuggingFace by the Polish Ministry of Digital Affairs (CYFRAGOVPL). Quantized for single-GPU consumer deployment: 5.53 GiB footprint fits comfortably on a single 32 GB R9700 with large KV cache headroom (22.22 GiB → 88.89× concurrency at max_seq_len=2048), served via vLLM. To the author's knowledge, this is the first public AWQ W4A16 (vLLM-native compressed-tensors) quantization of this Llama-PLLuM-8B-chat-2512 checkpoint (HuggingFace Hub check, 2026-05-26).| Parameter | Value |
|---|---|
| Method | AWQ W4A16 (activation-aware) |
| Library | llm-compressor |
| Format | compressed-tensors (vLLM-native) |
| Group size | 128 |
| Bits / weights | 4 |
| Bits / activations | 16 (bf16) |
| Calibration host | AMD Radeon AI PRO R9700 (gfx1201, RDNA 4), local workstation — 2× R9700 (64 GB VRAM combined), Ryzen 9 9950X3D, Kubuntu 24.04, ROCm 7.2.1 |
| Calibration corpus | mozarcik/clinical-pl-smpc-awq-calibration — 418 chunks (~512 tokens) polskich Charakterystyk Produktów Leczniczych (SmPC) z EMA; pulmonologia + onkologia klatki piersiowej; No PHI |
| Skipped layers | lm_head (standard) |
| Recipe | AWQModifier(scheme="W4A16", targets="Linear", ignore=["lm_head"]) — patrz recipe.yaml w repo |
navimed-umb/calibration/quantization/.max_seq_len=2048, enforce_eager=True:| Metric | Value |
|---|---|
| Footprint (model weights) | 5.53 GiB |
| KV cache available | 22.22 GiB |
| Maximum concurrency @ 2048 tok/req | 88.89× |
| Init engine time | 3.56 s |
| vLLM ready time (from cold start) | ~30 s |
| Sanity response time (5 PL prompts, max_tokens=64) | ~6 s/prompt |
navimed-umb/environment/sanity-tests/2026-05-26-Llama-PLLuM-8B-chat-2512-awq-sanity.json.Co to znaczy "envelope": to są twarde liczby pamięci/ładowania/pojemności KV — czyli "czy model w ogóle się mieści i odpowiada poprawnie na pojedyncze zapytanie". To nie są liczby throughput (tok/s, P50/P95, scaling z N) — te pozostają EMBARGOED do akceptacji paper'a (METHODOLOGY §11.2, stricter embargo dla modeli polskich §11.3).
/v1/completions, temperature=0.0, max_tokens=64. Vehicle-integrity per METHODOLOGY §8 — NIE ocena jakości modelu."Tiotropium to lek wziewny stosowany w leczeniu astmy oskrzelowej i przewlekłej obturacyjnej choroby płuc (POChP). Zawiera on substancję czynną tiotropiumbromek, która działa na receptory muskarynowe M3 znajdujące [...]"
⚠ Important — ignore Hugging Face autogenerated snippets. The "Use this model" button on the HF UI may suggest snippets likepipeline("text-generation", model="...")or a barevllm serve "..."— these are incomplete for this quantized AWQ release and will fail or misbehave without the flags below (ROCm env vars,--quantization compressed-tensorsorawq_marlin,--enforce-eageron gfx1201).
0.19.0+rocm721 PINNED + ROCm 7.2 + PyTorch 2.10 (regressions reported on newer ROCm wheels on gfx1201, cf. Capitelli #40980). Build wheel from ROCm vLLM fork or use AMD-provided container — NOT available on PyPI. Setup: navimed-umb environment/.1# Required env vars for gfx1201
2export VLLM_ROCM_USE_AITER=0 # AITER unsupported on gfx1201
3export AMD_SERIALIZE_KERNEL=1
4export HIP_LAUNCH_BLOCKING=1
5export ROCR_VISIBLE_DEVICES=0 # single GPU; adjust for your card
6
7vllm serve mozarcik/Llama-PLLuM-8B-chat-2512-awq \
8 --tensor-parallel-size 1 \
9 --enforce-eager \
10 --max-model-len 2048 \
11 --gpu-memory-utilization 0.90 \
12 --quantization compressed-tensors1vllm serve mozarcik/Llama-PLLuM-8B-chat-2512-awq \
2 --tensor-parallel-size 1 \
3 --max-model-len 4096 \
4 --gpu-memory-utilization 0.85 \
5 --quantization awq_marlinNote for NVIDIA users: drop--enforce-eager(CUDA graphs work on Hopper/Ampere/Ada); useawq_marlinquantization flag (faster on NVIDIA thancompressed-tensors); lowergpu_memory_utilizationto 0.85 to avoid OOM with fragmentation. Ifawq_marlinis rejected by your vLLM version, fall back to--quantization awq.
1curl http://localhost:8000/v1/chat/completions \
2 -H "Content-Type: application/json" \
3 -d '{
4 "model": "mozarcik/Llama-PLLuM-8B-chat-2512-awq",
5 "messages": [{"role": "user", "content": "Wyjaśnij krótko zasadę działania spirometrii."}],
6 "max_tokens": 256,
7 "temperature": 0.3
8 }'1from openai import OpenAI
2client = OpenAI(base_url="http://localhost:8000/v1", api_key="dummy")
3
4resp = client.chat.completions.create(
5 model="mozarcik/Llama-PLLuM-8B-chat-2512-awq",
6 messages=[{"role": "user", "content": "Wyjaśnij krótko zasadę działania spirometrii."}],
7 max_tokens=256, temperature=0.3,
8)
9print(resp.choices[0].message.content)awq_marlin performance jest typowa dla AWQ Marlin kernels (cf. vLLM docs).llm-compressor v_proj skip on GQA — known library artifact dla Grouped-Query Attention (Llama 3.1 family); patrz vllm-project/llm-compressor#issues. Wpływ na koherencję outputu zweryfikowany Gate 2 probe — patrz sekcja powyżej./v1/chat/completions z chat template z tokenizera.CYFRAGOVPL/Llama-PLLuM-8B-chat-2512 (Llama-PLLuM-8B family by the PLLuM consortium, published by the Polish Ministry of Digital Affairs) is derived from Meta's Llama 3.1 and is governed by the Llama 3.1 Community License Agreement and the Llama 3.1 Acceptable Use Policy.Llama 3.1 is licensed under the Llama 3.1 Community License, Copyright © Meta Platforms, Inc. All Rights Reserved.
CYFRAGOVPLCYFRAGOVPL/Llama-PLLuM-8B-chat-2512mozarcikmozarcik/clinical-pl-smpc-awq-calibration (License: see dataset card; EMA reproduction policy applies)0.19.0+rocm721navimed-umb v0.4.0 — github.com/kicrazom/navimed-umbllama3.1 — inherited from CYFRAGOVPL/Llama-PLLuM-8B-chat-2512.LICENSE)USE_POLICY.md)NOTICE) reproducing the required Meta copyright text1@software{minarowski_2026_pllum8b_chat_2512_awq,
2 author = {Minarowski, Łukasz},
3 title = {{Llama-PLLuM-8B-chat-2512} AWQ W4A16 quantization for vLLM on consumer AMD},
4 year = {2026},
5 publisher = {HuggingFace},
6 url = {https://huggingface.co/mozarcik/Llama-PLLuM-8B-chat-2512-awq},
7 doi = {10.5281/zenodo.19851346}
8}1@article{kocon2025pllum,
2 title = {PLLuM: A Family of Polish Large Language Models},
3 author = {Kocoń, Jan and others},
4 journal = {arXiv preprint arXiv:2511.03823},
5 year = {2025}
6}