loom-fun/bartowski-Qwen2.5-0.5B-Instruct-loom

Qwen2.5-0.5B-Instruct (Q4_K_M) for AMD Loom

A Loom package (.loom programs and a model.json card) for the unmodified GGUF checkpoint bartowski/Qwen2.5-0.5B-Instruct-GGUF (Qwen2.5-0.5B-Instruct-Q4_K_M.gguf), generated on an AMD Ryzen AI Max+ 395 "Strix Halo" with the loom-skills workflow.

Quick start

pip install "hrx-loom[hf]"
loom-fetch https://huggingface.co/loom-fun/bartowski-Qwen2.5-0.5B-Instruct-loom/resolve/main/model.json --materialize=copy --output=Qwen2.5-0.5B-Instruct
loom-serve --model=Qwen2.5-0.5B-Instruct/model.json --device=amdgpu --memory_bytes=5000000000

The server listens on 127.0.0.1:8080 and answers OpenAI-style /v1/chat/completions requests.

Performance vs llama.cpp

metric	llama.cpp	Loom	margin	roofline	Loom % of roofline	DevLab llama.cpp	DevLab Loom	local ÷ DevLab (llama.cpp, Loom)
decode tok/s, short prompt	284.8	380.6	+34 %	675.8	56 %	263.0	312.2 ¹	1.08, 1.22
decode tok/s, 2k context	245.8	543.2	+121 %	1181.6	46 %	266.1	435.9 ¹	0.92, 1.25
decode tok/s, 32k context	225.4	249.1	+11 %	728.0	34 %	220.8	206.7 ¹	1.02, 1.21
prefill tok/s, 1k document	11169.0	16879.0	+51 %	77768.6	22 %	11575.8	17097.2 ¹	0.96, 0.99
prefill tok/s, 2k document	9177.1	16644.3	+81 %	73596.4	23 %	10537.1	17641.3 ¹	0.87, 0.94
prefill tok/s, 8k document	8002.1	14605.9	+83 %	55675.2	26 %	9015.2	16542.5 ¹	0.89, 0.88
prefill tok/s, 16k document	5502.5	11529.9	+110 %	42029.3	27 %	5803.9	12692.8 ¹	0.95, 0.91
prefill tok/s, 32k document	3774.1	8008.7	+112 %	28203.9	28 %	3932.5	7848.4 ¹	0.96, 1.02
4 streams, aggregate decode tok/s	791.2	938.8	+19 %	2850.3	33 %	816.6	635.6 ¹	0.97, 1.48
speculative decoding, drafted tokens accepted	—	12 % ²	—	n/a	n/a	n/a	n/a	n/a

• llama.cpp: build llamacpp-rocm/b1369, server arguments -ngl 99 -cram 0 --spec-type ngram-simple -c 33792 -kvu -np 4.
• Loom: speculative decoding on (n-gram prompt lookup, depth 2).
• Same GGUF on both engines, tokens per second. Loom: client-side timing; llama.cpp: server-side timings. Runs per value: decode and prefill 2 (mean), 4 streams 3 (median).
• Measured 2026-11-07; device AMD RYZEN AI MAX+ 395 w/ Radeon 8060S, AMD Radeon 8060S Graphics, Ubuntu 26.04; hrx-loom 0.1.0 (hrx-system d01d1c1); source: local run.
• ¹ Measured 2026-11-07; device AMD RYZEN AI MAX+ 395 w/ Radeon 8060S, AMD Radeon 8060S Graphics, Ubuntu 24.04; hrx-loom 0.1.0 (hrx-system d01d1c1); source: DevLab Dispatch run 42342192972.
• ² Measured 2026-11-08; device AMD RYZEN AI MAX+ 395 w/ Radeon 8060S, AMD Radeon 8060S Graphics, Ubuntu 24.04; hrx-loom 0.1.0 (hrx-system d01d1c1); source: DevLab Dispatch run 42419939771.
• DevLab: the same measurement on an AMD DevLab Dispatch Strix Halo runner; local ÷ DevLab is the local value over the DevLab value of each engine, recorded for comparison, not a verdict.
• Drafted tokens accepted: the share of the drafter's proposed tokens the model accepted during the spec gate's requests.
• Roofline: the hardware ceiling for this GGUF on AMD Ryzen AI Max+ 395 "Strix Halo" (gfx1151) at 216 GB/s and 59 TFLOP/s, from loom-roofline roofline.py (loom-skills a05730e); decode and 4 streams with speculative decoding: verify passes × (1 + acceptance × depth 2) tokens per pass, the drafter's cost not counted; acceptance measured in the row's own session: decode 11 %, decode at 2k 64 %, decode at 32k 82 %, 4 streams 20 %.
• — not measured; n/a not applicable.

Features

feature	status	detail
API	✅	OpenAI-style /v1/chat/completions
tool calling	✅	OpenAI-style tools and tool_calls
reasoning / thinking	❌	not supported
reasoning effort	❌	not supported
speculative decoding	✅	n-gram prompt lookup, depth 2, on by default
image input	❌	not supported
audio input / output	❌	not supported
sampling	✅	temperature, top_p, top_k, seed
context size	✅	32,768 tokens per request
concurrency	✅	4 concurrent requests
max output tokens	✅	4,096 (preset max_tokens)

Correctness

gate	description	result
kernel_tests	Unit tests of the generated kernels against reference implementations	✅ pass (38/38) ¹
golden	llama.cpp reference recorded on the same GGUF: greedy tokens and top-5 log-probs, short prompts and 1k-32k documents	✅ pass ²
compare	Greedy output vs the llama.cpp reference, token by token; every divergence is a near-tie	✅ pass ²
vision	Image input vs llama.cpp with the same projector: identical prompt tokens, replies equal up to near-ties	— not applicable
consistency	Concurrent requests give the same output as the same requests run alone	✅ pass ²
spec	Speculative decoding on gives exactly the output of speculative decoding off	✅ pass ²
agent	Multi-turn tool-calling loop: exact prompt rendering, history reuse, valid tool calls, output vs llama.cpp	✅ pass ²
agent_think_off	The tool-calling loop with thinking off	— not applicable
sampling	temperature, top_k and seed behave as specified; first-token distribution matches llama.cpp	✅ pass ²
needle	Needle-in-a-haystack prompt at up to half the context, compared with llama.cpp; over-capacity prompt refused	✅ pass ³
needle_full	Needle-in-a-haystack prompt at the full context is served	✅ pass ²
decision	Decision answers (choice, score, noul probabilities) vs llama.cpp on the same prompts	— not applicable

• ¹ DevLab Dispatch run 42342192972, 2026-11-07; device AMD RYZEN AI MAX+ 395 w/ Radeon 8060S, AMD Radeon 8060S Graphics, Ubuntu 24.04; hrx-loom 0.1.0 (hrx-system d01d1c1).
• ² DevLab Dispatch run 42419939771, LEVEL=full, 2026-11-08; device AMD RYZEN AI MAX+ 395 w/ Radeon 8060S, AMD Radeon 8060S Graphics, Ubuntu 24.04; hrx-loom 0.1.0 (hrx-system d01d1c1).
• ³ DevLab Dispatch run 42342192972, LEVEL=full, 2026-11-07; device AMD RYZEN AI MAX+ 395 w/ Radeon 8060S, AMD Radeon 8060S Graphics, Ubuntu 24.04; hrx-loom 0.1.0 (hrx-system d01d1c1).
• Gates are the loom-benchmark correctness gates (gates.sh); a gate without a recorded run shows no evidence.

Files

• model.json / model.json.lock: the card and the pinned source files (sha256 checked by loom-fetch).
• program/*.loom, weights.loom: the generated Loom programs.
• tokenizer.json: the tokenizer.
• llm_config.json: the configuration the programs were generated from.
• info.json: the measurements, gate evidence and versions this card shows; everything else comes from the files above.
• REPORT.md: how the values above were measured: setup, commands and the value of every run.

Versions: card schema loom-card/2; loom-skills ca6ae88; hrx-loom 0.1.0 (hrx-system d01d1c1); card generator 1c7ab20.

Source files, fetched by loom-fetch from their own repositories:

• checkpoint: bartowski/Qwen2.5-0.5B-Instruct-GGUF @ 41ba88d, Qwen2.5-0.5B-Instruct-Q4_K_M.gguf

Next steps

• decode_32k is far from its roofline because the decode attention kernel streams the whole KV cache once per verified token, so a depth-2 verify pass costs about as much as decoding its tokens one by one (7 of 10.6 ms per pass at 32k is attention)
• a span kernel that reads each key block once per verification span (loom-skills PR #134, parked) measured decode_32k 250 -> 343 tok/s locally and 211 -> 261 on DevLab, with all gates passing
• it is not adopted: at short context it costs 4-stream decode 14 % on DevLab (the per-epoch leader list and one indirect launch per layer); it needs a device-side skip when no token is past the long-context threshold before this card is regenerated