[AaronStGeorge]

Overview

Llama 3.2 3B decode fails to produce perplexity because Loom cannot resolve @ggml_rope_f32_pair_packet (HRX issue #1089). This PR temporarily works around the linker ordering bug; kernel arithmetic is unchanged. Move rope_f32.loom after llm_attention_qkv_matmul_postops.loom in the vector Q/K recipes, keeping library_sources and compile_dependencies aligned.

CI evidence

Both completed runs cover all 25 smoke models, prefill and decode, on HRX/Vulkan gfx1151. The baseline already includes #119's IQ4 fix, so this pair isolates the ordering workaround.

Run	llama.cpp commit	Prefill passing	Decode passing
Before	682337098905cf69ea4b8ec00bc50a9993bd6ff8	25/25	22/25
After	12aa79f6b822998f2dc48cba97218e916e9f7869	25/25	25/25

The three missing decode measurements (Llama 3.2 1B, Llama 3.2 3B and SmolLM3 3B) all report the unresolved helper before and pass afterward. All previously available HRX PPL values are unchanged. Both workflows are red because expected failures now XPASS; the after run passes all 50 numerical comparisons.

Selected model: Llama-3.2-3B-Instruct-UD-Q4_K_XL.gguf.

Regime	Baseline HRX / reference PPL	Candidate HRX / reference PPL	Numerical verdict	Before / after CI links
decode-like	unavailable / 7.2515	7.2492 / 7.2515	invalid-measurement -> pass	before / after
prefill-like	10.5279 / 10.5638	10.5279 / 10.5638	pass -> pass	before / after

Complete before/after comparison: all 25 smoke models, both regimes

Model	Regime	Baseline HRX / reference PPL	Candidate HRX / reference PPL	Numerical verdict	Change	Failure causes
extra.Bonsai-1.7B-gguf	decode-like	20.2263 / 20.2263	20.2263 / 20.2263	pass -> pass	unchanged pass	none
extra.Bonsai-1.7B-gguf	prefill-like	21.3080 / 21.3049	21.3080 / 21.3049	pass -> pass	unchanged pass	none
extra.Bonsai-4B-gguf	decode-like	14.5323 / 14.5328	14.5323 / 14.5328	pass -> pass	unchanged pass	none
extra.Bonsai-4B-gguf	prefill-like	14.3731 / 14.3636	14.3731 / 14.3636	pass -> pass	unchanged pass	none
extra.Bonsai-8B-gguf	decode-like	13.1626 / 13.1621	13.1626 / 13.1621	pass -> pass	unchanged pass	none
extra.Bonsai-8B-gguf	prefill-like	12.5425 / 12.5495	12.5425 / 12.5495	pass -> pass	unchanged pass	none
extra.Jan-nano-128k-gguf	decode-like	10.4439 / 10.4295	10.4439 / 10.4295	pass -> pass	unchanged pass	none
extra.Jan-nano-128k-gguf	prefill-like	9.7897 / 9.9176	9.7897 / 9.9176	pass -> pass	unchanged pass	none
extra.Jan-v1-4B-GGUF	decode-like	10.8061 / 10.8138	10.8061 / 10.8138	pass -> pass	unchanged pass	none
extra.Jan-v1-4B-GGUF	prefill-like	9.6797 / 9.7696	9.6797 / 9.7696	pass -> pass	unchanged pass	none
extra.LFM2-1.2B-GGUF	decode-like	14.4871 / 14.5189	14.4871 / 14.5189	pass -> pass	unchanged pass	none
extra.LFM2-1.2B-GGUF	prefill-like	13.8624 / 13.8816	13.8624 / 13.8816	pass -> pass	unchanged pass	none
extra.LFM2.5-1.2B-Instruct-GGUF	decode-like	21.0922 / 20.9775	21.0922 / 20.9775	pass -> pass	unchanged pass	none
extra.LFM2.5-1.2B-Instruct-GGUF	prefill-like	18.1015 / 18.1351	18.1015 / 18.1351	pass -> pass	unchanged pass	none
extra.Llama-3.2-1B-Instruct-GGUF	decode-like	unavailable / 10.0121	10.0045 / 10.0121	invalid-measurement -> pass	improved	before hrx: invalid-measurement: llama-perplexity did not print a valid finite positive PPL estimate
extra.Llama-3.2-1B-Instruct-GGUF	prefill-like	14.4433 / 14.4604	14.4433 / 14.4604	pass -> pass	unchanged pass	none
extra.Llama-3.2-3B-Instruct-GGUF	decode-like	unavailable / 7.2515	7.2492 / 7.2515	invalid-measurement -> pass	improved	before hrx: invalid-measurement: llama-perplexity did not print a valid finite positive PPL estimate
extra.Llama-3.2-3B-Instruct-GGUF	prefill-like	10.5279 / 10.5638	10.5279 / 10.5638	pass -> pass	unchanged pass	none
extra.Meta-Llama-3.1-8B-Instruct-GGUF	decode-like	5.4161 / 5.4211	5.4161 / 5.4211	pass -> pass	unchanged pass	none
extra.Meta-Llama-3.1-8B-Instruct-GGUF	prefill-like	7.0713 / 7.0906	7.0713 / 7.0906	pass -> pass	unchanged pass	none
extra.MiniCPM-V-4.6-Thinking-gguf	decode-like	16.0936 / 16.1030	16.0936 / 16.1030	pass -> pass	unchanged pass	none
extra.MiniCPM-V-4.6-Thinking-gguf	prefill-like	14.2873 / 14.3064	14.2873 / 14.3064	pass -> pass	unchanged pass	none
extra.MiniCPM-V-4.6-gguf	decode-like	16.2314 / 16.2231	16.2314 / 16.2231	pass -> pass	unchanged pass	none
extra.MiniCPM-V-4.6-gguf	prefill-like	14.3866 / 14.4046	14.3866 / 14.4046	pass -> pass	unchanged pass	none
extra.MiniCPM3-4B-GGUF	decode-like	21.9671 / 21.9597	21.9671 / 21.9596	pass -> pass	unchanged pass	none
extra.MiniCPM3-4B-GGUF	prefill-like	14.6335 / 14.6525	14.6335 / 14.6581	pass -> pass	unchanged pass	none
extra.MiniCPM5-1B-GGUF	decode-like	23.8496 / 23.7958	23.8496 / 23.7958	pass -> pass	unchanged pass	none
extra.MiniCPM5-1B-GGUF	prefill-like	26.3941 / 26.4078	26.3941 / 26.4095	pass -> pass	unchanged pass	none
extra.Ministral-3-3B-Instruct-2512-GGUF	decode-like	8.1995 / 8.1871	8.1995 / 8.1871	pass -> pass	unchanged pass	none
extra.Ministral-3-3B-Instruct-2512-GGUF	prefill-like	7.8130 / 7.8116	7.8130 / 7.8116	pass -> pass	unchanged pass	none
extra.Phi-4-mini-instruct-GGUF	decode-like	7.5603 / 7.5497	7.5603 / 7.5497	pass -> pass	unchanged pass	none
extra.Phi-4-mini-instruct-GGUF	prefill-like	9.6331 / 9.7933	9.6331 / 9.7933	pass -> pass	unchanged pass	none
extra.PromptBridge-0.6b-Alpha-GGUF	decode-like	24.1013 / 24.0622	24.1013 / 24.0622	pass -> pass	unchanged pass	none
extra.PromptBridge-0.6b-Alpha-GGUF	prefill-like	23.8217 / 23.8592	23.8217 / 23.8592	pass -> pass	unchanged pass	none
extra.Qwen3-0.6B-GGUF	decode-like	19.2814 / 19.2797	19.2814 / 19.2797	pass -> pass	unchanged pass	none
extra.Qwen3-0.6B-GGUF	prefill-like	20.5002 / 20.4865	20.5002 / 20.4865	pass -> pass	unchanged pass	none
extra.Qwen3-1.7B-GGUF	decode-like	17.4745 / 17.5090	17.4745 / 17.5090	pass -> pass	unchanged pass	none
extra.Qwen3-1.7B-GGUF	prefill-like	15.9749 / 15.9850	15.9749 / 15.9850	pass -> pass	unchanged pass	none
extra.Qwen3-4B-GGUF	decode-like	14.0307 / 14.0311	14.0307 / 14.0311	pass -> pass	unchanged pass	none
extra.Qwen3-4B-GGUF	prefill-like	12.2655 / 12.2982	12.2655 / 12.2982	pass -> pass	unchanged pass	none
extra.Qwen3-4B-Instruct-2507-GGUF	decode-like	10.4159 / 10.4413	10.4159 / 10.4413	pass -> pass	unchanged pass	none
extra.Qwen3-4B-Instruct-2507-GGUF	prefill-like	9.0959 / 9.1689	9.0959 / 9.1689	pass -> pass	unchanged pass	none
extra.Qwen3.5-0.8B-GGUF	decode-like	16.5618 / 16.5637	16.5618 / 16.5637	pass -> pass	unchanged pass	none
extra.Qwen3.5-0.8B-GGUF	prefill-like	15.7516 / 15.7933	15.7516 / 15.7933	pass -> pass	unchanged pass	none
extra.Qwen3.5-2B-GGUF	decode-like	12.1638 / 12.1727	12.1638 / 12.1727	pass -> pass	unchanged pass	none
extra.Qwen3.5-2B-GGUF	prefill-like	10.8440 / 10.8548	10.8440 / 10.8548	pass -> pass	unchanged pass	none
extra.SmolLM3-3B-128K-GGUF	decode-like	unavailable / 7.4642	7.4737 / 7.4642	invalid-measurement -> pass	improved	before hrx: invalid-measurement: llama-perplexity did not print a valid finite positive PPL estimate
extra.SmolLM3-3B-128K-GGUF	prefill-like	8.9937 / 9.0182	8.9937 / 9.0182	pass -> pass	unchanged pass	none
extra.gemma-3-270m-it-GGUF	decode-like	53.9771 / 53.9679	53.9771 / 53.9679	pass -> pass	unchanged pass	none
extra.gemma-3-270m-it-GGUF	prefill-like	60.2556 / 60.2415	60.2556 / 60.2579	pass -> pass	unchanged pass	none

Smoke comparison: before / after.

Both automation trees differ only in the llama.cpp gitlink and use HRX c9ffcd25e74ff3580cb128a485c74f0c3d48fec4 with the same unfiltered model manifest. The baseline source tree is identical to integration merge 4e14145b37c11f5e9394e02868e81d6edae7cce3. Build checkouts and raw PPL logs were verified.

Settings: model_tier=smoke, skip_benchmarks=["lemonade-benchmark"]; WikiText-2, context/batch 512; prefill microbatch 512 with 32 chunks; decode microbatch 1 with two chunks; unchanged HRX/Vulkan PPL ratio limit 1.10. Different gfx1151 runners were used. Both exact Q/K recipes also reproduce the before/after link result locally on gfx1100.

Stability: ten additional Llama 1B/8B runs passed all 40 comparisons across nine runners with identical PPL. The earlier high-PPL/NaN results were likely runner/environment related; the original runner was not retested.

Requirements

• Contributing guidelines reviewed for this change; the contributor requested this submission after reviewing the diagnosis and ordering workaround.
• AI usage disclosure: YES. Codex assisted with diagnosis, the manifest edit, validation, and this PR.

[AaronStGeorge]

Merged as part of #121