Annier.fp4 refactor

[Annieren]

Overview

It dissolves the https://github.com/charlie12345/rocmfp4 into ggml's canonical file homes for the two rocmfp4 quant types on gfx1151 (RDNA3.5), so rocmfp4 becomes structurally indistinguishable from nvfp4/mxfp4.

Additional information

First Author: Charlie12345

Requirements

• I have read and agree with the contributing guidelines
• AI usage disclosure: Yes, some study assistance and document.

[mgehre-amd]

Can you please explain the validation that you have done in the PR description?

[Annieren]

Got catch. mmq-config-rdna3-5.cuh was changed accidentally. Added it back.

1. I have tested ppl for multiple rocmfp4 models test:

model	gpu (PPL ± err)	cpu (PPL ± err)
gpt-oss-20b-MXFP4.gguf	209.1671 ± 6.69884	206.9469 ± 6.62915
Qwen3.6-35B-A3B-Q4_0_ROCMFP4_STRIX_LEAN.gguf	6.0749 ± 0.14655	6.0730 ± 0.14638
ornith-1.0-35b-MTP-ROCmFP4-COHERENT.gguf	6.1894 ± 0.14964	6.1722 ± 0.14919
Qwen3.6-40B-Deckard-MTP-ROCmFP4-STRIX-embQ8-imatrix-headQ6.gguf	6.0991 ± 0.14321	FAIL(too big to load)
Qwen3-Coder-Next-ROCmFP4-STRIX-embQ8-imatrix-headQ6.gguf	6.3562 ± 0.18145	FAIL(too big to load)
Qwen3.6-35B-A3B-MTP-ROCmFP4-STRIX-embF16-headQ6.gguf	6.0612 ± 0.14725	6.0690 ± 0.14756
FastContext-1.0-4B-SFT-ROCmFP4-STRIX-embF16-imatrix.gguf	9.0616 ± 0.25846	9.0622 ± 0.25856
Nex-N2-mini-ROCmFP4-STRIX-embF16-imatrix-headQ6.gguf	6.0234 ± 0.14591	6.0219 ± 0.14597

2. Performance test on both rcomfp4 and non-rocmfp4 models, no performance impact as expected:
   fp4refactor-vs baseline

model	test	A	B	Δ (B/A)
Cosmos-Reason2-8B-Q8_0	pp512	1500.85	1474.00	-1.8%
Cosmos-Reason2-8B-Q8_0	tg128	26.15	26.11	-0.2%
DeepSeek-R1-Distill-Qwen-7B-Q4_K_M	pp512	1633.05	1621.45	-0.7%
DeepSeek-R1-Distill-Qwen-7B-Q4_K_M	tg128	46.64	45.82	-1.8%
Meta-Llama-3.1-8B-Instruct-Q4_K_M	pp512	1376.46	1367.60	-0.6%
Meta-Llama-3.1-8B-Instruct-Q4_K_M	tg128	43.29	42.56	-1.7%
Qwen3-1.7B-Q4_K_M	pp512	5399.19	5391.90	-0.1%
Qwen3-1.7B-Q4_K_M	tg128	152.71	152.51	-0.1%
Qwen3-30B-A3B-Q4_K_M	pp512	1768.02	1773.46	+0.3%
Qwen3-30B-A3B-Q4_K_M	tg128	79.30	79.44	+0.2%
Qwen3-4B-Q4_K_M	pp512	2552.13	2545.35	-0.3%
Qwen3-4B-Q4_K_M	tg128	75.44	75.39	-0.1%
Qwen3.5-35B-A3B-Q4_K_M	pp512	1738.59	1734.58	-0.2%
Qwen3.5-35B-A3B-Q4_K_M	tg128	59.19	58.72	-0.8%
Qwen3.5-4B-Q4_0	pp512	2325.64	2306.02	-0.8%
Qwen3.5-4B-Q4_0	tg128	66.67	66.46	-0.3%
Qwen3.5-4B-Q4_K_M	pp512	2179.13	2267.82	+4.1%
Qwen3.5-4B-Q4_K_M	tg128	63.18	62.33	-1.3%
Qwen3.5-9B-Q4_0	pp512	1380.28	1381.55	+0.1%
Qwen3.5-9B-Q4_0	tg128	40.09	40.15	+0.1%
Qwen3.5-9B-Q4_K_M	pp512	1241.35	1266.67	+2.0%
Qwen3.5-9B-Q4_K_M	tg128	35.24	37.52	+6.5%
Qwen3.6-35B-A3B-UD-Q4_K_XL	pp512	1732.68	1707.55	-1.5%
Qwen3.6-35B-A3B-UD-Q4_K_XL	tg128	57.62	58.26	+1.1%
Qwen3.6-35B-A3B-UD-Q5_K_XL	pp512	1578.18	1612.38	+2.2%
Qwen3.6-35B-A3B-UD-Q5_K_XL	tg128	55.82	56.19	+0.7%
SmolLM2-1.7B-Instruct-Q4_K_M	pp512	4421.65	4374.87	-1.1%
SmolLM2-1.7B-Instruct-Q4_K_M	tg128	148.78	156.17	+5.0%
gemma-4-12b-it-Q4_K_M	pp512	916.41	899.97	-1.8%
gemma-4-12b-it-Q4_K_M	tg128	27.93	27.76	-0.6%
gemma-4-26B-A4B-it-UD-Q4_K_M	pp512	1703.79	1736.18	+1.9%
gemma-4-26B-A4B-it-UD-Q4_K_M	tg128	49.61	47.93	-3.4%
gemma-4-31B-it-Q4_K_M	pp512	349.48	355.00	+1.6%
gemma-4-31B-it-Q4_K_M	tg128	11.21	11.27	+0.5%
mean Δ = -0.2%

3. ISA diff, no change as expected