Independent Linux and Windows testing will determine whether llama.cpp’s ROCm 7.14 targets make AMD’s TheRock-based production stack reliable for local inference.

state: resolvedheat: lowuncertainty: mediumconvergesscott: mediumllama-cpp rocm local-inferencellama.cpp AMDsuperm1

What is this?

Contributor superm1 opened llama.cpp pull request #25775 to add CI targets for AMD ROCm 7.14, described in the supplied artifact as the first production release built using TheRock. AMD’s release material says llama.cpp is available across Ryzen Linux and Windows and emphasizes TheRock’s public PRs and CI pipelines, while AMD’s installation docs provide unit-test validation steps. However, the supplied snippets contain no independent Linux/Windows test results for these new targets, so cross-platform reliability and performance are not yet established here; some third-party material instead describes AMD support as less consistent on Windows or on unsupported consumer GPUs.

Why it matters to Scott

llama.cpp adding Linux and Windows CI for ROCm 7.14 converges with Scott’s position that stack portability and production capability must be demonstrated through representative, repeatable tests rather than vendor claims. It bears on his hardware-aware local-inference work and CUDA-based gamepc stack as a potential AMD alternative, but without independent results it does not yet justify a stack change; the radar already tracks adjacent llama.cpp/ROCm reliability and performance questions.

ip:framework.sovereign-software-assuranceip:concept.capability-auditdev:concept.hardware-aware-local-inferencedev:project.gamepcradar:concept.rocmradar:concept.llama-cppradar:person.amdradar:person.llama-cppradar:llama-cpp-rocm-q2k-speedups

queries asked of Scott's wikis

AMD vs CUDA local inference strategy
cross-platform GPU CI and hardware test matrices
ROCm llama.cpp build and deployment experience
local inference hardware sovereignty
Linux versus Windows inference reliability
TheRock open-source accelerator stack

Measured heat

no measured readings yet — the hourly heat pass fills this in

How the heat travelled

no chain yet — the hourly chain pass fills this in

Evidence (10) — ⭐ canonical anchor

source	object	author	score	comments
🟠 reddit	Add CI targets for ROCm 7.14 by superm1 · Pull Request #25775 · ggml-org/llama.cpp LocalLLaMA	pmttyji	16	2
🟧 echo.github ⭐	The original artifact is the llama.cpp pull request itself. Its description states: “ROCm 7.14 is the first production release using TheRock	Mario Limonciello (superm1)	—	—
🟠 reddit	Running Qwen 3.6 35B A3B-Q8_0 gguf on a cheap radeon 7600 at 18 token/s * update increased to 21 t/s LocalLLaMA	Sweaty_Perception655	17	15
🟠 reddit	Llama.cpp ROCm 7.2->7.14 upgrade, Radeon 780m iGPU benchmarks: ROCm vs Vulkan LocalLLaMA	MaximusSenior	22	20
🟧 hn	Native vLLM and ROCm 7.15 for RX 6000 (RDNA2) on Windows 11 – 26 Tflops FP16	esseba-dev	1	0
🟠 reddit	Did AMD just fucking fix ROCm and nobody told me? LocalLLaMA	smellof	12	55
🟠 reddit	I noticed llama.cpp started putting out 7.14rocm nightlies again after going dark fixing a bug for a while. LocalLLaMA	W61k3r	2	0
🟠 reddit	Re-done benchmarks for V620 on Windows/ROCm & Vulkan LocalLLaMA	Brave_Load7620	7	14
🟠 reddit	Llama.cpp with ROCm 7.14 on Radeon 780m - fast, but unstable. Workaround LocalLLaMA	MaximusSenior	6	23
🟠 reddit	ROCm 10.0: A Decade of Open Compute, Built for the Age of Agentic AI LocalLLaMA	pmttyji	293	116

Interpretation history

2026-08-28T19:36:42Z
ROCm 10 and its pending llama.cpp integration have overtaken the version-specific window for validating ROCm 7.14. The accumulated evidence leaves 7.14 Linux-viable but hardware-sensitive and not established as a reliable cross-platform production stack; further testing belongs in a successor case.
2026-08-28T19:23:52Z
evidence attached: reddit.post.1w0yfmn — The ROCm 10 release and pending llama.cpp integration materially contextualize the open case about AMD's evolving local-inference software stack.
2026-08-27T19:51:08Z
The refreshed comments add adjacent gfx1103 user interest and troubleshooting suggestions, but no independent reproduction, confirmed root cause, or new Windows ROCm 7.14 result. The hardware-specific instability remains credible enough to constrain production-readiness claims, yet the case is otherwise unchanged and cold.
2026-08-27T16:34:34Z
A technically specific comment links the 780M crashes to an open gfx1103 MES issue and clarifies that full kernel serialization is a performance-costly throttle rather than a fix. This sharpens the suspected failure mode but remains unverified and adds neither independent reproduction nor representative Windows ROCm 7.14 validation.
2026-08-27T14:41:49Z
The case now contains a direct ROCm 7.14 reliability failure rather than merely an absence of validation: one Radeon 780M deployment crashes frequently unless a costly serialization workaround is enabled. This strengthens evidence that Linux performance is real but production readiness remains hardware-sensitive and unproven, while representative Windows 7.14 testing is still missing.
2026-08-27T14:25:25Z
evidence attached: reddit.post.1vzud1q — Independent local deployment confirms promising ROCm 7.14 preprocessing performance but exposes instability and a substantial workaround cost.
2026-08-27T00:30:10Z
The new same-silicon Linux benchmark uses older ROCm 6.4.3, while the other comments only suggest controls and tuning. It adds context but no representative Windows or cross-platform ROCm 7.14 validation, so the production-reliability claim remains open and cold.
2026-08-25T23:41:12Z
The refreshed V620 comments add no completed benchmark or reproducible Windows result beyond the existing preliminary ROCm 7.15 data point. Linux viability remains corroborated, but representative ROCm 7.14 cross-platform reliability is still unresolved and cold.
2026-08-25T18:39:44Z
The refreshed V620 discussion proposes useful ROCm-versus-Vulkan and flash-attention controls but supplies no new run or result. Windows viability remains a preliminary ROCm 7.15 data point, leaving representative ROCm 7.14 cross-platform reliability unresolved.
2026-08-25T16:47:03Z
The V620 benchmark supplies the case’s first independent Windows llama.cpp data point, showing that a newer TheRock ROCm nightly can run alongside Vulkan on Windows 11. Because it uses ROCm 7.15 rather than the production 7.14 target, covers one specialized GPU, and remains preliminary, it narrows the platform gap without establishing cross-platform production reliability.
2026-08-25T16:25:00Z
evidence attached: reddit.post.1vy3t5h — Independent V620 benchmarks using a newer TheRock ROCm nightly provide relevant evidence about AMD local-inference support, though they remain preliminary.
2026-08-24T18:26:03Z
Resumed ROCm 7.14 nightlies show that llama.cpp’s bug remediation and binary publication are progressing, but they do not add independent reliability evidence. Linux viability remains corroborated; reproducible Windows testing and cross-platform production readiness remain unresolved.
2026-08-24T18:23:05Z
evidence attached: reddit.post.1vx9r6j — The resumed ROCm 7.14 nightly release provides first-party evidence that llama.cpp remediation is progressing, though validation remains outstanding.
2026-08-23T04:27:34Z
The refresh is repetitive amplification of existing Linux usability anecdotes and adds no reproducible TheRock benchmark, packaging validation, or Windows llama.cpp result. Linux viability remains corroborated, while cross-platform production reliability is still unresolved and cold.
2026-08-21T03:28:06Z
The refreshed comments remain repetitive amplification of general ROCm usability and Vulkan alternatives, adding no reproducible TheRock test or Windows llama.cpp result. Linux viability is corroborated, but cross-platform production reliability remains unresolved and cold.
2026-08-21T00:25:37Z
The refreshed comments add general claims that ROCm works on modern hardware and that Vulkan remains a practical alternative, but no reproducible TheRock result or Windows llama.cpp validation. Linux usability is still corroborated while cross-platform production reliability remains unresolved and cold.
2026-08-20T20:38:17Z
A first-hand Fedora/RDNA2 report extends the Linux evidence from raw compatibility to substantially easier installation through TheRock’s multi-architecture packages, including on older consumer hardware. The disputed version labeling, lack of reproducible testing, and continued absence of Windows llama.cpp validation leave cross-platform production reliability unresolved.
2026-08-20T18:23:28Z
evidence attached: reddit.post.1vtqrv1 — First-hand testing provides useful independent corroboration that TheRock's multi-architecture ROCm packaging may materially improve AMD local-inference usability.
2026-08-20T11:29:34Z
The staleness check adds only negligible engagement and no new testing. Linux viability remains corroborated, but reproducible Windows llama.cpp validation and cross-platform production-reliability evidence are still absent.
2026-08-18T10:37:34Z
The Windows/RDNA2 vLLM claim is adjacent evidence that native ROCm on Windows may be becoming practical, but it uses ROCm 7.15 and a different inference engine and supplies no independent validation. It therefore does not close the case’s missing Windows llama.cpp 7.14 reliability evidence or elevate Linux viability into cross-platform production readiness.
2026-08-18T10:22:55Z
evidence attached: hn.story.49343550 — A first-party implementation claim about vLLM and ROCm on Windows/RDNA2 materially extends the open case on practical AMD local-inference support.
2026-08-17T11:30:39Z
The refreshed comments add no reproducible benchmark, packaging improvement, or Windows result; they only repeat interest and isolated reliability concerns. Linux viability remains corroborated, while cross-platform production reliability stays unresolved and cold.
2026-08-17T06:28:36Z
The refreshed discussion adds no reproducible result, packaging improvement, or Windows test, so it does not extend the existing Linux viability evidence. Cross-platform production reliability remains unproven and the case stays cold.
2026-08-17T02:30:13Z
The comment refresh adds no reproducible benchmark, packaging result, or Windows test. Linux viability remains corroborated, while cross-platform production reliability is unresolved and cold.
2026-08-17T01:24:09Z
The refreshed discussion adds no independent benchmark, packaging result, or Windows test. Linux viability remains corroborated, but cross-platform production reliability is still unproven and the case remains cold.
2026-08-16T23:26:19Z
The refreshed comments add only acknowledgment and version-number confusion, with no new benchmark, implementation detail, or Windows validation. Linux viability remains corroborated, while cross-platform production reliability stays unresolved and cold.
2026-08-16T22:29:00Z
A second independent Linux data point now corroborates that ROCm 7.14 works with llama.cpp across both integrated and discrete consumer Radeon hardware, but the 780M result also exposes source-build packaging gaps and mixed performance versus Vulkan. This validates Linux viability rather than production reliability; repeatability across hardware and any independent Windows evidence remain missing.
2026-08-16T22:22:49Z
evidence attached: reddit.post.1vqa7qh — Independent Radeon 780M benchmarks provide concrete evidence about ROCm 7.14 support and performance versus Vulkan.
2026-08-15T00:23:21Z
No new validation has appeared beyond the same anecdotal Linux result; the case still lacks reproducible ROCm 7.14 testing and any independent Windows evidence. The portability and production-reliability hypothesis remains open but cold.
2026-08-12T23:32:32Z
The refreshed discussion adds another low-end AMD performance anecdote, but it does not identify ROCm 7.14 or provide a controlled comparison, so it broadens Linux interest without independently validating the new stack. Windows reliability and repeatable cross-platform results remain absent.
2026-08-12T21:34:17Z
The case now has preliminary Linux field evidence that ROCm 7.14 can provide usable llama.cpp performance on low-end Radeon hardware, moving it beyond upstream enablement alone. The reports remain uncontrolled and provide no Windows validation, so they do not establish cross-platform reliability or production readiness.
2026-08-12T21:22:56Z
evidence attached: reddit.post.1vmqpnm — A low-cost Radeon report provides weak field evidence that ROCm 7.14 and llama.cpp can deliver usable Qwen inference performance, though the result is anecdotal.
2026-08-11T07:47:35Z
No independent Linux or Windows validation has appeared; the small engagement increase is merely amplification and does not change the portability-enablement story into evidence of reliability.
2026-08-11T07:28:43Z
grounded: converges/medium — llama.cpp adding Linux and Windows CI for ROCm 7.14 converges with Scott’s position that stack portability and production capability must be demonstrated throug
2026-08-11T07:25:46Z
origin walked (codex/luna, conf 0.98): anchor reddit.post.1vl9t2k -> echo.github.8d912a0336 by Mario Limonciello (superm1)
2026-08-11T07:24:53Z
case created — The upstream integration is a concrete, testable change affecting AMD local inference on two operating systems.

Decision trace

08-29 05:36resolveROCm 10 and its pending llama.cpp integration have overtaken the version-specific window for validating ROCm 7.14. The accumulated evidence leaves 7.14 Linux-viable but hardware-sensitive and not esta
08-29 05:36alert_silentThe ROCm 10 release is established but adds no demonstrated llama.cpp compatibility, stability, or performance improvement, and the integration remains pending. It can wait for normal briefing or conc
08-29 05:36alert_routeThe ROCm 10 release is established but adds no demonstrated llama.cpp compatibility, stability, or performance improvement, and the integration remains pending. It can wait for normal briefing or conc
08-29 05:24alert_silentAMD’s first-party ROCm 10.0 release is a real event, and a llama.cpp integration PR exists, but the supplied evidence establishes neither a material local-inference improvement nor completed llama.cpp
08-29 05:24surface_candidateAMD’s first-party ROCm 10.0 release is a real event, and a llama.cpp integration PR exists, but the supplied evidence establishes neither a material local-inference improvement nor completed llama.cpp
08-29 05:24alert_routeAMD’s first-party ROCm 10.0 release is a real event, and a llama.cpp integration PR exists, but the supplied evidence establishes neither a material local-inference improvement nor completed llama.cpp
08-29 05:23attachThe ROCm 10 release and pending llama.cpp integration materially contextualize the open case about AMD's evolving local-inference software stack.
08-29 05:22propose_attachThe ROCm 10 release and pending llama.cpp integration materially contextualize the open case about AMD's evolving local-inference software stack.
08-28 05:51repriceThe refreshed comments add adjacent gfx1103 user interest and troubleshooting suggestions, but no independent reproduction, confirmed root cause, or new Windows ROCm 7.14 result. The hardware-specific
08-28 05:51alert_silentThe new discussion does not materially extend the known 780M failure report or validate its suspected MES cause, so it can wait for the normal briefing.
08-28 05:51alert_routeThe new discussion does not materially extend the known 780M failure report or validate its suspected MES cause, so it can wait for the normal briefing.
08-28 03:22sensor_dirtycomment_update
08-28 02:34repriceA technically specific comment links the 780M crashes to an open gfx1103 MES issue and clarifies that full kernel serialization is a performance-costly throttle rather than a fix. This sharpens the su
08-28 02:34alert_silentThe refreshed diagnosis makes the existing instability report more actionable, but it is still one commenter’s unverified attribution rather than a reproduced result or confirmed upstream finding, so
08-28 02:34alert_routeThe refreshed diagnosis makes the existing instability report more actionable, but it is still one commenter’s unverified attribution rather than a reproduced result or confirmed upstream finding, so
08-28 02:22sensor_dirtycomment_update
08-28 00:41repriceThe case now contains a direct ROCm 7.14 reliability failure rather than merely an absence of validation: one Radeon 780M deployment crashes frequently unless a costly serialization workaround is enab
08-28 00:41alert_silentThe instability report bears directly on the case, but it remains a single hardware-specific anecdote without reproduction or a confirmed root cause; it changes the routine interpretation but does not
08-28 00:41alert_routeThe instability report bears directly on the case, but it remains a single hardware-specific anecdote without reproduction or a confirmed root cause; it changes the routine interpretation but does not
08-28 00:25alert_silentA single Radeon 780M user reports frequent ROCm 7.14 crashes and a stabilizing workaround, AMD_SERIALIZE_KERNEL=3, that substantially reduces the reported prefill advantage over Vulkan. This is a usef
08-28 00:25surface_candidateA single Radeon 780M user reports frequent ROCm 7.14 crashes and a stabilizing workaround, AMD_SERIALIZE_KERNEL=3, that substantially reduces the reported prefill advantage over Vulkan. This is a usef
08-28 00:25alert_routeA single Radeon 780M user reports frequent ROCm 7.14 crashes and a stabilizing workaround, AMD_SERIALIZE_KERNEL=3, that substantially reduces the reported prefill advantage over Vulkan. This is a usef
08-28 00:25attachIndependent local deployment confirms promising ROCm 7.14 preprocessing performance but exposes instability and a substantial workaround cost.
08-28 00:22propose_attachIndependent local deployment confirms promising ROCm 7.14 preprocessing performance but exposes instability and a substantial workaround cost.
08-27 10:30repriceThe new same-silicon Linux benchmark uses older ROCm 6.4.3, while the other comments only suggest controls and tuning. It adds context but no representative Windows or cross-platform ROCm 7.14 validat
08-27 10:30alert_silentNo new completed ROCm 7.14 test, packaging result, or representative Windows validation appeared; adjacent older-stack results and benchmarking advice can wait for routine review.
08-27 10:30alert_routeNo new completed ROCm 7.14 test, packaging result, or representative Windows validation appeared; adjacent older-stack results and benchmarking advice can wait for routine review.
08-27 10:29drop_targetsquiet through full ladder or over cap 8
08-26 09:41repriceThe refreshed V620 comments add no completed benchmark or reproducible Windows result beyond the existing preliminary ROCm 7.15 data point. Linux viability remains corroborated, but representative ROC
08-26 09:41alert_silentThe discussion refresh contains testing suggestions or removed content rather than new results, so it does not alter the case or merit attention before the next briefing.
08-26 09:41alert_routeThe discussion refresh contains testing suggestions or removed content rather than new results, so it does not alter the case or merit attention before the next briefing.
08-26 06:21sensor_dirtycomment_update
08-26 04:39repriceThe refreshed V620 discussion proposes useful ROCm-versus-Vulkan and flash-attention controls but supplies no new run or result. Windows viability remains a preliminary ROCm 7.15 data point, leaving r
08-26 04:39alert_silentThe new comment is a request for stronger testing rather than evidence; it does not change the case or warrant attention before a routine briefing.
08-26 04:39alert_routeThe new comment is a request for stronger testing rather than evidence; it does not change the case or warrant attention before a routine briefing.
08-26 04:22sensor_dirtycomment_update
08-26 02:47repriceThe V620 benchmark supplies the case’s first independent Windows llama.cpp data point, showing that a newer TheRock ROCm nightly can run alongside Vulkan on Windows 11. Because it uses ROCm 7.15 rathe
08-26 02:47alert_silentThis is useful incremental Windows evidence, but it does not validate the case’s ROCm 7.14 target or provide representative, repeatable results that would affect Scott’s local-inference stack decision
08-26 02:47alert_routeThis is useful incremental Windows evidence, but it does not validate the case’s ROCm 7.14 target or provide representative, repeatable results that would affect Scott’s local-inference stack decision
08-26 02:25alert_silentA single self-reported Windows 11 benchmark on a V620 adds an incremental portability data point, but it uses an unofficial ROCm 7.15 TheRock nightly rather than the production 7.14 target, remains un