AMD Radeon RX 7800 XT 16GB

Reported on GitHub by HarlonOna — posted 2026-08-09 Original report

Not reproduced by Bitworks

RX 7800 XT Vulkan Qwen3.5-9B field report

Method (as reported): llama-bench -pg 512,128 -r 20; separate server chat with MTP2 and one slot.

• Reported GPU setup RX 7800 XT 16GB
• GPU count 1
• Operating system Linux (Fedora 44, Mesa RADV)
• Engine llama.cpp
• Backend Vulkan
• Model Qwen3.5-9B Q4_K_M
• Configured context 102,400 tokens
• Actual input input length not reported
• Reported context Reported sessions occupied about 20000 to 30000 tokens, not the configured full window.
• Simultaneous clients 1
• Server slots 1
• Workers 1
• Topology single
• Speculative decoding MTP2 on server; off in llama-bench
• CPU Ryzen 5 7500F

• 71.95 tok/s ± 0.03 (as reported, n = 20) — generation speed (per request) · Synthetic benchmark (not API serving) · 1 client
  llama-bench tg128; MTP off; author reports ±0.03
  This row: configured context: not reported for this row · speculative decoding: off
• 1700.48 tok/s ± 4.65 (as reported, n = 20) — prompt speed (per request) · Synthetic benchmark (not API serving) · 1 client · input 512 tokens
  llama-bench pp512; MTP off; author reports ±4.65
  This row: configured context: not reported for this row · speculative decoding: off
• ≈86 tok/s — generation speed (per request) · agent session · 1 client
  Approximate average across ten chat questions with MTP2
  This row: speculative decoding: MTP2

Not reported: system memory, Exact llama.cpp revision, GGUF digest, Driver revision, Physical GPU residency, Raw server timings

• Configured context was not filled in the reported sessions.
• Synthetic no-MTP and chat with MTP are different workloads.
• Author describes a concurrent slowdown but not a controlled aggregate throughput rate.
• Linux RADV does not qualify native Windows Vulkan.

Setup guide (external)

Reported on GitHub by Niko1221 / Strata maintainers Original report

Not reproduced by Bitworks Project documentation — capability, not a measurement

Strata AMD HIP guide

Method (as reported): Maintainer documents Windows bundled HIP runtime, per-card checks and Linux layer split; no discrete AMD Windows model-load validation.

HEAD seen at retrieval: 6f32ec070f23 (not necessarily the commit the author measured)

• Reported GPU setup Windows: one GPU per model; Linux: selectable layer split
• GPU count not reported
• Operating system Windows and Linux (Windows HIP and Linux HIP)
• Engine Strata HIP
• Backend HIP
• Model Qwen3.8-Flash-Next pack
• Actual input input length not reported
• Topology not reported — Windows one GPU per model; Linux selectable layer split

Not reported: model quantisation, configured context, actual input length, simultaneous clients, topology, CPU, system memory, release archive digest and revision, Windows discrete-card load, Windows dual-card execution, quality and memory curves

• Maintainer says ready-made Windows archive has not run a model on a discrete AMD card in documented validation.
• Capability documentation, not a measured compatibility or performance result.
• RX 9060 XT memory variant is unspecified in the guide, so neither variant is mapped.

Setup guide (external)

Reported on Reddit by Specific-Pomelo-5455 Original report

Not reproduced by Bitworks

Corrected RX 7800 XT Qwen3.8 occupancy benchmarks

Method (as reported): Structured server harness; parallel one, FA, q4_0 K/V; four to six repeats per point.

• Reported GPU setup RX 7800 XT 16GB
• GPU count 1
• Operating system Windows (Windows, version not reported)
• Engine llama.cpp
• Backend HIP/ROCm
• Model Qwen3.8-27B Unsloth-Dynamic-derived IQ4_XS mixed tensors
• Configured context 65,536 tokens
• Actual input input length not reported
• Reported context One row occupied about 7000 tokens; the other occupied about 48000.
• Simultaneous clients 1
• Server slots 1
• Workers 1
• Topology single
• Speculative decoding MTP off for context-depth comparison

• 21.85 tok/s — generation speed (per request) · API server · 1 client
  About 7000 occupied tokens at configured 65536; q4_0 K/V
  This row: speculative decoding: off
• 8.77 tok/s — generation speed (per request) · API server · 1 client
  About 48000 occupied tokens at configured 65536; q4_0 K/V
  This row: speculative decoding: off
• 22.15 tok/s — generation speed (per request) · API server · 1 client
  Compared with 18.62 tok/s (MTP on versus off for one long generation); reported difference: speculative decoding
  Separate reported comparison at configured 32K; not the context-depth pair
  This row: configured context: 32,768 tokens · speculative decoding: MTP on

Not reported: Exact GGUF digest, llama.cpp build, Driver revision, CPU and RAM, Physical WDDM residency

• The configured window is not the occupied input.
• Earlier author measurements were corrected; use these corrected points only.
• The default fit setting had silently moved layers to CPU in some runs.
• MTP worsened the author's multi-step agent wall time; the long-generation comparison is not an agent guarantee.

Setup guide (external)