Qwen3.8-27B Q8_0 on two Radeon RX 7900 XTX 24GB cards (Linux)

Reported on Reddit by nasone32 Original report

Not reproduced by Bitworks

Dual RX 7900 XTX tensor parallel on limited PCIe

Method (as reported): Author used rocprofv3 on a saturated chipset link and combined transfer compression with a larger patch set.

HEAD seen at retrieval: 15995a12d1d5 (not necessarily the commit the author measured)

• Reported GPU setup RX 7900 XTX
• GPU count 2
• Operating system Linux (Ubuntu 24; ROCm 7.14)
• Engine custom llama.cpp HIP fork
• Backend ROCm
• Model Qwen3.8-27B Q8_0
• Actual input input length not reported
• Reported context Prompt row has known input length; decode range may use another condition.
• Workers 1
• Topology dual-split (tensor split; one chipset-connected PCIe x4 card)
• Speculative decoding adaptive MTP in decode range
• PCIe one chipset-connected PCIe x4 card

• ≈1400 tok/s — prompt speed (per request) · input 8,192 tokens
  Compared with 1100 tok/s (baseline patched AllReduce; reported Q8 transfer); changed between the two runs: engine build, other
  Approximate author comparison on the same asymmetric host.
  This row: topology: dual-split
• ≈60–65 tok/s — generation speed (per request)
  Combined RDNA Boost and adaptive MTP patch set; not isolated transfer-only data.
  This row: speculative decoding: adaptive MTP · topology: dual-split

Not reported: configured context, simultaneous clients, CPU, system memory, exact tested fork commit, artifact digest, replications, quality delta, full launch arguments

• Custom patch set differs from the linked recipe and is not a stock Windows path.
• Same author and host as the topology report; not an independent replication.
• Author had not evaluated quality impact of Q8 inter-card transfer.
• Fork author says combined build will not be maintained.

Setup guide (external)

Related link (external)

Reported on Reddit by nasone32 Original report

Not reproduced by Bitworks

Notes on dual RX 7900 XTX tensor split topology

Method (as reported): Author compared layer and patched tensor split on an asymmetric dual-card host; exact test counts not supplied.

• Reported GPU setup RX 7900 XTX
• GPU count 2
• Operating system Linux (ROCm Linux; Windows failure also described)
• Engine custom llama.cpp HIP
• Backend ROCm
• Model Qwen3.8-27B Q8_0
• Actual input input length not reported
• Workers 1
• Topology dual-split (PCIe 4.0 x16 CPU plus x4 chipset)
• Speculative decoding MTP3 in generation comparison
• PCIe PCIe 4.0 x16 CPU plus x4 chipset

• ≈1150 tok/s — prompt speed (per request)
  Compared with 1500 tok/s (baseline layer split; reported patched tensor split); changed between the two runs: split mode, engine build
  Approximate author-reported comparison on asymmetric PCIe topology.
  This row: topology: dual-split
• ≈50 tok/s — generation speed (per request)
  Compared with 30 tok/s (baseline layer split; reported patched tensor/MTP3); changed between the two runs: split mode, engine build, speculative decoding
  Approximate prose comparison; split and speculation both changed.
  This row: speculative decoding: MTP3 · topology: dual-split

Not reported: configured context, actual input length, simultaneous clients, CPU, system memory, exact build and patch digest, prompt and output counts, repetition and variance, Windows version and driver

• Custom internal AllReduce runtime differs from the linked recipe.
• The generation comparison changes both split and MTP settings.
• Author-reported Windows P2P failure does not establish a native-Windows tensor recipe.
• Predicted symmetric-slot gains are not measurements.