[ppkrol]

What & why

edge0-145b is documented as NVIDIA-only. edge0-strata still carries upstream Strata's HIP backend, but the
latent-MoE additions did not build or run with it. With the changes below Edge0-145B-q3k loads, reads prompts and
decodes on a Radeon RX 7900 XTX (gfx1100, 24 GB) under Windows 11.

Build (HIP only):

• CMakeLists.txt: the HIP MMQ branch now compiles lat_q3k_*.cu and moe_fused_q3k.cu (their entry points are
  already stubs under __HIPCC__), and strata_engine gets latent_dense.cpp and latent_moe.cpp on HIP too; they
  were only added inside if(STRATA_ENABLE_CUDA).
• moe_fused.cu, moe_fused_q3k.cu: the latent_kernel / latent_kernel_q3k launches were outside the
  !__HIPCC__ guard that holds the kernels. They are not reached on HIP (lat_dev().ok is false).
• verify.cpp, prefill.cpp: two debug dumps use cudaGraph* / cudaPointer* names the HIP shim does not map.

Run (HIP only):

• verify.cpp: the D stage's two D2H cudaMemcpyAsync calls and the "verify top3" dump (a synchronous
  cudaMemcpy) ran inside the verify window's graph capture. HIP then fails every request with
  verify: end capture: operation failed due to a previous error during capture. On HIP the two copies go through
  copy_bytes_from_mapped and the dump is skipped.

Every code change is inside #if defined(STRATA_USE_HIP) / __HIPCC__ or in the HIP branch of the CMake file, so
the CUDA build's sources are unchanged after preprocessing.

Docs: EDGE0.md gets a "Build (Windows, AMD, experimental)" section.

Limits:

• The fused latent kernels stay CUDA-only. On HIP the model runs with STRATA_LATENT_FUSED=0,
  STRATA_LATENT_FUSED_DU=0, STRATA_LAT_FUSED_Q3K=0, so the prompt speed is far below the NVIDIA figures.
• edge0-infer does not start this build; the server is started by hand with a configuration file.
• labels.bin is located by the last / of STRATA_LATENT_A, so a path with only \ fails. Not changed here;
  the documentation says to write the path with /.

Area

• Engine, CLI or server (edge0-infer/)
• Model asset tools (edge0-infer/tools/assets/)
• Chat app examples (edge0-infer/examples/)
• Model docs (docs/)
• Other (edge0-strata/)

How was it tested?

Environment: Windows 11 (10.0.26300), Ryzen 7 5800X3D, 64 GB RAM, Radeon RX 7900 XTX 24 GB, driver
32.0.31041.1004, ROCm 10.2.0a20260930 (TheRock wheels), Visual Studio Build Tools, base commit b7900cb.

Passed:

• HIP build: set STRATA_HIP_ARCHS=gfx1100, set STRATA_NO_VISION=1, tools\hip\build_windows.bat (targets
  strata, strata-device, packaged zip).

• Run: serve\server.py --engine strata --config <file> with Edge0-145B-q3k, D/U Q8, --vram-reserve-mib 3072
  (about 4100 experts resident). Chat requests through the web page and /v1/chat/completions return answers.

• tools\strata_bench.py (medians):

  A stage of the prompt	Chat decode tok/s	4K prompt tok/s	16K prompt tok/s	Agent decode tok/s
  default (native_expert_grouped)	37.9	243.6	241.3	43.6
  STRATA_LATENT_MMQ=1	40.4	335.0	319.3	47.2
  STRATA_LATENT_MMQ=1, STRATA_ROUTE_TAIL_SKIP=7, STRATA_ROUTE_PRIOR=1.5	53.7	337.3	319.5	57.4

Not run:

• The CUDA build (no NVIDIA card here).
• The HIP unit tests (build_windows.bat tests) and the latent_*_verify tools.
• Linux, other Radeon cards, Edge0 120B, the NVFP4 folders.
• Any quality evaluation beyond reading the answers.

Before the changes: the HIP build stopped at verify.cpp, moe_fused.cu, moe_fused_q3k.cu and prefill.cpp;
once it built, every request failed with the capture error above.

Checklist

• ctest --preset cpu -L "L0|L1" passes locally in edge0-infer/ (engine changes) - not applicable, no change in edge0-infer/
• Ran the higher test levels this change affects (see edge0-infer/docs/testing.md) - not run, see above
• Added/updated tests for new behavior - none added
• Updated relevant docs / READMEs (incl. translations if user-facing) - edge0-strata/EDGE0.md (English only)

Co-Authored-By: Claude Fable 5.2

[brycesub]

Linux / RDNA2 data point: this PR's commit (78fe2d4) on gfx1030, 16 GB

Before finding this PR I hit and fixed the same issues on my own: the HIP CMake sources, the latent launch guards, the unmapped debug-dump names, and the verify "top3" copy failing graph capture with end capture: operation failed due to a previous error during capture. Then I built and ran this PR's exact commit, unmodified. Results are below.

Environment: CachyOS (Linux 7.3), Ryzen 7 5800X, 64 GB RAM, RX 6950 XT 16 GB (gfx1030, RDNA2), ROCm TheRock 7.13.0a20260515 wheels. Plain cmake build, no build_windows.bat equivalent:

cmake -S . -B build-hip -G Ninja -DCMAKE_BUILD_TYPE=Release -DSTRATA_ENABLE_HIP=ON -DSTRATA_ENABLE_CUDA=OFF \
  -DCMAKE_HIP_ARCHITECTURES=gfx1030 -DSTRATA_PREFILL_MMQ=ON -DSTRATA_AMD_COMMUNITY_TUNING=ON \
  -DSTRATA_GGML_DIR=$PWD/third_party/llama.cpp   # + CMAKE_HIP_COMPILER / CMAKE_PREFIX_PATH for the TheRock wheels
cmake --build build-hip --target strata strata-device

Passed

• Build of strata and strata-device for gfx1030: no errors. CMake prints the existing "gfx1030 builds, but it is not validated" warning. strata-device --selftest: OK.

• Run with Edge0-145B-q3k, D/U Q8, --vram-reserve-mib 3072, started by the unmodified edge0-infer serve launcher: --strata-dir points at this worktree, edge0-strata/build is symlinked to build-hip, and the ROCm libraries come from LD_LIBRARY_PATH. I kept the stock template, so STRATA_LATENT_FUSED=1, STRATA_LATENT_FUSED_DU=1 and STRATA_LAT_FUSED_Q3K=1 were left on. They fall through to the generic path because lat_dev().ok is false on HIP. So the "edge0-infer does not start this build" limitation does not apply on Linux.

• /v1/chat/completions through the edge0-infer proxy: the answers were correct, and a tool request returned a well-formed tool_calls entry.

• tools/strata_bench.py --url <engine> with the defaults (3 runs, thinking off). No failed requests.

  chat decode tok/s	4K prompt tok/s	16K prompt tok/s	agent decode tok/s	agent turn first token	agent cache hit
  23.7 (23.0–31.8)	173.1 (171.1–173.3)	173.1 (171.5–173.3)	23.6 (17.5–26.4)	19.5 s (10.6–31.5)	65.5%

  Prompt reading is flat at about 173 tok/s from 4K to 16K. With 1407 experts resident on 16 GB, the CPU share is active, so the D-stage copies this PR moves to copy_bytes_from_mapped do run.

Failed: none observed on this commit.

Notes from my own patch set (base da9cc52, not this PR):

• I never touched the two D-stage cudaMemcpyAsync calls, and they were captured without error on Linux even with the CPU share active. They may only fail on Windows, but I haven't verified that. Either way, the copy-kernel version in this PR works here.
• STRATA_HIP_PROMPT_F16=1 read a 3.7K-token prompt in 24–26 s instead of 30 s (single runs, same server). Decode speed was unchanged.

Not run: ctest, any CUDA build, Windows, and this PR with the fused variables set to 0 as EDGE0.md describes. I only ran the template-driven launch.