Netra MXFP4 kernels for gfx1151

Hand-written AMDGCN kernels for Qwen3.6-35B-A3B on the AMD Ryzen AI Max+ PRO 395 (gfx1151). The repository is MXFP4-only. HIP and Python code is limited to launching, correctness checking, checkpoint conversion, profiling, and serving integration.

All build, validation, profiling, and benchmark commands must run inside the Netra LXC.

Repository layout

scripts/rocm/
  kernels/        Canonical raw gfx1151 ASM for newly reorganized work
  harness/        HIP launch, graph, correctness, and timing bridges
  tools/          Reproducible build, benchmark, and profiling entry points
docs/netra/notes/ Current findings, disassemblies, and machine-readable evidence
kernels/gfx1151/mxfp4/
  decode/        M=1 decode kernels and retained experiments
  verify/        M=12 speculative-verification kernels
  prefill/       WMMA prefill kernels, shared include, and experiments
  lm_head/       MXFP4 LM-head kernels
  serving/       SGLang ABI kernels
  epilogue/      Raw assembly epilogues
kernels/gfx1151/attention/
  ...            Standard-attention kernels and rejected experiments
kernels/gfx1151/gdn/
  ...            Gated-delta-network kernels and rejected experiments
kernels/gfx1151/moe/
  ...            MoE routing/packing kernels and rejected experiments
harness/gfx1151/mxfp4/
  ...            HIP launch, timing, and correctness harnesses
harness/gfx1151/{attention,gdn,moe}/
  ...            Non-MXFP4 launch, timing, and correctness harnesses
scripts/rocm/integrations/sglang/
  ...            SGLang bridge, launch script, and consolidated patch
tools/
  build/         Reproducible gfx1151 builds
  benchmark/     Kernel and serving benchmarks
  checkpoint/    One-time real-checkpoint extraction and repacking
  profiling/     rocprofv3 request and trace analysis tools
  triton/        MXFP4 reference-baseline preparation
tests/gfx1151/   Real-checkpoint integration correctness gates
docs/notes/      Measurements, negative results, and machine-readable data
results/         Selected compact kernel, profiler, and serving evidence

New mission work is canonical under scripts/rocm/ and docs/netra/notes/. The older top-level kernel/tool and docs/notes/ paths remain in place for reproducibility while they are migrated incrementally; new sources must not be duplicated into those legacy trees.

Files under an experiments/ directory are retained measured negative results; they are built for reproducibility but are not selected as shipping kernels.

Build

Run inside Netra:

cd /root/netra-mxfp4-gfx1151
bash tools/build/build_gfx1151_mxfp4_raw.sh
bash scripts/rocm/tools/build/build_netra_sglang_gfx1151.sh

The raw build writes to build/raw by default. The serving build writes to build/sglang and produces the raw-ASM HSACOs plus the HIP launch-only shared library.

SGLang integration

The validated integration base is SGLang commit 1eee8fbdcc25b44e13bc097d5ff6ac24e8c24af4.

cd /root/work/sglang-main
git apply \
  /root/netra-mxfp4-gfx1151/scripts/rocm/integrations/sglang/sglang-gfx1151-integration.patch
git apply \
  /root/netra-mxfp4-gfx1151/scripts/rocm/integrations/sglang/sglang-gfx1151-qkvz-ba-fusion.patch
git apply \
  /root/netra-mxfp4-gfx1151/scripts/rocm/integrations/sglang/sglang-gfx1151-extend-attention.patch
git apply \
  /root/netra-mxfp4-gfx1151/scripts/rocm/integrations/sglang/sglang-gfx1151-gdn-chunk-o.patch
git apply \
  /root/netra-mxfp4-gfx1151/scripts/rocm/integrations/sglang/sglang-gfx1151-gdn-recompute-w-u.patch
git apply \
  /root/netra-mxfp4-gfx1151/scripts/rocm/integrations/sglang/sglang-gfx1151-causal-conv1d.patch
git apply \
  /root/netra-mxfp4-gfx1151/scripts/rocm/integrations/sglang/sglang-gfx1151-qk-mrope-kv-fusion.patch
git apply \
  /root/netra-mxfp4-gfx1151/scripts/rocm/integrations/sglang/sglang-gfx1151-mamba-track-host-flag.patch
git apply \
  /root/netra-mxfp4-gfx1151/scripts/rocm/integrations/sglang/sglang-gfx1151-gdn-syncfree-chunk-metadata.patch
git apply \
  /root/netra-mxfp4-gfx1151/scripts/rocm/integrations/sglang/sglang-gfx1151-bf16-lm-head.patch
cp /root/netra-mxfp4-gfx1151/scripts/rocm/integrations/sglang/netra_gfx1151_sglang.py \
  python/sglang/srt/layers/quantization/netra_gfx1151.py

cd /root/netra-mxfp4-gfx1151
bash scripts/rocm/integrations/sglang/launch.sh

The launch script defaults to the real transformed checkpoint at /root/models/qwen36-sgl-mxfp4. Environment variables in the script allow overriding the SGLang tree, virtual environment, model, host, port, memory fraction, context length, and visible device.

Correctness

Run the real-checkpoint SGL ABI gates inside Netra:

cd /root/netra-mxfp4-gfx1151
NETRA_TEST_M=65 NETRA_TEST_N=64 \
  /root/sglvenv1151/bin/python tests/gfx1151/test_netra_mxfp4_sgl_linear.py
/root/sglvenv1151/bin/python \
  tests/gfx1151/test_netra_mxfp4_sgl_decode.py

Both tests compare gfx1151 outputs with staged fp64 dequantized-MXFP4 references using real checkpoint weights.

Results and implementation notes

• Kernel development and negative results
• Machine-readable kernel results
• SGLang integration and serving results
• Machine-readable SGLang results
• Raw ASM versus Triton MXFP4 comparison
• Fast real-checkpoint shard loading
• Accepted extend-attention K LDS swizzle

Every reported performance number in those records is labeled measured or estimated and is specific to gfx1151.