Phase 4A2 Model-B public validation for AMD RDNA4 / gfx1201
This release freezes the public Phase-4A2 Model-B integration of the
rocWMMA Width-64 fused MLP path with the AMD gsplat image path on an
AMD Radeon AI PRO R9700.
Qualified source
- Tag:
phase4a2-model-b-public-gfx1201-pass - Tag object:
636dfc01673c97c43815c0205f53f5ed937b3249 - Commit:
b98bdcc6b2878f6cb6c10a2141e50867cec6d96a - Git tree:
a8ffaaa3f509400c40f6de58e8a74fb047f8e16e - Target:
gfx1201 - ROCm: 7.2
- PyTorch:
2.13.0+rocm7.2 - HIP reported by PyTorch:
7.2.53211
Qualification result
- Native public build: PASS
- Staged runtime import provenance: PASS
- Qualified-kernel immutability: PASS
- Decoder and rasterizer forward/backward: PASS
- No CPU fallback: PASS
- Real optimizer update: PASS
- 100-step finite/loss/signal/render/VRAM gates: PASS
- Local replay determinism: PASS
- Resume equivalence: PASS
- Fused-backend runtime attestation: PASS
- Backend negative tests: PASS
- Overall gate:
RDNA4_FUSED_MLP_PHASE4A2_AMD_GSPLAT_FUSED_ROCWMMA: PASS
Independent GitHub fresh-clone validation
An independent recursive clone from GitHub reproduced the exact public commit,
tree, and recursive submodule pins. It built successfully and passed the same
complete public GPU demo on the R9700.
Fresh-clone functional gate:
GITHUB_FRESHCLONE_FUNCTIONAL_VALIDATION: PASS
Binary fingerprint boundary
The SHA256 values below identify concrete build instances:
| Build instance | tiny-rdna4-nn | amd-gsplat |
|---|---|---|
| Qualified checkout | 883f89efdad7bb909a4a3899ab79b2defe9713fdb5c7cf22cf4882c626b3efc4 |
b5eab3c002cd622aa08b47094b03e642babd86b15a42318c5a465b5932882360 |
| GitHub fresh clone | aeeb40781cb16bfb2d31d4a5d95b2f550b54e50d30615323b6214599b78fc08d |
ad931a31572cde8d4b794683387cc11b5fa2cbc9bd7ca081a94db4f39367fb7c |
Cross-checkout binary reproducibility is not qualified.
Absolute checkout/build paths are embedded in the generated ELF/HIP binaries.
The fresh clone therefore does not produce bit-identical binaries at a
different absolute path. In the section-level comparison, selected
amd-gsplat code objects also had differing .text and .rodata bytes.
This release claims successful source provenance, buildability, and functional
GPU validation—not checkout-path-independent bit reproducibility.
Release assets
- Deterministic source archive with recursive submodule contents
- Public sanitized evidence archive
- Git bundle preserving the qualified public commit and annotated tag
- SHA256 manifest for all uploaded assets
The private qualification freeze is retained locally and is not uploaded
because it contains local validation paths.