Known issues

Known issues (HIP)

Gaps that do not block default ./h3 T2VA. Build with HIP_ARCH=gfx1151 or HIP_ARCH=gfx90a.

• gfx1151: BF16 activations, INT8 DiT weights, wave32 rocWMMA SDPA.
• gfx90a: BF16 DiT weights (hipBLAS BF16 GEMM), wave64 MFMA flash. Restore INT8 with H3_INT8_MLP=1.

In-tree copy: docs/KNOWN_ISSUES.md.

KI-001: CPU Euler sampler

Status: intentional until a GPU Euler path is proven.

h3_gpu_is_m5() returns 0, so DiT uses the host Euler sampler. Default CLI works; timing and exact step numerics differ from M5 Metal.

KI-002: Nearest-neighbor host scale

Status: HIP host fallback (no Accelerate / vImage).

--ref-image / first-last when the still is scaled: slightly sharper/blockier than Metal. Same-size canvases are unaffected.

KI-003: MLX fixture numerical parity

Status: pending without misc/fixtures/.

HIP correctness is gated by CPU-oracle kernel tests, make HIP_ARCH=… test, make hip-functional, and end-to-end smokes.

KI-004: Apple TensorOps kernels

Status: deferred. HIP uses its own INT8 tiles (sdot4) and fused MLP kernels.

Backlog

ID	Topic	Status
P	Remaining E2E	gfx1151 fox-s2 wall is mostly NVMe I/O. gfx90a long T2VA is still DiT SDPA. See Performance.
V	--ref-video-audio VIDEO AUDIO E2E clip	Same kernels as --ref-video; not separately showcased