Haruto Watanabe · GPU & AI Inference Performance · XR / Game Engineering
Founder & Hands-on Technical Lead, Wildman Inc. · Tokyo, Japan
I build and optimize systems across real-time graphics, AI inference, and AMD Radeon / ROCm, with 15+ years of software development and XR work from Oculus Rift DK1 through Apple Vision Pro. My public work connects application-level bottlenecks to runtime behavior, numerical correctness, and measurable end-to-end results.
Recent results: three merged PRs in ROCmFPX; 20–50% faster end-to-end processing across my measured Qwen Image 2.1 and MiniMax H3 workflows using ROCmFPX; and MiniMax H3 audio-VAE processing reduced from approximately 5.0 s to 0.3 s for a 5-second video workload.
Important
Code related to commercial games developed at Wildman Inc. is managed separately by the company and is not included in this public GitHub account or its commit history. Therefore, the history shown here is not a complete record of my career or the full scope of my implementation work on commercial projects.
| Scope | Measured result | Interpretation |
|---|---|---|
| Qwen Image 2.1 and MiniMax H3 workflows using ROCmFPX | 20–50% faster E2E processing across measured workflows | Workload-specific application results; the reported speedup is not a universal hardware or model claim. |
| MiniMax H3 audio VAE for a 5-second video workload | ~5.0 s → ~0.3 s | ~16.7× stage speedup / ~94% less processing time, calculated from the rounded timings. This is the audio-VAE stage, not total video-generation time. |
These are my local measurements for the tested configurations. They are separate from the PR-specific correctness and compatibility validations below; the three merged PRs are not presented as individually causing all of these performance gains. The E2E range is reported as a speed improvement, not as a 20–50% reduction in elapsed time.
Related public update: ROCmFPX / generative-AI performance work.
All three PRs below were submitted from this account and merged into
charlie12345/ROCmFPX on September 24, 2026 (JST).
| PR | Contribution | Validation recorded in the PR |
|---|---|---|
| #114 — Preserve probability tails in entropy-bound decoding | Use double-precision accumulation for normalization, CDF, and entropy to avoid losing small positive probability tails in a 262,144-token vocabulary. | Analytic entropy, vocabulary-order stability, confidence boundaries, tail sampling, ties, and underflow regression coverage. This is a numerical-correctness fix, not a speed claim. |
| #115 — Stop truncating valid repeated output before EOG | Replace repetition-based truncation with end-of-generation-token handling so valid repeated digits and JSON-array content are not cut prematurely. | 14 model-free regression cases, including a 33-token fixture previously cut to 10 tokens. Prevents postprocessing corruption; does not guarantee valid model-generated JSON. |
| #116 — Support Qwen3VL hidden-state checkpoints | Enable pruned text encoders used by MiniMax H3 through optional hidden-state-only metadata while preserving ordinary Qwen3VL behavior and the public C ABI; include an upstream reduction-race fix. | On Windows / Radeon 8060S (gfx1151): 220/220 SOFT_MAX/GROUP_NORM tests; 280/280 bitwise-identical softmax repetitions; 30/30 bitwise-identical real-H3 hidden-state encodes. |
Development provenance: these PRs used OpenAI Codex for investigation,
implementation, tests, and/or validation, as disclosed in each PR. The
reduction-race fix in #116 is an unchanged backport of
ggml-org/llama.cpp commit 9bd4c09, with its original
authorship preserved. The validation results above are the recorded PR
results, not claims of broader model-quality or hosted-CI coverage.
- End-to-end generative-AI optimization on AMD GPUs: ROCmFPX, ROCm/HIP, PyTorch, Triton, low-precision inference, text encoders, and VAE stages
- Numerical correctness, reproducible benchmarks, regression testing, and clearly scoped performance claims
- AI agents and tools built around evidence, reproducibility, and local execution
- Interactive web experiences with TypeScript, Next.js, and Three.js
- Games and VR interactions with Unity, Unreal Engine, C++, and C#
The following projects were selected from this account's original public repositories.
- KoushiKo Vault — A local-first media investigation agent that produces answers grounded in document pages and video timecodes, together with an evidence reel
- ComfyUI RDNA35 Attention — Implementation and benchmarking research for ComfyUI attention on RDNA 3.5 using PyTorch ROCm and Triton
- ComfyUI AMD Video Upscaler —
A ComfyUI video upscaling node powered by AMD AMF and FFmpeg's
sr_amf - ModelDuel — A learning experience that compares predictions from two models under the same conditions and uses evidence to refine understanding
| Period | Area | Public repositories / contributions |
|---|---|---|
| 2012–2014 | HTML5 / JavaScript, games, and input interactions | RGBWars, blastbuster-vr |
| 2026 | Web / 3D products with integrated AI | ModelDuel |
| 2026 | AMD Radeon / ROCm, ComfyUI, and GPU optimization | ComfyUI AMD Video Upscaler, ComfyUI RDNA35 Attention |
| 2026 | Local AI and evidence-grounded agents | KoushiKo Vault |
| September 2026 | Merged inference-runtime correctness and model-compatibility contributions | ROCmFPX #114, #115, #116 |
Wherever possible, each project's README documents its design decisions, runtime requirements, testing procedures, and benchmark assumptions.
- Languages:
C++·Rust·Python·TypeScript·JavaScript·C# - Product / Web:
Unity·Unreal Engine·React·Next.js·Three.js·FastAPI - GPU / Media:
ROCmFPX·ROCm / HIP·PyTorch·Triton·ComfyUI·FFmpeg / AMF - Engineering: end-to-end profiling · numerical correctness · regression tests · AI-assisted development