Domestic GPU Enablement & Joint Optimization
Source-level inference-stack adaptation and measured validation across AMD MI308X, Huawei Ascend 910B, MetaX N260 and other platforms — turning domestic / non-NVIDIA accelerators into production capacity.
Core Capabilities and Measured Basis
- Source-level vLLM / LMCache builds and patches on ROCm: the LMCache parallel-read patch cuts TTFT by 4.1× (measured, R1)
- Multi-platform benchmark coverage: AMD MI308X ×8 (R1–R4), Huawei Ascend 910B ×8 (R9), MetaX N260 (R5 methodology)
- Multimodal adaptation: all 7 ComfyUI + LTX-Video 2.3 models run end-to-end on MI308X / ROCm 7.2 with no operator or kernel errors (R6/R7)
- Code-level reproducible delivery: git patches, load clients, orchestration and forensic scripts, full raw-data export (R8)
Supporting Evidence
LMCache parallel-read patch, multi-GPU KV tiering scale-out, concurrent model loading, model-switch effective TPS, training checkpoint concurrent writes (Qwen2.5-32B/7B).
Download report PDF ↓Seven-way comparison (GDS direct read / two re-compute HBM tiers / low-HBM direct read); three-step HBM-equivalence argument. Platform: MetaX N260 single GPU — methodology portable across platforms, numbers not mixed with MI308X.
Download report PDF ↓All 7 delivered models (3 VAE / fp8 text encoder / text projection / 2 LoRA) running end-to-end on MI308X / ROCm 7.2 with no dtype, operator, or kernel errors.
Download report PDF ↓Per-model runs and adaptation notes for 3 VAE + Singularity LoRA.
Download report PDF ↓LMCache parallel-read patch (git patch + full before/after), load clients, orchestration and forensic scripts, raw data — enables full third-party reproduction (contact us for access).
Contact us for access →Huawei Atlas 910B ×8 (Kunpeng-920): model-serving load (DeepSeek-32B/70B), training weight/checkpoint I/O (Qwen-7B), training-data acceleration (YOLOv8/COCO) — three groups against an NFS baseline (Ascend platform, labeled as such; contact us for access).
Contact us for access →The Engagement Path
Joint test first, decisions second: gate-based acceptance with built-in stop-loss. The full costing model is provided as reproducible Python after NDA — customers can rerun it with their own parameters. Every key figure on this site carries a report ID and is open to third-party verification.