feat(rocm): AMD ROCm runtime for RDNA3/4 with native Qwen3.5-MoE GGUF decode parity - #217

Open

samuelishida wants to merge 20 commits into

FlashML-org:mainfrom

samuelishida:feat/amd-rocm-gfx1100-support

Commits on Aug 25, 2026

feat(rocm): AMD ROCm support (gfx1100) + qwen35moe GGUF loader
Samuel Ishida
committed
9741646

Commits on Aug 26, 2026

fix(gguf): register unmerged CONTROL/USER_DEFINED tokens for atomic encoding
Samuel Ishida
committed
6f4974e
fix(server): route Qwen3.6 to the qwen3_coder tool-call parser
Samuel Ishida
committed
bc62324
chore(local): ROCm serve scripts + VS Code tasks/debug configs
Samuel Ishida
committed
25d7bd8
fix(qwen3.5-moe): GGUF lm_head last-token gather + presence/frequency penalties
Samuel Ishida
committed
729789a
chore(rocm): drop CI workflow + tinygrad fallback; review cleanup
Samuel Ishida
committed
1f76e88
Merge remote-tracking branch 'origin/main' into feat/amd-rocm-gfx1100-support
Samuel Ishida
committed
8a9d2df

Commits on Aug 30, 2026

Merge remote-tracking branch 'origin/main' into feat/amd-rocm-gfx1100-support
Samuel Ishida
committed
9ee03a2
fix(rocm): retain hybrid attention capability flag
Samuel Ishida
committed
ce11a07

Commits on Aug 31, 2026

Commits on Sep 2, 2026

feat(rocm): complete Qwen GGUF decode parity path
Samuel Ishida
committed
93819f7
Merge remote-tracking branch 'origin/main' into feat/amd-rocm-gfx1100-support
Samuel Ishida
committed
ac9c965