Skip to content

Lowvram 4gb - #456

Open
BenMohStem wants to merge 5 commits into
FlashML-org:mainfrom
BenMohStem:lowvram-4gb
Open

BenMohStem wants to merge 5 commits into
FlashML-org:mainfrom
BenMohStem:lowvram-4gb

Conversation

@BenMohStem

Copy link
Copy Markdown

No description provided.

… diag)

- add --kv-dtype fp8 KV-cache storage (compute stays BF16; query dtype
  threaded separately through the flashinfer backend; pool slab priced
  with the storage dtype)
- FREETOKEN_CPU_EMBED=1 moves the 970 MiB embed_tokens table to host RAM
  (requires --cuda-graph-max-bs 0)
- fix: respect explicit --moe-cache-size and --disable-moe-prefill-overlap
  with the CPU MoE backend (stock 0.1.2 silently doubled 256 -> 512 slots
  and forced overlap back on)
- add [VRAM-DIAG]/[WEIGHT-DIAG] init diagnostics; right-size the
  flashinfer workspace floor (32 MiB) for small GPUs

Verified on RTX 3050 4 GiB laptop with nvidia/Qwen3.6-35B-A3B-NVFP4:
64K-token context pool, 65,349-token request completed, ~206-230 tok/s
prefill, ~10-14 tok/s decode. See README-4GB.md on the lowvram-4gb branch
(based on v0.1.2) for the full recipe and engineering notes.
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant