bloomery good first issues (Rust)

experiment: an opt-in lever that drops low-ranked missed experts in the verify window (speed vs KL) 2 days ago

help wanted measure

19

Hybrid GPU + CPU inference for mixture-of-experts models, written in Rust down to the CUDA kernels. A drop-in llama-server.

Rust

#cuda#deepseek#gguf#glm#inference#llama-cpp#llm#moe#qwen#rust

gpu-gates: SeatFlags::bind takes the levers, not one bool per lever 5 days ago

good first issue gpu serve

midagedev/bloomery

19

Hybrid GPU + CPU inference for mixture-of-experts models, written in Rust down to the CUDA kernels. A drop-in llama-server.

Rust

#cuda#deepseek#gguf#glm#inference#llama-cpp#llm#moe#qwen#rust

qdot: AVX-512 (VNNI) host kernels, and one owner that picks the ISA 6 days ago

enhancement help wanted host

midagedev/bloomery

19

Hybrid GPU + CPU inference for mixture-of-experts models, written in Rust down to the CUDA kernels. A drop-in llama-server.

Rust

#cuda#deepseek#gguf#glm#inference#llama-cpp#llm#moe#qwen#rust

tools: a prose mention in ds41_prefill.py over-selects weekly-gpu-ds41-flowcounts 6 days ago

good first issue tools

midagedev/bloomery

19

Hybrid GPU + CPU inference for mixture-of-experts models, written in Rust down to the CUDA kernels. A drop-in llama-server.

Rust

#cuda#deepseek#gguf#glm#inference#llama-cpp#llm#moe#qwen#rust

One owner for a file's sha256 (four copies today) 6 days ago

good first issue tools

midagedev/bloomery

19

Hybrid GPU + CPU inference for mixture-of-experts models, written in Rust down to the CUDA kernels. A drop-in llama-server.

Rust

#cuda#deepseek#gguf#glm#inference#llama-cpp#llm#moe#qwen#rust

measure: GLM-5.3-Flash with the MTP draft on and off (same binary) 8 days ago

help wanted measure

midagedev/bloomery

19

Hybrid GPU + CPU inference for mixture-of-experts models, written in Rust down to the CUDA kernels. A drop-in llama-server.

Rust

#cuda#deepseek#gguf#glm#inference#llama-cpp#llm#moe#qwen#rust

측정: 익명 mmap의 cudaHostRegister 비용 (s/GiB)과 등록 버퍼의 H2D 11 days ago

help wanted hardware measure

midagedev/bloomery

19

Hybrid GPU + CPU inference for mixture-of-experts models, written in Rust down to the CUDA kernels. A drop-in llama-server.

Rust

#cuda#deepseek#gguf#glm#inference#llama-cpp#llm#moe#qwen#rust

dgx-spark: decode step timeline (nsys) and sampled vs greedy decode rate on GB10 11 days ago

help wanted hardware dgx-spark

midagedev/bloomery

19

Hybrid GPU + CPU inference for mixture-of-experts models, written in Rust down to the CUDA kernels. A drop-in llama-server.

Rust

#cuda#deepseek#gguf#glm#inference#llama-cpp#llm#moe#qwen#rust

measure: next-layer routing prediction recall (the premise of paged read-ahead) 12 days ago

help wanted measure

midagedev/bloomery

19

Hybrid GPU + CPU inference for mixture-of-experts models, written in Rust down to the CUDA kernels. A drop-in llama-server.

Rust

#cuda#deepseek#gguf#glm#inference#llama-cpp#llm#moe#qwen#rust

측정: MoE 모델의 검증 pass 비용 r(m) = pass(m)/pass(1), lookup draft 폭과 expert 묶기를 정함 12 days ago

help wanted hardware measure

midagedev/bloomery

19

Hybrid GPU + CPU inference for mixture-of-experts models, written in Rust down to the CUDA kernels. A drop-in llama-server.

Rust

#cuda#deepseek#gguf#glm#inference#llama-cpp#llm#moe#qwen#rust

gpu-gates: required-features 뒤의 죽은 stub main 정리 12 days ago

help wanted good first issue gpu

midagedev/bloomery

19

Hybrid GPU + CPU inference for mixture-of-experts models, written in Rust down to the CUDA kernels. A drop-in llama-server.

Rust

#cuda#deepseek#gguf#glm#inference#llama-cpp#llm#moe#qwen#rust

gpu-gates: 카드 이름 토큰 .replace(' ', "_")가 7곳에 복사됨 12 days ago

help wanted good first issue gpu

midagedev/bloomery

19

Hybrid GPU + CPU inference for mixture-of-experts models, written in Rust down to the CUDA kernels. A drop-in llama-server.

Rust

#cuda#deepseek#gguf#glm#inference#llama-cpp#llm#moe#qwen#rust

GLM step: the card fills the next step's inputs (no host round trip) 13 days ago

help wanted

midagedev/bloomery

19

Hybrid GPU + CPU inference for mixture-of-experts models, written in Rust down to the CUDA kernels. A drop-in llama-server.

Rust

#cuda#deepseek#gguf#glm#inference#llama-cpp#llm#moe#qwen#rust

측정: 샘플러를 단일 확률 임계값(w = p/p_max)으로 바꾸면 얼마나 빨라지나 13 days ago

help wanted no-box

midagedev/bloomery

19

Hybrid GPU + CPU inference for mixture-of-experts models, written in Rust down to the CUDA kernels. A drop-in llama-server.

Rust

#cuda#deepseek#gguf#glm#inference#llama-cpp#llm#moe#qwen#rust

serve: a start screen for people 13 days ago

help wanted good first issue

midagedev/bloomery

19

Hybrid GPU + CPU inference for mixture-of-experts models, written in Rust down to the CUDA kernels. A drop-in llama-server.

Rust

#cuda#deepseek#gguf#glm#inference#llama-cpp#llm#moe#qwen#rust

model: qwen35moe PlanInputs::describe가 호스트 room을 직접 읽음 — describe_in(split, room)으로 분리 13 days ago

good first issue no-box

midagedev/bloomery

19

Hybrid GPU + CPU inference for mixture-of-experts models, written in Rust down to the CUDA kernels. A drop-in llama-server.

Rust

#cuda#deepseek#gguf#glm#inference#llama-cpp#llm#moe#qwen#rust

tools: recipes.py가 한 줄짜리 속성 붙은 mod x;를 못 봄 14 days ago

good first issue no-box

midagedev/bloomery

19

Hybrid GPU + CPU inference for mixture-of-experts models, written in Rust down to the CUDA kernels. A drop-in llama-server.

Rust

#cuda#deepseek#gguf#glm#inference#llama-cpp#llm#moe#qwen#rust

tools: lcpp-warm.sh --self-test가 Python stub 서버를 남김 14 days ago

good first issue no-box

midagedev/bloomery

19

Hybrid GPU + CPU inference for mixture-of-experts models, written in Rust down to the CUDA kernels. A drop-in llama-server.

Rust

#cuda#deepseek#gguf#glm#inference#llama-cpp#llm#moe#qwen#rust

serve 테스트: mock::Hold 복사본 제거 14 days ago

good first issue no-box

midagedev/bloomery

19

Hybrid GPU + CPU inference for mixture-of-experts models, written in Rust down to the CUDA kernels. A drop-in llama-server.

Rust

#cuda#deepseek#gguf#glm#inference#llama-cpp#llm#moe#qwen#rust

serve: 요청의 samplers 필드를 무시하지 말고, 우리 순서와 다르면 거부 14 days ago

good first issue no-box

midagedev/bloomery

19

Hybrid GPU + CPU inference for mixture-of-experts models, written in Rust down to the CUDA kernels. A drop-in llama-server.

Rust

#cuda#deepseek#gguf#glm#inference#llama-cpp#llm#moe#qwen#rust

측정: Qwen3.8 expert 라우팅 국소성 (층별 LRU-64/128 적중률) 14 days ago

help wanted no-box

midagedev/bloomery

19

Hybrid GPU + CPU inference for mixture-of-experts models, written in Rust down to the CUDA kernels. A drop-in llama-server.

Rust

#cuda#deepseek#gguf#glm#inference#llama-cpp#llm#moe#qwen#rust

tools: ptx-scan에 ld.global 폭(스칼라/.v2/.v4) 열 추가 15 days ago

help wanted good first issue

midagedev/bloomery

19

Hybrid GPU + CPU inference for mixture-of-experts models, written in Rust down to the CUDA kernels. A drop-in llama-server.

Rust

#cuda#deepseek#gguf#glm#inference#llama-cpp#llm#moe#qwen#rust

plan 레코드: 호스트에 고정된 layer와 그 이유(host_only_reason)가 출력에 없음 15 days ago

help wanted good first issue

midagedev/bloomery

19

Hybrid GPU + CPU inference for mixture-of-experts models, written in Rust down to the CUDA kernels. A drop-in llama-server.

Rust

#cuda#deepseek#gguf#glm#inference#llama-cpp#llm#moe#qwen#rust

serve: 지원하지 않는 샘플링 필드(typical_p, mirostat, dry_*, xtc_* 등)를 무시하지 말고 거부 15 days ago

good first issue no-box

midagedev/bloomery

19

Hybrid GPU + CPU inference for mixture-of-experts models, written in Rust down to the CUDA kernels. A drop-in llama-server.

Rust

#cuda#deepseek#gguf#glm#inference#llama-cpp#llm#moe#qwen#rust

serve: /tokenize가 content가 없거나 문자열이 아니면 ""로 읽음 15 days ago

good first issue no-box

midagedev/bloomery

19

Hybrid GPU + CPU inference for mixture-of-experts models, written in Rust down to the CUDA kernels. A drop-in llama-server.

Rust

#cuda#deepseek#gguf#glm#inference#llama-cpp#llm#moe#qwen#rust

serve: n_predict가 -1보다 작으면 조용히 -1로 바뀜 15 days ago

good first issue no-box

midagedev/bloomery

19

Hybrid GPU + CPU inference for mixture-of-experts models, written in Rust down to the CUDA kernels. A drop-in llama-server.

Rust

#cuda#deepseek#gguf#glm#inference#llama-cpp#llm#moe#qwen#rust

Blackwell probe: does the block-scaled FP4 MMA (mxf8f6f4) give GGUF MXFP4's values bit for bit? 15 days ago

help wanted hardware dgx-spark

midagedev/bloomery

19

Hybrid GPU + CPU inference for mixture-of-experts models, written in Rust down to the CUDA kernels. A drop-in llama-server.

Rust

#cuda#deepseek#gguf#glm#inference#llama-cpp#llm#moe#qwen#rust

dgx-spark: is prefill byte-bound or tensor-bound on GB10? (grouped GEMM time, pp512/pp4096) 15 days ago

help wanted hardware dgx-spark

midagedev/bloomery

19

Hybrid GPU + CPU inference for mixture-of-experts models, written in Rust down to the CUDA kernels. A drop-in llama-server.

Rust

#cuda#deepseek#gguf#glm#inference#llama-cpp#llm#moe#qwen#rust

fuzz: 신뢰할 수 없는 텍스트를 읽는 파서에 cargo-fuzz 타깃 추가 15 days ago

good first issue no-box

midagedev/bloomery

19

Hybrid GPU + CPU inference for mixture-of-experts models, written in Rust down to the CUDA kernels. A drop-in llama-server.

Rust

#cuda#deepseek#gguf#glm#inference#llama-cpp#llm#moe#qwen#rust

midagedev/bloomery

Open Issues Need Help