feat: support GLM-5.3-Flash NVFP4 - #292

Open

earlvanze wants to merge 25 commits into

FlashML-org:mainfrom

earlvanze:feat/glm53-flash

Commits on Aug 30, 2026

feat: start GLM-5.3-Flash support
Earl Co
committed
68e6242
feat: add GLM-5.3-Flash NVFP4 support
Earl Co
committed
a69e144
perf: speed up NVFP4 expert bank loading
Earl Co
committed
a74edf6
Add Vast Serverless integration for GLM 5.3 Flash
Earl Co
committed
c4feb45
Install Python headers for Vast worker builds
Earl Co
committed
4fa5984
Retry resumable Vast model downloads
Earl Co
committed
dd7036f
Overlap Vast checkpoint and dependency setup
Earl Co
committed
d4caac7
Keep Vast cold starts inside load deadline
Earl Co
committed
dcdd195
Add bounded Vast worker concurrency controls
Earl Co
committed
06693d5
Satisfy Vast worker lint checks
Earl Co
committed
37f02c1
feat: add persistent Vast GLM provisioner
Earl Co
committed
5bc4e4f
fix: pin persistent provisioner via advertised ref
Earl Co
committed
b1c7ce3
fix: install CUDA cuRAND headers for FlashInfer JIT
Earl Co
committed
5092dfe
fix: use NVIDIA cuRAND package name
Earl Co
committed
ab26800
fix: gate Vast readiness on engine health and JIT warmup
Earl Co
committed
615d58e
feat: allow Vast MoE backend selection
Earl Co
committed
09cda4a
fix: install cuRAND headers in serverless provisioner
Earl Co
committed
e49cfd4
fix: verify Serverless checkout pin
Earl Co
committed
3edf168
fix(glm5): materialize token-major KDA prefill inputs
Earl Co
committed
c61ef42
fix(vast): fast-path cached serverless resumes
Earl Co
committed
3d5354c
feat(serverless): add DeepSeek V4 Vast bootstrap
Earl Co
committed
2fb464b
fix(serverless): fetch shared Vast provisioner
Earl Co
committed
a0d1144

Commits on Aug 31, 2026

feat(vast): persist DeepSeek FTW cache
Earl Co
committed
753af37
fix(vast): detect completed FTW checkpoints
Earl Co
committed
bafb029
feat(vast): provision Qwen3.8 27B serverless
Earl Co
committed
55120d9

Navigation Menu

feat: support GLM-5.3-Flash NVFP4 - #292

feat: support GLM-5.3-Flash NVFP4#292
earlvanze wants to merge 25 commits into
FlashML-org:mainfrom
earlvanze:feat/glm53-flash

Commits on Aug 30, 2026

feat: start GLM-5.3-Flash support

feat: add GLM-5.3-Flash NVFP4 support

perf: speed up NVFP4 expert bank loading

Add Vast Serverless integration for GLM 5.3 Flash

Install Python headers for Vast worker builds

Retry resumable Vast model downloads

Overlap Vast checkpoint and dependency setup

Keep Vast cold starts inside load deadline

Add bounded Vast worker concurrency controls

Satisfy Vast worker lint checks

feat: add persistent Vast GLM provisioner

fix: pin persistent provisioner via advertised ref

fix: install CUDA cuRAND headers for FlashInfer JIT

fix: use NVIDIA cuRAND package name

fix: gate Vast readiness on engine health and JIT warmup

feat: allow Vast MoE backend selection

fix: install cuRAND headers in serverless provisioner

fix: verify Serverless checkout pin

fix(glm5): materialize token-major KDA prefill inputs

fix(vast): fast-path cached serverless resumes

feat(serverless): add DeepSeek V4 Vast bootstrap

fix(serverless): fetch shared Vast provisioner

Commits on Aug 31, 2026

feat(vast): persist DeepSeek FTW cache

fix(vast): detect completed FTW checkpoints

feat(vast): provision Qwen3.8 27B serverless