Issues

model-serving: gemma-4-31b at 32k on the L40S needs an fp8 KV cache — prove it before the preset carries it

#611

· teemow opened

on Sep 19, 2026

model-serving: the predictor's emptyDir volumes carry no sizeLimit — Kyverno's require-emptydir-requests-and-limits audits every model pod

#610

· teemow opened

on Sep 19, 2026

model-serving: the gpt-oss presets answer 500 after Ready — harmony fetches the tiktoken vocab at the first request and the serving egress is closed

#606

· teemow opened

on Sep 19, 2026

model-serving: the cache claim carries its tier so its price survives its StorageClass

#605

· teemow opened

on Sep 19, 2026

llmRouting: served models on the LLM endpoint — model router, external hostname, API keys

#603

· teemow opened

on Sep 19, 2026

model-serving: admit model-manager's interface read, refuse --disable-fastapi-docs in presets

#602

· teemow opened

on Sep 19, 2026

model-serving: a Muse Glimmer 30B preset once the runtime carries vLLM 0.28 or later

#597

· teemow opened

on Sep 19, 2026

model-serving: refresh the shipped presets to the September 2026 open-weight line-up

#591

· teemow opened

on Sep 19, 2026

kagent, Substrate, CloudNativePG and the bundled Flux controllers from gsoci

#580

· teemow opened

on Sep 18, 2026

model-serving: page cache after a large pull vs the CUDA free figure

#579

· teemow opened

on Sep 18, 2026

model-serving: a preset that serves eager loses Triton's kernel cache with every pod — put TRITON_CACHE_DIR on the cache claim beside vLLM's compiled graphs

#572

· teemow opened

on Sep 18, 2026

model-serving: the KServe predictor pod's emptyDir volumes have no sizeLimit — Kyverno's require-emptydir-requests-and-limits audits it every ten seconds; bound them in the serving-pod mutation

#556

· teemow opened

on Sep 18, 2026

Navigation Menu

All issues

model-serving: gemma-4-31b at 32k on the L40S needs an fp8 KV cache — prove it before the preset carries it

model-serving: the predictor's emptyDir volumes carry no sizeLimit — Kyverno's require-emptydir-requests-and-limits audits every model pod

model-serving: the gpt-oss presets answer 500 after Ready — harmony fetches the tiktoken vocab at the first request and the serving egress is closed

model-serving: the cache claim carries its tier so its price survives its StorageClass

llmRouting: served models on the LLM endpoint — model router, external hostname, API keys

model-serving: admit model-manager's interface read, refuse --disable-fastapi-docs in presets

model-serving: a Muse Glimmer 30B preset once the runtime carries vLLM 0.28 or later

model-serving: refresh the shipped presets to the September 2026 open-weight line-up

kagent, Substrate, CloudNativePG and the bundled Flux controllers from gsoci

model-serving: page cache after a large pull vs the CUDA free figure

model-serving: a preset that serves eager loses Triton's kernel cache with every pod — put TRITON_CACHE_DIR on the cache claim beside vLLM's compiled graphs

model-serving: the KServe predictor pod's emptyDir volumes have no sizeLimit — Kyverno's require-emptydir-requests-and-limits audits it every ten seconds; bound them in the serving-pod mutation

Search results