Skip to content
View dundysm's full-sized avatar

Block or report dundysm

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Content in all repositories owned by your account will be closed.
Maximum 250 characters. Please don’t include any personal information such as legal names or email addresses. Markdown is supported. This note will only be visible to you.
Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
dundysm/README.md

vllm-project/production-stack#1089: prefix-aware routing fell back to windowed qps, which pinned sessions to saturated backends; now least-inflightgoogle-deepmind/mujoco_warp#1813: contact sensors asking for linvel or angvel raised NotImplementedError; now computed in the velocity stage, matching mujoco cray-project/ray#66307: handle-metric staleness compared clocks across machines; now aged by the controller's monotonic receive timegoogle-deepmind/mujoco_menagerie#334: go2 mjx back-thigh range and calf torque (24 vs 45.43 nm) had drifted from the urdf; now they matchai-dynamo/dynamo#15371: p/d decode dropped multi_modal_data, so same-size images collided in the prefix cache and reused each other's kvmoveit/moveit2#3853: execution_complete_ was a plain bool read outside its mutex, a c++ data race; now std::atomicNVIDIA/TensorRT-Model-Connect#1284: zephyr and tinyllama chat templates were misdetected as phi, so the first decode step returned eosnewton-physics/newton#4780: mjcf import parsed a mesh file again for every geom using it; now once per import (go2: 1924 → 972 ms)ai-dynamo/aiperf#1436: late warmup histogram samples leaked into profiling server metrics; warmup now waits for the flush before its last scrapeRoboparty/rpo_description#5: 15 mjcf joint ranges and every motor limit had drifted from the urdf; synced, plus a script that fails on driftvllm-project/router#456: proxy-connection, a hop-by-hop alias of connection, was forwarded to workers; now stripped both waysRoboparty/human-humanoid-tools#14: the wheel never shipped the reference poses, so lafan retargeting fell back to a synthetic rest pose; now packagedkubernetes-sigs/kueue#16630: a new leader could schedule from a half-filled cache, so occupied nodes looked free and cohort quota went missing; now it waits for the initial syncflashinfer-ai/flashinfer#5335: the blackwell gdn kernel freed a tmem stage while loads from it were in flight, so the next mma could overwrite it; now fencedNVIDIA/gpu-operator#2843: the operator and its helm hook jobs ran without a securityContext; restricted defaults for k8s and openshiftNVIDIA/TransformerEngine#3565: non-fp8 gemm never set cublaslt's bias dtype, so a bf16 bias read as fp32 ran out of bounds; now rejectedMellanox/network-operator#3189: the rdma shared device plugin assumed /var/lib/kubelet; the kubelet root is now configurable through helmhuggingface/candle#3973: rms_norm always failed on an unconditional bias lookup and lazy VarMap init broke; both work againMellanox/k8s-rdma-shared-dev-plugin#329: when daemonset pods overlapped, the exiting one unlinked the new socket and rdma capacity dropped to zero; now only the owner removes itNVIDIA-NeMo/Automodel#3999: nemotron-h checkpoints saved with a singular embedding key failed dtensor load; both key forms now mapMellanox/maintenance-operator#243: the controller oomkilled at 128Mi on ~141 nodes (idle 130-156Mi); raised to 256Mi with sizing guidancetraining & kernelsin progress: open prs upstream; builtr1 sparring gym: mjlab footwork policies for a simulated unitree r1, sparring a second r1 on its own motors; 127/128 hard held-out bouts survivedocclubench: a mocap plate slides into one camera mid-episode; every occluded rollout is paired with a clean one, on mujoco warpinference under pressure: pre-registered vllm + dynamo experiments on kv budget, session affinity and kv-aware routing, from one a40 to 2× l40sgaitkeeper: checks obs → action → effort → next state against a golden trace to tell contract, physics and task failures apartinference doctor: welch t gate on repeated runs; now also $/1m tokens, idle-gpu $ and ranked cost fixes from vllm and sglang metricscenter of mass: mujoco wasm at 200 hz runs unitree's g1 policy at 50 hz; a gaussian-splat scan becomes collision boxes with a* pathsforgetting: replays vllm v1's lru block manager over measured runs; each kv token slot is a point on an mri-derived cortex in three.jssea of fog: typesafe's new jev model takes napoleon's bait at austerlitz 79% of the time; one warning line cuts that to 44%

Pinned Loading

  1. gaitkeeper gaitkeeper Public

    find why a humanoid locomotion policy fails between simulators: contract, physics, or task. MuJoCo, CPU.

    Python 2 1

  2. inference-doctor inference-doctor Public

    Open-source inference cost doctor for self-hosted LLM serving: $/1M tokens from Prometheus /metrics, API price comparison, and a CI gate (PASS/FAIL/INCONCLUSIVE) for performance and cost regression…

    Python 1

  3. inference-under-pressure inference-under-pressure Public

    Reproducible performance investigations into LLM inference, KV-cache memory pressure, scheduling, and distributed serving.

    Python 1

  4. occlubench occlubench Public

    OccluBench: GPU-native benchmark for VLA robustness to mid-episode physical occlusion (MuJoCo Warp)

    Python 1

  5. sea-of-fog sea-of-fog Public

    napoleon's battles on a sand table, with an ai general deciding at the key moment. austerlitz: typesafe's jev takes the bait 79% of the time.

    Python