Ilker ManapandClaude Opus 5.5 057be16a92 Do not call a GPU shard ready while it computes on the CPU ...

llama.cpp whose CUDA backend cannot load runs on the CPU without an
error. The CUDA build is now asked for its devices when installed, a
shard on GPUs must hold memory on each before it is ready, and
engine-check reports the devices. Shards get CUDA_DEVICE_ORDER=PCI_BUS_ID
so CUDA numbers GPUs as placement does.

Refs #2 (item 2).

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

2026-10-01 11:54:39 +02:00

..

discover_test.go

Report node memory even when GPU discovery fails

2026-09-16 23:50:53 +02:00

discover.go

Document what changed since v0.9.2, and fabrics for the sites that have them

2026-09-29 10:14:58 +02:00

fabrics_test.go

Stop trusting hardware facts that were never checked

2026-09-17 17:29:15 +02:00

fabrics.go

Add llmgw doctor: what a node's GPUs, NVLink and fabric need, checked

2026-09-29 03:51:09 +02:00

gpu_fabric_test.go

Find multi-node NVLink domains, place across them, and show them as racks

2026-09-29 09:07:59 +02:00

gpu_fabric.go

Find multi-node NVLink domains, place across them, and show them as racks

2026-09-29 09:07:59 +02:00

gpu_nonvml.go

Phase 2 (2/2): NVML backend behind a build tag

2026-09-16 02:05:25 +02:00

gpu_nvml.go

Phase 2 (2/2): NVML backend behind a build tag

2026-09-16 02:05:25 +02:00

gpu_smi.go

Read NVLink and GPU-to-adapter distance from nvidia-smi topo

2026-09-29 00:09:26 +02:00

gpu_test.go

Export what the cluster is made of, not only what it is serving

2026-09-17 15:09:27 +02:00

gpu.go

Find multi-node NVLink domains, place across them, and show them as racks

2026-09-29 09:07:59 +02:00

gpuprocs_test.go

Do not call a GPU shard ready while it computes on the CPU

2026-10-01 11:54:39 +02:00

gpuprocs.go

Do not call a GPU shard ready while it computes on the CPU

2026-10-01 11:54:39 +02:00

host.go

Install vLLM on every GPU node that can run it, as llama.cpp is

2026-09-29 13:15:54 +02:00

ib_test.go

Watch each fabric's subnet managers, and keep a fabric one through a failover

2026-09-29 07:47:03 +02:00

ib.go

Tell independent fabrics apart by their subnet manager, not their name

2026-09-29 07:29:16 +02:00

json_test.go

Phase 2 (1/2): node discovery, engine adapters, process supervisor

2026-09-16 02:01:59 +02:00

load_test.go

Measure each node's CPU and memory, and give every node a page

2026-09-28 23:21:23 +02:00

load.go

Measure each node's CPU and memory, and give every node a page

2026-09-28 23:21:23 +02:00

memory_test.go

Place on nodes that can actually hold the model

2026-09-16 23:12:44 +02:00

memory.go

Place on nodes that can actually hold the model

2026-09-16 23:12:44 +02:00

perftest_test.go

Measure two nodes' fabric rail by rail: llmgw ctl fabric-test

2026-09-29 21:29:23 +02:00

perftest.go

Measure two nodes' fabric rail by rail: llmgw ctl fabric-test

2026-09-29 21:29:23 +02:00

sensors_test.go

Read nodes' heat and power: GPUs, CPUs, board inlet, RAPL

2026-09-30 19:49:08 +02:00

sensors.go

Read nodes' heat and power: GPUs, CPUs, board inlet, RAPL

2026-09-30 19:49:08 +02:00

sminfo.go

Map every InfiniBand fabric, draw it, and place replicas by it

2026-09-29 08:46:23 +02:00

storage_test.go

Let a node say how much room it has left

2026-09-18 09:38:57 +02:00

storage.go

Let a node say how much room it has left

2026-09-18 09:38:57 +02:00

topo_test.go

Read NVLink and GPU-to-adapter distance from nvidia-smi topo

2026-09-29 00:09:26 +02:00

topo.go

Add llmgw doctor: what a node's GPUs, NVLink and fabric need, checked

2026-09-29 03:51:09 +02:00

weights_test.go

Know where the weights are, without owning the storage

2026-09-17 08:21:16 +02:00

weights.go

Know where the weights are, without owning the storage

2026-09-17 08:21:16 +02:00