Commits

Commits on Aug 8, 2026

Merge pull request #7 from BudEcosystem/claude/qwen-memory-requirements-a3f401
dittops
authored
9c1a0dc
fix(memory): address review — three defects the fixture corpus could not see

root
and
claude
committed
df00566
fix(memory): hybrid architectures have eleven config dialects, and we read one

root
and
claude
committed
0b168b5

Commits on Aug 1, 2026

Merge pull request #6 from BudEcosystem/claude/token-budget-and-lora-scratch
dittops
authored
7b5e74e
feat(memory): own the sampler working set, so deployment code stops reading configs

root
and
claude
committed
f8d41cb
fix(memory): the LoRA scratch is LINEAR in tokens, not convex

root
and
claude
committed
cfe036e

Commits on Jul 31, 2026

fix(memory): size the LoRA scratch from its mechanism, and by device

root
and
claude
committed
39fdab3
fix(memory): the LoRA prefill scratch is architecture-specific, so measure it

root
and
claude
committed
17f240b

Commits on Jul 30, 2026

fix(memory): a shard index states the weight total before the shards arrive

root
and
claude
committed
5b50faa
feat(memory): measure weights from the checkpoint, estimate only as fallback

root
and
claude
committed
5bedb33
fix(memory): experts own their down-projection unless a config key says otherwise

root
and
claude
committed
58250dc

Commits on Jul 29, 2026

test(memory): fix the note assertion the margin commit broke

root
and
claude
committed
6e299b2
fix(memory): add margin past the calibrated token ceiling

root
and
claude
committed
a4a1bb7

Commits on Jul 28, 2026

fix(memory): the LoRA token slope has a knee; only rank 64 is validated

root
and
claude
committed
6428c72
fix(memory): rank scaling is superlinear, so say which way the fit errs

root
and
claude
committed
698771e

Commits on Jul 27, 2026

feat(memory): size the LoRA prefill scratch, not just adapter storage

root
and
claude
committed
002a778

Commits on Jul 25, 2026

Merge pull request #5 from BudEcosystem/claude/kv-cache-sizing-fixes
dittops
authored
3a38a42

Commits on Jul 24, 2026

fix(memory): a disabled sliding window must not crash the per-layer KV path

dittops
and
claude
committed
3a22e5e
fix(memory): treat a null head_dim as absent in every reader, not just the KV ones

dittops
and
claude
committed
eed0588
fix(memory): detect interleaved attention through a multimodal text_config

dittops
and
claude
committed
cd991c1
fix(memory): correct KV-cache geometry, MLA, SSM state, and TP sharding

dittops
and
claude
committed
dcf17ae

Commits on Jul 13, 2026

Merge pull request #4 from BudEcosystem/claude/gpt-oss-120b-slo-check-97f11c
dittops
authored
3302a5f
fix(genz): address PR review — EP-shard MoE activations + defensive guards

root
and
claude
committed
d9852b8

Commits on Jul 12, 2026

fix(genz): correct MoE/PP inference and training memory/MFU models to match measured reality

root
and
claude
committed
22151dc

Commits on Jul 2, 2026

Merge pull request #3 from BudEcosystem/calib/cpu-memory-bound-decode
dittops
authored
044958d
docs(cpu): drop c10/int8kv from data; add per-user throughput columns

nived-bud
and
claude
committed
b5eb71a
docs(cpu): drop unfounded INT8 attribution from the +60% target CSV

nived-bud
and
claude
committed
600c1d2

Commits on Jul 1, 2026

docs(cpu): recommend TP=2 in best-config CSV (drop fp8-KV pick) + add +60% target

nived-bud
and
claude
committed
5283140
docs(cpu): add best-config-per-workload CSV (vLLM tuning result)

nived-bud
and
claude
committed
c1b2388
docs(cpu): add real-run CSVs (force-add; repo .gitignore blanket-ignores *.csv)

nived-bud
and
claude
committed
817ec68
docs(cpu): add raw vLLM benchmark data + repro scripts behind the calibration

nived-bud
and
claude
committed
c16144c
feat(cpu): memory-bound decode floor + NUMA all-reduce latency, calibrated to Xeon 6767P

nived-bud
and
claude
committed
527b186

Commits on Jun 28, 2026

Commits on Jun 27, 2026

feat(inference-engines): read decoder dims from text_config so the ORT activation generalizes to multimodal

jithinAB
and
claude
committed
be6cacd