[nasser1941]

Builds on #260 and merges after it. This branch has #260's commit, then main merged in, then its own two commits. Until #260 merges, the diff below includes #260's changes too; this PR's own are 784f8dd (CHANGELOG placement after merging main) and f3395e9 (the change).

What changes

lcode bench read /api/ps directly. With Ollama 0.40, lcode-qwen3.6-35b at 32K on a 12 GB GPU printed Loaded in 18s · 10.4 GB, 98% on the GPU, because Ollama leaves out the ~12 GB of experts that stay mapped from the model's file in RAM. The summary and the --markdown table said the same.

It now uses residency.placement (which matches lcode-qwen3.6-35b:latest) and #260's residency.with_estimate, so it says what the line after a load says:

  Loaded in 18s · ~23 GB, 42% on the GPU and the rest in system RAM · reads prompts at 905 tok/s

and ~23 GB, 42% on GPU in the summary's Memory row. When Ollama counts the whole model (all of it on the GPU, or a model outside the catalog), both show Ollama's numbers as before, in the load line's words: 6.6 GB, all on the GPU.

The JSON

--json results keep Ollama's count under the names it always had: memory_gb (10.4) and gpu_percent (98). docs/bench.md promises that existing fields keep their meaning until schema changes, and older results only have Ollama's count, so comparing those two fields across results stays like for like. lcode's estimate goes next to them:

Field	Example	Meaning
estimated_memory_gib	22.7	The whole model by the catalog's estimate, in GiB, as lcode setup shows it (~23 GB)
estimated_gpu_percent	42	The share of that on the GPU

Both are null when Ollama counts the whole model. lcode bench --compare reads only --harbor results, which record no memory, so nothing changes there.

Docs and CHANGELOG

• docs/bench.md: the example's Loaded line, What's measured, the 12 GB laptop table (with a note that 0.58.0 printed Ollama's 10.5 GB, 98% on GPU), the JSON example and the field table.
• docs/models.md: names lcode bench next to the load line and lcode doctor.
• CHANGELOG.md: a Fixed entry under the new Unreleased.

Merging main

Main released 0.62.0 after #260 branched, which renamed the old Unreleased section. Merging main filed #260's two CHANGELOG entries under [0.62.0] without a conflict. 784f8dd puts them back under Unreleased, where this PR's entry follows them. After #260 merges, merging main here again will likely conflict in Unreleased, since both sides add #260's entries there: keep main's and this PR's entry after them.

Testing

• uv run pytest: 1156 passed, 2 skipped
• uv run ruff check . && uv run ruff format --check .: clean
• uv run basedpyright: 0 errors
• uv run --group docs mkdocs build --strict: builds
• New tests in tests/test_bench.py, with FakeOllama: Ollama 0.40's numbers for lcode-qwen3.6-35b:latest (the line, the JSON and the Memory row), a model all on the GPU and one outside the catalog (Ollama's count stands), and a model measured after its tasks because the load failed. All three fail on the old bench.py.
• Not run against a real Ollama: no model was loaded on the GPU, because of the gate nights.

🤖 Generated with Claude Code