Name the setup steps as the updated Image Classification lab does, call
the environment LiteRT, and drop a leftover image filename that rendered
as a caption under the training output.
Removing /usr/lib/python3.11/EXTERNALLY-MANAGED does not apply on
Trixie (Python 3.13) and contradicts the venv rule stated right after
it. Also rename the sampling variable in avg_temp_power.sh from
SECONDS, which bash reserves and increments automatically, to SAMPLES.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Ultralytics prints the plural of the custom class 'box' as 'boxs'
('3 boxs, 4 wheels'); the lab quotes that output verbatim.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
'/set nothink' is an Ollama session command that turns off the
thinking mode; it is not a misspelling of 'nothing'.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Sync with the author's EdgeML Made Easy book: ai-edge-litert replaces
tflite_runtime, venv with system site packages, Picamera2 test, and
recaptured Edge Impulse figures.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Sync with the author's EdgeML Made Easy book: Trixie and Imager 2.0,
rpi-swap and zram (with Zero 2 W notes), rpicam-hello, and a new
temperature and power measurement section.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Sync with the author's EdgeML Made Easy book: LiteRT for SSD and the
live-stream app, expanded IoU/mAP, and a rewritten YOLO section with
YOLO11 and YOLO26, Trixie installation, NCNN/ONNX export measurements,
and the imgsz rule for exported models.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Sync with the author's EdgeML Made Easy book: Gemma 3/3n/4 replace
Gemma 2, MoonDream, Ollama commands and model inspection, an expanded
Python library section, direct REST API, and function calling with
Ollama's native tools instead of instructor/OpenAI.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
labs/tools/check_browser_wheels.py requires the committed mlsysim wheel to match the runtime source. Replace mlsysim/hardware/data/cloud/T4.yaml inside mlsysim-0.3.1-py3-none-any.whl with the updated file and its RECORD entry; every other entry is unchanged, and all 508 runtime files now match the source.
The T4 entry stored only an int8 rate, so Engine.solve with precision='fp32' fell through to peak_flops, the 65 TFLOP/s FP16 tensor-core rate, and warned "Results may be inaccurate". FP32 work on the T4 was modelled about 8x too fast, and INT4 4x too slow.
Add the datasheet rates: FP32 8.1 TFLOP/s on the CUDA cores and INT4 260 TOPS. peak_flops (FP16 tensor, 65) and int8 (130) are unchanged, so FP16 and INT8 results, the appendix table and the hardware chapter's ridge point stay the same.
test_precision_keys_use_canonical_vocabulary catches an aliased key such as 'fp32_cuda', but not a chip with no fp32 key at all, which is how the T4 slipped through. Add a T4 regression test and an Engine check that a compute-bound FP32 solve is limited by the FP32 rate without the fallback warning.
Option B from #2155: the vault schema is unchanged, and each exercise lives
under staffml/exercises/<exercise>/ with README.md (task and napkin math),
starter.py, solution.py and a test file. The tests grade the reference
solution by default, or the starter with STAFFML_IMPL=starter.
- kv-cache-decode: incremental single-head attention with a pre-allocated
KV cache. Tests check the outputs against full recompute to 1e-5, that
each token is projected once (3T rows, not 3T(T+1)/2), that the cache is
written in place, and the KV-cache size of Llama-2-7B.
- per-channel-int8: per-channel weight scales, a per-tensor activation
scale, and an INT8 x INT8 -> INT32 matmul with one rescale per channel.
Tests check dtypes and scale shapes, the half-step round-trip error, the
analytic matmul error bound, per-channel vs per-tensor accuracy, and
INT32 headroom.
NumPy and pytest only, CPU only. 13 tests pass on the solutions.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01SxFzWDUc3UeWQ1QBGjw6UM
The three drafts reviewed in #2156, with topic: mixture-of-experts on the
mobile track (status: draft, provenance: llm-draft):
- mobile-202611-77dc (L4, memory): sizing a phone for Qwen3-30B-A3B by
active vs total parameters.
- mobile-202611-1e6f (L5, memory): expert offloading to UFS 4.0 for
Mixtral-8x7B at 0% and 90% expert-cache hit rates.
- mobile-202611-2872 (L5, latency): why prefill touches every expert in an
offloaded MoE, and the time-to-first-token that follows.
IDs allocated with vault new; vault check --strict passes on the full
corpus.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01SxFzWDUc3UeWQ1QBGjw6UM
- Add configurable PopClip-style selection bar (Explain this, Ask, Check me, Taxonomy, Simplify, Example, Study tools) with gear menu and Learning preferences sync
- Add security-hardened BYOK and 1-Click OpenRouter OAuth PKCE SSO module (src_shared/byok-provider.js) with strict HTTPS host allowlist, secret redaction, session-only default storage, and shared .github.io origin protection
- Add Free shared model status bar and Connect your own Gemini / Claude controls in both Focused Reader and Study Tools
- Extract interactive figure explorer ([data-mlsys-explore]) control parameters in reader and study context extractors
- Update SocratiQ user guide prose and 2x HiDPI screenshots and add unit and Playwright regression coverage
- Wrap secondary live meters, calculation tables, and step-math workspaces inside the collapsible 'Inspect the calculation' drawer by default across original-player.js and convolution.js so every figure presents a uniform 4-part layout:
1. Guiding Question + 1-sentence Intro + compact [Assumptions] button
2. Connected 3-Tier Player Frame (Canvas -> Tinted Controls Bar -> Flush Blue Step Narration)
3. Compact [Inspect the calculation] dropdown-button box at the bottom
- Refresh runtime_sha1 in figure-registry.json for all 111 figures.
- Style closed Assumptions and Inspect-the-calculation disclosures as compact fit-content dropdown buttons that expand into full-width bordered cards when opened.
- Unify the Stage Canvas, tinted Transport Controls Dock, and flush Step Narration into a single connected 3-tier player frame.
- Replace the outer 4-sided box on .mlsys-original-details with a top hairline divider to eliminate nested box-in-box clutter.