Update tutorials off models slated for retirement on 2026-06-12#732
Open
nealwu wants to merge 16 commits into
Open
Update tutorials off models slated for retirement on 2026-06-12#732nealwu wants to merge 16 commits into
nealwu wants to merge 16 commits into
Conversation
Swaps each tutorial's model (and any hardcoded paired renderer) to the recommended replacement from the deprecation page: - Qwen3-4B-Instruct-2507 -> Qwen3.5-4B (qwen3_5_disable_thinking) - Llama-3.x base -> Qwen3.5-9B-Base (role_colon) - Qwen3-30B-A3B -> Qwen3.6-35B-A3B (qwen3_5) - Qwen3-VL-235B-A22B-Instruct -> Qwen3.5-397B-A17B (vision-capable) - Kimi-K2.5 -> Kimi-K2.6 (kimi_k26_disable_thinking) Also refreshes the renderer table and supported-models prose in 201 and 102 so they don't reference retired families. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
derek-tml
approved these changes
May 22, 2026
Qwen3.5-4B is a chat-tuned thinking model; feeding it raw tokens (the intentional pattern in 101) drops into thinking mode and produces garbled output. Switch to Qwen3.5-9B-Base -- base models are the natural match for the raw tokenizer.encode -> sample API the tutorial demonstrates. Drop the "\n" stop arg and tune temperatures so the completion reads naturally. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Remove the 'name=' argument from save_weights_and_get_sampling_client_async()
calls in tutorials 102 and 302 -- the parameter is now ignored and warns
in newer Tinker SDK versions.
Add a 'tutorials' optional-deps extra ('marimo', 'matplotlib') so users
can install everything needed for the tutorials with
'uv pip install "tinker-cookbook[tutorials]"' instead of remembering each
package by hand. Update tutorials/README.md to use the new install line.
Includes marimo auto-formatting noise from a 0.21.1 -> 0.23.8 session
(expanded function signatures, reordered cell-return tuples) and a small
prose addition to 102 noting that the Kimi K2.6 stage is slower than the
Qwen3.5 stage.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
The narration described the synchronous sample() + .result() futures API, but every code cell uses await sample_async() + asyncio.gather. Rewrite the prose (intro, sequential/concurrent sections, batch-eval, summary, and the section header) to describe the actual mechanism: asyncio.gather schedules the sample_async() coroutines onto the event loop so all requests are in flight before any completes. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Metadata-only change from running the notebook locally; brings 104 in line with the other tutorials already at 0.23.8. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
- Fix parse_response demo: the hardcoded fake_tokens were written for the old Qwen3-30B-A3B vocab, where 151645 was <|im_end|>. After the tokenizer swap that id decodes to unrelated text and the parse returns malformed. Build the tokens by encoding a plausible reply + the stop token instead. - Vision section: stay on Qwen3.6-35B-A3B and use the same qwen3_5 renderer with an image processor, rather than switching models and framing VL as a separate renderer (Qwen3.5/3.6 are natively vision-capable). - Renderer-switching cell now actually switches: loops over Qwen3.6 and Kimi K2.6 so the "each model family needs its own tokenizer" point is real. - Note that qwen3_5 is a thinking renderer (open <think> tag) and document the qwen3_5_disable_thinking alternative. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
loss_fn_outputs[...]["logprobs"] is a TensorData wrapper, not a plain list, so slicing it (ce_logprobs[-3:]) raised "TensorData object is not subscriptable". Convert with .tolist() before slicing. Also aligns the loss:sum print columns; remainder is marimo formatting churn (version bump, trimmed cell-return tuples). Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
asyncio.run() can't be called inside marimo's running event loop. Switch the five completer calls to top-level await in async cells, and drop the now-unused asyncio / MessageCompleter / TokenCompleter / TokensWithLogprobs imports. Remainder is marimo serialization churn (version bump, trimmed cell-return tuples). Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
The triple-quoted judge prompt picked up 4-space indentation from marimo serialization, which became part of the string sent to the model. Restore the original flush-left text. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
- save_weights_for_sampler / save_state: drop _async (those return the
future un-awaited; the sync method awaited once is the SDK's idiom).
- RestClient list/TTL/publish cells: .result() -> await *_async in async
cells (sync-from-async warning / deadlock risk in marimo's loop).
- list_checkpoints: pass the full model_id, not model_id.split(":")[0]
(the endpoint wants a ModelID; truncating caused 404 Model not found).
- weights.download: run via asyncio.to_thread so the sync helper's
internal .result() doesn't block marimo's event loop.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
- NLLEvaluator: output["logprobs"] and datum.loss_fn_inputs["weights"] are TensorData, so torch.tensor(...) raised "Could not infer dtype". Use TensorData.to_torch() instead. - Fix dead Inspect AI link: https://inspect.ai -> https://inspect.aisi.org.uk/ (the inspect-ai package's actual docs site, by the UK AI Security Institute). Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Member
Author
|
Going to split this up into multiple PRs for simplicity, one for each section of the tutorials (100s, 200s, etc.) |
This was referenced Jun 1, 2026
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Swaps each tutorial's model (and any hardcoded paired renderer) to the recommended replacement from the deprecation page:
Also refreshes the renderer table and supported-models prose in 201 and 102 so they don't reference retired families.