Update dependencies, examples, and docs for GPT-OSS #6
Conversation
Signed-off-by: Mustafa Eyceoz <[email protected]>
Signed-off-by: Mustafa Eyceoz <[email protected]>
Signed-off-by: Mustafa Eyceoz <[email protected]>
WalkthroughUpdates documentation to mark OSFT as implemented, adds two GPT-OSS 20B example scripts (SFT and OSFT) with CLI and checkpoint handling, documents a new OSFT parameter (osft_memory_efficient_init), and revises dependencies in pyproject.toml including new/updated packages and optional group changes. Changes
Sequence Diagram(s)sequenceDiagram
autonumber
actor U as User
participant CLI as CLI Args
participant Script as SFT/OSFT Script
participant Hub as training_hub.sft/osft
participant Trainer as Trainer/Backend
participant Store as Checkpoint Storage
U->>CLI: Provide --data-path, --ckpt-output-dir, ...
CLI->>Script: Parse args
Script->>Script: Build params (model, batch, dist, OSFT/SFT)
Script->>Hub: Invoke sft()/osft()(params)
Hub->>Trainer: Initialize & run training
Trainer-->>Store: Write checkpoints
Script->>Script: find_most_recent_checkpoint(output_dir)
Store-->>Script: Latest checkpoint path
Script-->>U: Print duration and checkpoint
rect rgba(220,255,220,0.3)
Note right of Script: New: OSFT example script and checkpoint finder
Note right of Hub: New doc param osft_memory_efficient_init (OSFT)
end
Estimated code review effort🎯 3 (Moderate) | ⏱️ ~25 minutes Poem
Pre-merge checks and finishing touches❌ Failed checks (1 warning)
✅ Passed checks (2 passed)
✨ Finishing touches
🧪 Generate unit tests
Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out. Comment |
Signed-off-by: Mustafa Eyceoz <[email protected]>
RobotSail
left a comment
RobotSail
left a comment
There was a problem hiding this comment.
A few comments but otherwise LGTM
Co-authored-by: Oleg Silkin <[email protected]>
![coderabbitai[bot]](https://avatars.githubusercontent.com/in/347564?s=60&v=4){kind=small}
There was a problem hiding this comment.
Actionable comments posted: 1
🧹 Nitpick comments (15)
examples/docs/osft_usage.md (2)
145-173: Deduplicate/relocate parameter docs for clarity.
unfreeze_rank_ratioappears under Required Parameters and again under OSFT‑Specific Parameters. Keep it only under Required and move the range guidance there to avoid conflicting edits later.
269-273: Tighten the memory guidance wording.Clarify that memory‑efficient init helps at load time, not step time. Suggested tweak:
-2. **Memory Management**: OSFT doesn't reduce memory requirements compared to SFT, so adjust `max_tokens_per_gpu` accordingly. For memory-constrained environments or OOMs during model loading, set `osft_memory_efficient_init=True`. +2. **Memory Management**: OSFT does not reduce step‑time memory vs SFT. Tune `max_tokens_per_gpu` to avoid OOMs during training. For OOMs while loading the base model, enable `osft_memory_efficient_init=True` (affects load‑time memory only).pyproject.toml (3)
39-46: Duplicate dependency: filelock listed twice.Drop the duplicate to avoid resolver noise.
"filelock>=3.0", "psutil>=6.0", "urllib3>=2.4", "frozenlist>=1.7", "xxhash>=3.0", "requests>=2.32.5", "attr>=0.3.2", - "filelock>=3.19.1", "mpmath>=1.3.0", "pytest>=8.0"
47-49: Move pytest to dev extras (avoid shipping test runner to end users).- "mpmath>=1.3.0", - "pytest>=8.0" + "mpmath>=1.3.0" ] @@ dev = [ "ipykernel", - "ipython" + "ipython", + "pytest>=8.0" ]
14-17: Build tools in runtime deps.
wheel(and usuallysetuptools) shouldn’t be a runtime requirement given they’re already in [build-system]. Keeppackagingif used at runtime.dependencies = [ - "setuptools>=80.0", "packaging>=24.2", - "wheel>=0.43", + # wheel/setuptools are provided via [build-system]examples/scripts/sft_gpt_oss_example.py (5)
1-1: Shebang mismatch with file perms.Either make the file executable or drop the shebang to appease linters.
-#!/usr/bin/env python3
17-22: Remove unused imports.
osanddatetimearen’t used.-import os import sys import time -from datetime import datetime import argparse
61-61: Drop unusedresultassignment.- result = sft( + sft(
74-79: Avoid hardcoding/dev/shm; make it configurable with a safe default.@@ - parser.add_argument('--nproc-per-node', type=int, default=8, + parser.add_argument('--nproc-per-node', type=int, default=8, help='Number of GPUs (default: 8)') + parser.add_argument('--data-output-dir', default=None, + help='Optional processed data dir (default: system temp). Use /dev/shm on Linux for speed.') @@ - data_output_dir="/dev/shm", # Use RAM disk for speed + data_output_dir=(args.data_output_dir or os.getenv("TMPDIR") or "/tmp"),Add
import osback if you apply this change.
103-114: Catching broad Exception; prefer specific exceptions first.- except Exception as e: + except (ValueError, RuntimeError) as e: end_time = time.time() duration = end_time - start_time @@ - sys.exit(1) + sys.exit(1) + except Exception as e: + # Fallback catch-all to surface unexpected errors cleanly + print("=" * 50) + print(f"Unexpected error: {e}") + sys.exit(1)examples/scripts/osft_gpt_oss_example.py (5)
1-1: Shebang mismatch with file perms.Either make the file executable or drop the shebang.
-#!/usr/bin/env python3
22-22: Remove unused import.
datetimeis unused.-from datetime import datetime
28-52: Make checkpoint discovery more robust (mtime and dir check).-def find_most_recent_checkpoint(output_dir): +def find_most_recent_checkpoint(output_dir): @@ - checkpoint_dirs = glob.glob(checkpoint_pattern) + checkpoint_dirs = [p for p in glob.glob(checkpoint_pattern) if os.path.isdir(p)] @@ - # Find the most recently created checkpoint - most_recent_checkpoint = max(checkpoint_dirs, key=os.path.getctime) + # Pick most recently modified checkpoint (portable across filesystems) + most_recent_checkpoint = max(checkpoint_dirs, key=os.path.getmtime)
118-122: Avoid hardcoding/dev/shm; expose CLI override with safe default.@@ def main(): - parser.add_argument('--unmask-messages', action='store_true', default=False, + parser.add_argument('--unmask-messages', action='store_true', default=False, help='Unmask messages during training (default: False)') + parser.add_argument('--data-output-dir', default=None, + help='Optional processed data dir (default: system temp). Use /dev/shm on Linux for speed.') @@ - 'data_output_dir': "/dev/shm", # Use RAM disk for speed + 'data_output_dir': (args.data_output_dir or os.getenv("TMPDIR") or "/tmp"),Add
import osat top if removed.
159-171: Catching broad Exception; prefer specific exceptions first.- except Exception as e: + except (ValueError, RuntimeError) as e: end_time = time.time() duration = end_time - start_time @@ - sys.exit(1) + sys.exit(1) + except Exception as e: + print("=" * 50) + print(f"Unexpected error: {e}") + sys.exit(1)
📜 Review details
Configuration used: CodeRabbit UI
Review profile: CHILL
Plan: Pro
📒 Files selected for processing (5)
examples/README.md(2 hunks)examples/docs/osft_usage.md(3 hunks)examples/scripts/osft_gpt_oss_example.py(1 hunks)examples/scripts/sft_gpt_oss_example.py(1 hunks)pyproject.toml(2 hunks)
🧰 Additional context used
🧬 Code graph analysis (2)
examples/scripts/sft_gpt_oss_example.py (1)
src/training_hub/algorithms/sft.py (1)
sft(177-249)
examples/scripts/osft_gpt_oss_example.py (2)
src/training_hub/algorithms/osft.py (1)
osft(439-500)examples/scripts/sft_gpt_oss_example.py (1)
main(26-113)
🪛 Ruff (0.12.2)
examples/scripts/sft_gpt_oss_example.py
1-1: Shebang is present but file is not executable
(EXE001)
61-61: Local variable result is assigned to but never used
Remove assignment to unused variable result
(F841)
75-75: Probable insecure usage of temporary file or directory: "/dev/shm"
(S108)
103-103: Do not catch blind exception: Exception
(BLE001)
examples/scripts/osft_gpt_oss_example.py
1-1: Shebang is present but file is not executable
(EXE001)
46-46: Avoid specifying long messages outside the exception class
(TRY003)
119-119: Probable insecure usage of temporary file or directory: "/dev/shm"
(S108)
159-159: Do not catch blind exception: Exception
(BLE001)
🔇 Additional comments (8)
examples/README.md (3)
27-27: Formatting-only change.No action needed.
29-30: New SFT GPT‑OSS entry looks good.Path and description align with the added script.
62-63: New OSFT GPT‑OSS entry looks good.Consistent with the OSFT example script name.
pyproject.toml (4)
32-32: No action required — PyPI releases exist for these min versions. Verified fsspec (2025.9.0) and regex (2025.9.1) exist on PyPI, so keepingfsspec>=2025.0andregex>=2025.0in pyproject.toml is valid.
45-45: Confirm package name:attrvsattrs.pyproject.toml lists
attr>=0.3.2(pyproject.toml: line 45). Repo search returned no imports ofattr/attrs; verify the intended package and update pyproject.toml or the code to match.
21-21: Confirm instructlab-training (>=0.12.0) doesn't require a narrower transformers range.Wheel metadata shows instructlab-training depends on transformers>=4.41.2, and upstream issues report breakage with transformers >4.51.0 — verify 0.12.0's Requires‑Dist/pyproject so your top-level transformers pin won't cause resolver conflicts. (wheelodex.org)
55-61: Confirm flash‑attn / PyTorch / CUDA compatibilitypyproject.toml extras ('cuda') include torch>=2.6.0 and flash-attn>=2.8 — confirm prebuilt flash-attn wheels exist for the exact PyTorch+CUDA you target (many flash-attn releases on PyPI are source-only and require ninja/CUDA to build; prebuilt wheels are provided separately, e.g. flash-attn-wheels). (pypi.org)
Location: pyproject.toml lines 54–61.
examples/docs/osft_usage.md (1)
90-96: Confirm backend support for osft_memory_efficient_initDocs/examples add osft_memory_efficient_init, but a repo search shows it only appears in examples/docs and examples/scripts; MiniTrainerOSFTBackend.execute_training’s renames dict (src/training_hub/algorithms/osft.py) does not include this key and there’s no other handling — verify the backend accepts this kwarg (or add a rename/TrainingArgs field and/or validation entry) to avoid runtime "unexpected keyword" or validation errors.
| "rhai-innovation-mini-trainer>=0.2.0", | ||
| "torch>=2.6.0", | ||
| "transformers>=4.55.0", | ||
| "numba>=0.50", |
There was a problem hiding this comment.
🧩 Analysis chain
Numba/llvmlite/numpy compatibility risk.
The trio numba>=0.50, llvmlite>=0.42, and numpy<2.3 is likely to produce resolver or ABI mismatches. Prefer matching Numba’s documented pairings (Numba pins llvmlite tightly and constrains NumPy).
- "numba>=0.50",
- "llvmlite>=0.42",
- "numpy>=1.26.4,<2.3",
+ # Align versions to avoid ABI conflicts (NumPy 2+ support)
+ "numba>=0.60,<0.61",
+ "llvmlite>=0.43,<0.44",
+ "numpy>=1.26.4,<2.3",Please validate against your CI matrix and the GPUs you target.
Also applies to: 38-39
🏁 Script executed:
#!/bin/bash
set -euo pipefail
echo "=== locate pyproject.toml ==="
fd -HI --hidden pyproject.toml || true
for f in $(fd -HI --hidden pyproject.toml || true); do
echo "---- $f ----"
nl -ba "$f" | sed -n '1,240p'
done
echo
echo "=== Search for numba/llvmlite/numpy occurrences across repo ==="
rg -n --hidden -S -C2 '\b(numba|llvmlite|numpy)\b' || trueLength of output: 110
🌐 Web query:
Numba llvmlite compatibility matrix
Numba supported NumPy versions site:numba.pydata.org
Numba 0.60 release notes llvmlite
llvmlite 0.43 release notes
💡 Result:
Here are the sources you requested and the key compatibility facts:
- llvmlite v0.43.0 release notes (includes date and highlights). [1] (llvmlite.readthedocs.io)
- llvmlite compatibility / historical matrix (README) — shows which LLVM versions each llvmlite series supports. [2] (github.com)
- Numba 0.60.0 release notes (highlights) — Numba 0.60.0 released 2024-06-13 and documents NumPy/llvmlite compatibility. See the release notes for details. [3] (numba.readthedocs.io)
- Numba 0.60 installation / version-support table (canonical compatibility matrix): Numba 0.60.0 — Python 3.9–<3.13, NumPy 1.22 ≤ version < 1.27 (and == 2.0), llvmlite 0.43.x, LLVM 14.x. [4] (numba.readthedocs.io)
- For archived Numba docs showing supported NumPy ranges for earlier Numba releases (examples: 0.48, 0.43, etc.). Use these pages if you need per-release NumPy support details. [5][6] (numba.pydata.org)
If you want, I can extract the full compatibility table (Numba release → supported Python, NumPy, llvmlite, LLVM) from the 0.60 docs into a CSV or formatted table.
Citations:
- 1: https://llvmlite.readthedocs.io/en/v0.43.0/release-notes.html?utm_source=openai
- 2: https://github.com/numba/llvmlite/blob/main/README.rst?utm_source=openai
- 3: https://numba.readthedocs.io/en/latest/release/0.60.0-notes.html?utm_source=openai
- 4: https://numba.readthedocs.io/en/0.60.0/user/installing.html?utm_source=openai
- 5: https://numba.pydata.org/numba-doc/0.48.0/user/installing.html?utm_source=openai
🏁 Script executed:
#!/bin/bash
set -euo pipefail
echo "=== locate pyproject.toml ==="
fd -HI --hidden pyproject.toml || true
for f in $(fd -HI --hidden pyproject.toml || true); do
echo "---- $f ----"
sed -n '1,240p' "$f" || true
done
echo
echo "=== Search for numba/llvmlite/numpy occurrences across repo ==="
rg -n --hidden -S -C2 '\b(numba|llvmlite|numpy)\b' || true
echo
echo "=== Search for CI/workflows that reference numba/llvmlite/numpy or GPUs ==="
rg -n --hidden -S --glob '.github/**' 'numba|llvmlite|numpy|cuda|gpu|nvidia|rocm' || true
echo
echo "=== List .github/workflows files (first 240 lines) ==="
fd -HI --hidden .github/workflows || true
for f in $(fd -HI --hidden .github/workflows || true); do
echo "---- $f ----"
sed -n '1,240p' "$f" || true
doneLength of output: 3531
Align Numba/llvmlite with supported NumPy to avoid ABI/resolver conflicts.
pyproject.toml currently has "numba>=0.50" (line 21) and "llvmlite>=0.42" (line 38) while NumPy is ">=1.26.4,<2.3". Numba releases tightly pair with specific llvmlite and NumPy ranges (e.g., Numba 0.60 ↔ llvmlite 0.43); loose/mismatched ranges risk resolver or ABI failures. Pin numba/llvmlite to a tested pairing or adjust NumPy bounds and validate on CI + target GPUs.
- "numba>=0.50",
- "llvmlite>=0.42",
- "numpy>=1.26.4,<2.3",
+ # Align versions to avoid ABI conflicts (NumPy 2+ support)
+ "numba>=0.60,<0.61",
+ "llvmlite>=0.43,<0.44",
+ "numpy>=1.26.4,<2.3",pyproject.toml: lines 21, 38.
📝 Committable suggestion
‼️ IMPORTANT
Carefully review the code before committing. Ensure that it accurately replaces the highlighted code, contains no missing lines, and has no issues with indentation. Thoroughly test & benchmark the code to ensure it meets the requirements.
| "numba>=0.50", | |
| # Align versions to avoid ABI conflicts (NumPy 2+ support) | |
| "numba>=0.60,<0.61", | |
| "llvmlite>=0.43,<0.44", | |
| "numpy>=1.26.4,<2.3", |
🤖 Prompt for AI Agents
pyproject.toml lines 21 and 38: the numba and llvmlite ranges are too loose
compared to the pinned NumPy range and can cause ABI/resolver conflicts; update
these pins to a tested pairing (for example pin numba to the release that
matches your llvmlite and NumPy matrix, e.g. numba 0.60.x with llvmlite 0.43.x)
or alternatively relax/expand the NumPy bounds to match the chosen
numba/llvmlite combo; change the version specifiers to explicit, compatible
ranges (or exact patch pins) and run CI matrix tests on target Python/NumPy/GPU
combos to validate.
There was a problem hiding this comment.
Actionable comments posted: 0
♻️ Duplicate comments (2)
examples/scripts/osft_gpt_oss_example.py (1)
41-51: Checkpoint discovery: prefer mtime or numeric sorting; keep glob fixGood that the pattern is now
samples_*(matches prior review). Usingctimecan be unreliable across filesystems; usemtimeor parse the numeric suffix.- most_recent_checkpoint = max(checkpoint_dirs, key=os.path.getctime) + # Prefer modification time; alternatively parse the numeric suffix after 'samples_' + most_recent_checkpoint = max(checkpoint_dirs, key=os.path.getmtime)pyproject.toml (1)
48-48: Do not ship pytest as a runtime dependencyMove
pytest>=8.0to a dev/test extra. Shipping it increases install size and build time for end users.- "pytest>=8.0" + # moved to optional dev/test dependenciesThen add under
[project.optional-dependencies]:dev = [ "ipykernel", - "ipython" + "ipython", + "pytest>=8.0" ]
🧹 Nitpick comments (14)
examples/README.md (2)
29-29: Add a short note about CUDA extras next to GPT-OSS 20B scriptReaders may miss they need
pip install training-hub[cuda]for these scripts.Apply this minimal addition after the line:
- [SFT with GPT-OSS 20B](scripts/sft_gpt_oss_example.py) - Single-node multi-GPU training example with GPT-OSS 20B + [SFT with GPT-OSS 20B](scripts/sft_gpt_oss_example.py) - Single-node multi-GPU training example with GPT-OSS 20B (requires `pip install training-hub[cuda]`)
62-62: Mirror the CUDA extras hint for OSFT scriptSame rationale as SFT entry.
- [OSFT with GPT-OSS 20B](scripts/osft_gpt_oss_example.py) - Single-node multi-GPU training example with GPT-OSS 20B + [OSFT with GPT-OSS 20B](scripts/osft_gpt_oss_example.py) - Single-node multi-GPU training example with GPT-OSS 20B (requires `pip install training-hub[cuda]`)examples/docs/osft_usage.md (3)
172-173: Parameter reference: clarify default and compatibilityAdd “default: False” and, if applicable, note any backend/version requirements (e.g., requires mini-trainer ≥X.Y).
- - `osft_memory_efficient_init` (bool): Enable memory-efficient initialization to reduce memory usage during model loading (recommended for OOMs) + - `osft_memory_efficient_init` (bool, default: False): Enable memory‑efficient initialization to reduce memory usage during model loading (recommended if you hit OOMs). Requires backend support.
271-271: Best Practices: small clarity tweakCall out explicitly that it affects load-time memory only; training-time memory still governed by
max_tokens_per_gpu.-2. **Memory Management**: OSFT doesn't reduce memory requirements compared to SFT, so adjust `max_tokens_per_gpu` accordingly. For memory-constrained environments or OOMs during model loading, set `osft_memory_efficient_init=True`. +2. **Memory Management**: OSFT doesn't reduce training‑time memory vs SFT, so adjust `max_tokens_per_gpu` accordingly. For OOMs during model loading specifically, set `osft_memory_efficient_init=True`.
93-95: Document default and scope for osft_memory_efficient_initosft_memory_efficient_init is referenced only in docs/examples and is not declared on the public osft() API — document that it defaults to False and that it’s forwarded to backends via **kwargs unless you make it an explicit osft(...) parameter.
Occurrences: examples/docs/osft_usage.md (lines 93, 172–173, 271–272); examples/scripts/osft_gpt_oss_example.py (line 125). Public API: src/training_hub/algorithms/osft.py (def osft around lines 434–444) — parameter not present.examples/scripts/sft_gpt_oss_example.py (4)
1-1: Shebang present but file not executableEither make the file executable in the repo (
chmod +x) or drop the shebang.-#!/usr/bin/env python3 +# (shebang removed; invoke with `python sft_gpt_oss_example.py ...`)
17-22: Remove unused imports
osanddatetimeare unused.-import os import sys import time -from datetime import datetime import argparse
61-90: Avoid assigning unusedresult; narrow exception or re-raise with traceback
- Drop the unused assignment.
- Consider catching expected exception types or re-raising after logging.
- result = sft( + sft( @@ - except Exception as e: + except Exception as e: # consider narrowing to expected exceptions end_time = time.time() duration = end_time - start_time
75-79: Hard-coded/dev/shmcan be brittleExpose as a CLI flag with a safe default (e.g., omit or use
tempfile.gettempdir()).- data_output_dir="/dev/shm", # Use RAM disk for speed + # Consider exposing via --data-output-dir; default None lets backend decide + # data_output_dir="/dev/shm",examples/scripts/osft_gpt_oss_example.py (4)
1-1: Shebang present but file not executableSame as SFT script: make executable or remove shebang.
-#!/usr/bin/env python3 +# (shebang removed; invoke with `python osft_gpt_oss_example.py ...`)
19-25: Remove unused import
datetimeis unused.-from datetime import datetime import argparse import glob
119-121: Hard-coded/dev/shmcan be brittleSame suggestion as SFT script: make it configurable or omit.
- 'data_output_dir': "/dev/shm", # Use RAM disk for speed + # 'data_output_dir': "/dev/shm", # Consider exposing via CLI; default None is safer
159-171: Catching broad ExceptionConsider narrowing to expected exceptions and/or printing a traceback before exiting.
- except Exception as e: + except Exception as e: # consider narrowing to expected exceptions end_time = time.time() duration = end_time - start_timepyproject.toml (1)
40-47: Duplicate dependency: filelock listed twiceKeep the stricter one (
>=3.19.1) and drop the other to avoid confusion.- "filelock>=3.0", @@ - "filelock>=3.19.1", + "filelock>=3.19.1",
📜 Review details
Configuration used: CodeRabbit UI
Review profile: CHILL
Plan: Pro
📒 Files selected for processing (6)
README.md(1 hunks)examples/README.md(2 hunks)examples/docs/osft_usage.md(3 hunks)examples/scripts/osft_gpt_oss_example.py(1 hunks)examples/scripts/sft_gpt_oss_example.py(1 hunks)pyproject.toml(2 hunks)
🧰 Additional context used
🧬 Code graph analysis (2)
examples/scripts/osft_gpt_oss_example.py (2)
src/training_hub/algorithms/osft.py (1)
osft(439-500)examples/scripts/sft_gpt_oss_example.py (1)
main(26-113)
examples/scripts/sft_gpt_oss_example.py (2)
src/training_hub/algorithms/sft.py (1)
sft(177-249)examples/scripts/osft_gpt_oss_example.py (1)
main(54-171)
🪛 Ruff (0.12.2)
examples/scripts/osft_gpt_oss_example.py
1-1: Shebang is present but file is not executable
(EXE001)
46-46: Avoid specifying long messages outside the exception class
(TRY003)
119-119: Probable insecure usage of temporary file or directory: "/dev/shm"
(S108)
159-159: Do not catch blind exception: Exception
(BLE001)
examples/scripts/sft_gpt_oss_example.py
1-1: Shebang is present but file is not executable
(EXE001)
61-61: Local variable result is assigned to but never used
Remove assignment to unused variable result
(F841)
75-75: Probable insecure usage of temporary file or directory: "/dev/shm"
(S108)
103-103: Do not catch blind exception: Exception
(BLE001)
🔇 Additional comments (7)
README.md (1)
9-9: Support Matrix: OSFT marked Implemented — looks goodStatus change aligns with the new examples and docs. No further action from this file.
examples/README.md (1)
27-27: Formatting-only tweak — OKNo behavioral impact.
examples/scripts/sft_gpt_oss_example.py (1)
80-82: Checkpointing choice: confirm behavior matches backendIf
accelerate_full_state_at_epoch=False, resume behavior changes. Confirm that’s intended for users copying this example.examples/scripts/osft_gpt_oss_example.py (1)
63-78: Confirm HF model idEnsure
--model-pathdefaultopenai/gpt-oss-20bmatches the actual HF repo name.pyproject.toml (3)
54-61: Extras naming: confirm docs align withcudaextraThe README installs
training-hub[cuda], which matches this extra. If any docs mention agpt-ossextra, update them for consistency.
45-46: Verifyattrvsattrsin pyproject.tomlpyproject.toml lists "attr>=0.3.2" (pyproject.toml: lines 45–46). The canonical package is usually "attrs" — confirm which package you intend to require and update pyproject.toml if necessary. Repository search returned no import matches for either; manual verification required.
17-23: Version alignment sanity check for core training stackLarge version bumps in pyproject.toml (instructlab-training, rhai-innovation-mini-trainer, transformers, datasets, numpy upper bound) can conflict with torch/flash-attn/bitsandbytes — ensure CI uses a pinned lockfile or constraints file and verify resolver stability across Linux/macOS/GPU images.
Automated search returned no output; run locally/CI to confirm:
- rg -n -C2 -S 'instructlab-training|rhai-innovation-mini-trainer|transformers|torch|flash-attn|bitsandbytes|numpy|accelerate|peft' pyproject.toml
- ls | rg 'poetry.lock|requirements.txt|constraints.txt|Pipfile.lock' || true
- python -m pip install -U pip && python -m pip install 'training-hub[cuda] @ .'
Adds two new example scripts, updated documentation, and reorganized dependencies for GPT-OSS support
Summary by CodeRabbit
New Features
Documentation
Chores