Add Qwen 3.5 (4B/9B, Base/Instruct) to supervised_finetuning recipes

[dgallitelli]

Add Qwen 3.5 (4B / 9B, Base / Instruct) to supervised_finetuning recipes

Summary

Adds four notebooks and eight recipe YAMLs to 0_model_customization_recipes/supervised_finetuning/ for fine-tuning Qwen 3.5 — both Base and post-trained ("Instruct") variants in 4B and 9B sizes. Each notebook supports both QLoRA (4-bit) and full fine-tuning, selectable via a strategy toggle.

Notebook	Variants covered	QLoRA default	Full default
finetune--Qwen--Qwen3.5-4B-Base.ipynb	Pretrained	ml.g5.2xlarge (1× A10G 24 GB)	ml.g7e.2xlarge (1× RTX PRO 6000 96 GB)
finetune--Qwen--Qwen3.5-4B.ipynb	Instruct (post-trained)	ml.g5.2xlarge	ml.g7e.2xlarge
finetune--Qwen--Qwen3.5-9B-Base.ipynb	Pretrained	ml.g5.2xlarge	ml.g7e.12xlarge (4× RTX PRO 6000)
finetune--Qwen--Qwen3.5-9B.ipynb	Instruct	ml.g5.2xlarge	ml.g7e.12xlarge

Naming note: On HuggingFace, post-trained variants are published as Qwen/Qwen3.5-{4B,9B} with no -Instruct suffix — the -Base suffix denotes the pretrained checkpoint. Both share the same qwen3_5 architecture, so the same DLC and dependency pins apply; only weights and chat template differ.

What this PR does NOT change

Zero source code modifications. No edits to sagemaker_code/sft.py, sagemaker_code/utils/merge_adapter_weights.py, or sm_accelerate_train.sh. Qwen 3.5 works with the existing shared scaffolding as-is.

What this PR DOES change

• 4 new finetune--Qwen--Qwen3.5-*.ipynb notebooks
• 8 new recipe YAMLs under sagemaker_code/hf_recipes/Qwen/

Each notebook overrides sagemaker_code/requirements.txt via a %%writefile cell at job-submit time (with .bak backup, restored by the final cell). The override:

Package	Shared default	Required	Why
transformers	4.57.0	5.2.0	qwen3_5 architecture not in 4.x
peft	0.17.0	0.18.1	HybridCache removed in transformers 5.x; 0.17 hardcodes the import
bitsandbytes	0.46.1	0.49.2	First version with a CUDA 13.0 binary (DLC ships CUDA 13)
liger-kernel	0.6.1	0.7.0	Same HybridCache compatibility issue as peft

trl == 0.21.0 and the rest of the toolchain are unchanged — the lockstep TRL 1.x bump that Gemma 4 needs is not required here, so existing sibling recipes are unaffected.

Validation

This PR's smoke tests (us-east-1, 2026-05-22)

Two SageMaker training jobs run from a clean upstream checkout with only this PR's notebooks + YAMLs + the %%writefile-applied requirements.txt:

#	Notebook	Strategy	Instance	Status	Billable	Loss
1	finetune--Qwen--Qwen3.5-4B.ipynb	QLoRA	ml.g5.2xlarge	Completed	688 s	1.66 → 1.41
2	finetune--Qwen--Qwen3.5-4B-Base.ipynb	QLoRA	ml.g5.2xlarge	Completed	471 s	1.57 → 1.41

Both jobs trained 20 steps over 100 rows of Josephgflowers/Finance-Instruct-500k, saved a PEFT adapter, and ran the upstream merge_adapter_weights.py to completion (the merge step works for Qwen 3.5 — no Gemma-4-style skip needed).

Reference-repo validation matrix

All eight recipe × instance combinations have additionally been validated end-to-end with real SageMaker training jobs in the reference repo. Highlights:

#	Variant	Strategy	Instance	GPU(s)	Billable
T1	4B Instruct	QLoRA	ml.g5.2xlarge	1× A10G 24 GB	~21 min
T2	4B Base	Full SFT	ml.g7e.2xlarge	1× RTX PRO 6000 96 GB	~29 min
T3	4B Instruct	Full SFT	ml.g7e.2xlarge	1× RTX PRO 6000 96 GB	~30 min
T4	9B Base	Full SFT	ml.g7e.12xlarge	4× RTX PRO 6000 (384 GB total)	~49 min
T5	9B Instruct	Full SFT	ml.g7e.12xlarge	4× RTX PRO 6000 (384 GB total)	~46 min
T6	9B Instruct	QLoRA	ml.g5.2xlarge	1× A10G 24 GB	~28 min
T7	9B Instruct	QLoRA	ml.g6e.2xlarge	1× L40S 48 GB	~22 min

QLoRA recipes also validated portable across ml.g5.2xlarge / ml.g6.4xlarge / ml.g7e.2xlarge without any recipe changes.

Rationale

Qwen 3.5 has been generally available on Hugging Face for several months and is a strong general-purpose model that customers ask about regularly. The qwen3_5 architecture isn't in transformers 4.57 (the shared requirements.txt pin), and that single bump cascades to peft / bitsandbytes / liger-kernel — but none of the cascade reaches trl or the source code, so the change is fully contained in requirements.txt.

Submitting this as a notebook-only PR for now (matching the precedent of the Gemma 4 recipe submitted earlier today) so reviewers can merge it without touching shared code or other recipes. If requirements.txt is bumped repo-wide in a future PR, the %%writefile cell becomes a no-op — the notebook continues to work without change.

Files added

0_model_customization_recipes/supervised_finetuning/
├── finetune--Qwen--Qwen3.5-4B-Base.ipynb
├── finetune--Qwen--Qwen3.5-4B.ipynb
├── finetune--Qwen--Qwen3.5-9B-Base.ipynb
├── finetune--Qwen--Qwen3.5-9B.ipynb
└── sagemaker_code/hf_recipes/Qwen/
    ├── Qwen3.5-4B-Base--vanilla-peft-qlora.yaml
    ├── Qwen3.5-4B-Base--vanilla-full.yaml
    ├── Qwen3.5-4B--vanilla-peft-qlora.yaml
    ├── Qwen3.5-4B--vanilla-full.yaml
    ├── Qwen3.5-9B-Base--vanilla-peft-qlora.yaml
    ├── Qwen3.5-9B-Base--vanilla-full.yaml
    ├── Qwen3.5-9B--vanilla-peft-qlora.yaml
    └── Qwen3.5-9B--vanilla-full.yaml