Qwen3.5 has a 248,320-token vocabulary. At 8K context the full logits
tensor is 8192 x 248320 x 4 = 7.68 GB, which causes OOM on T4/P100.
The unsloth compiler already applies fused CE via apply_fused_lm_head,
but this adds an explicit FastQwen3_5Model dispatch for cleaner routing
and better error messages when Qwen3.5 is not supported.
Changes:
- Add unsloth/models/qwen3_5.py with FastQwen3_5Model that patches
Qwen3_5ForConditionalGeneration and Qwen3_5ForCausalLM forwards to
use unsloth_fused_ce_loss directly from hidden_states
- Add loader dispatch for model_type == "qwen3_5" before "qwen3"
- Version gate uses >= 5.0.0 (qwen3_5 only exists in transformers 5.x)
- Guarded import in loader.py with try/except fallback
- GDN layers intentionally left unpatched (flash-linear-attention)
- 23 unit tests covering all 4 code paths
Fixes from original PR #4331 by @vitalis:
- Add explicit _get_dtype import (wildcard import skips _-prefixed names)
- Single-token fast path now checks labels is None before returning early
- Default model name corrected to Qwen/Qwen3.5-9B (8B does not exist)
- Test assertion on nn.Linear removed (not a mock)
- Unused imports removed
Tested: Qwen3.5-0.8B 4bit training, 1.38 GB peak memory, 23/23 tests pass.
Backwards compatible: import unsloth works on transformers 4.57.6.
6974e4d37b