945c8aeb10
* fix: saving clones state dict
* fix: apply fix for only CP mode
* fix: add dropout check when using lora target param
* fix: re-add patch from transformers PR #39866
* feat: add moe quant to test by ved
* fix: try match target param properly end with
* fix: clear cache per param quant
* fix: attempt on-load quantize experts instead of post-load
* fix: attempt disable async load
* chore: add log
* chore: adjust log
* fix: remove cuda alloc for moe and enable async load
* chore: remove leftover logs
* chore: add extra empty cache
* fix(doc): clarify support
* fix: handle fsdp2 for paramwrapper dtensor
* feat: attempt to quant experts in 8bit mode too
* feat: attempt to release bf16 experts from vram
* feat: upgrade cce
* fix: fsdp2 init_sharded_param load int8/uint4 dtensor as
require_grad=true on init
* fix: remove unnecessary gc and empty cache
* Revert "fix: remove unnecessary gc and empty cache"
This reverts commit 1d54518990.
* fix: do not call full_tensor on non-dtensors
* fix: attempt to address fsdp2 with quant exp high loss
* fix: attempt lora quant experts wrong dim
* fix: ensure require_grad patch applied for lora 8bit
* fix: attempt lora 8bit fsdp2
* fix: attribute access on save for lora 8bit fsdp2
* fix: wrong weight attrib access
* chore(refactor): add config, re-arrange position of patches, clean
comments
* feat: add example docs
* chore: cherry pick trinity fixes from PR 3399
* chore: comments refactor; add guards
* fix: guard using wrong key
* fix: mamba save does not accept main process param
* fix: guard prevent double hook
* fix: move gc to upper scope
* chore: add comment on proxy forward patch
* fix: add comment to clarify
* feat: add test idempotency
* fix: AttributeError: `e_score_correction_bias` is not an nn.Parameter
* fix: AttributeError: 'NoneType' object has no attribute 'to'
* fix: update docs on cpu_ram_efficient_loading
Finetune Z.ai's GLM-4.7-Flash with Axolotl
GLM-4.7-Flash is a 30B-A3B MoE model by Z.ai.
This guide shows how to fine-tune it with Axolotl.
Getting started
-
Install Axolotl following the installation guide.
-
Install Cut Cross Entropy to reduce training VRAM usage.
-
Run the finetuning example:
# QLoRA
# - no target experts (1x48GB @ ~24GiB/GPU)
# - target experts (1x48GB @ ~34GiB/GPU)
axolotl train examples/glm4.7-flash/qlora.yaml
# QLoRA FSDP2 no target experts (2x48GB @ ~29GiB/GPU)
axolotl train examples/glm4.7-flash/qlora_fsdp.yaml
# LoRA
# - no target experts (1x48GB @ ~35GiB/GPU)
# - target experts (1x48GB @ OOM. Projected ~45-50GiB/GPU)
axolotl train examples/glm4.7-flash/lora.yaml
# LoRA FSDP2 no target experts (2x48GB @ ~43GiB/GPU)
axolotl train examples/glm4.7-flash/lora_fsdp.yaml
Expert LoRA
To also apply LoRA adapters to expert weights, add lora_target_parameters to your config.
Note: lora_dropout must be 0 when using lora_target_parameters.
lora_target_parameters:
- mlp.experts.gate_up_proj
- mlp.experts.down_proj
# - mlp.gate.weight # router, untested but should work, not normally targeted
Limitations
- FSDP VRAM: FSDP2 may use more VRAM per GPU than single GPU training. We suspect not all layers are properly sharded across ranks.
- FSDP initial spike: FSDP LoRA (8-bit) may have a large initial VRAM spike at the first 1-2 steps that then drops. FSDP QLoRA (4-bit) does not exhibit this.
- cpu_ram_efficient_loading: Must be set to
falsewith FSDP2 — causes hang otherwise. - lora_target_linear: Incompatible for this model.
- LoRA kernels: Incompatible with this model due to non-standard attention projections (DSA). Must be explicitly disabled (
lora_*_kernel: false).
TIPS
- For inference, the official Z.ai team recommends these default settings (most tasks):
temperature: 1.0top_p: 0.95max_new_tokens: 131072
- You can run a full finetuning by removing
adapter: qlora,load_in_4bit: true, andquantize_moe_experts: truefrom the config. This is heavy, so we have not tested this. - Read more on how to load your own dataset at docs.
Optimization Guides
Please check the Optimizations doc.