Verify each finding against the current code and only fix it if needed.

Inline comments:
In `@examples/glm4.7-flash/README.md`:
- Line 63: The README's full-finetune tip is incomplete: when instructing users
to remove adapter: qlora and load_in_4bit: true from the FSDP2 config, also
disable or remove quantize_moe_experts because it currently requires adapter to
be lora or qlora and will fail validation; update the sentence to say to set
quantize_moe_experts: false (or remove that key) in the FSDP2 config when doing
a full finetune so validation passes.

In `@scripts/cutcrossentropy_install.py`:
- Line 32: The pip-install line in scripts/cutcrossentropy_install.py currently
pins the repository with a short commit hash ("a668583") in the string
f'{UV_PREFIX}pip install "cut-cross-entropy[transformers] @
git+https://github.com/axolotl-ai-cloud/ml-cross-entropy.git@a668583"'; replace
that abbreviated SHA with the repository's full 40-character commit SHA to
ensure deterministic dependency resolution (look up the full SHA for the
intended commit in the ml-cross-entropy repo and update the string accordingly).

In `@src/axolotl/monkeypatch/fsdp2_qlora.py`:
- Around line 166-193: The runtime monkeypatch re-wraps methods on every call;
modify apply_linear8bitlt_save_patch to be idempotent by detecting and skipping
if already applied: capture and store the original function once (e.g., a
module-level variable or by checking an attribute on
bnb.nn.Linear8bitLt._save_to_state_dict), return immediately if the method is
already the patched wrapper, and when installing the wrapper, mark the patched
function with a sentinel attribute (e.g., _is_axolotl_patched = True) so
repeated calls do nothing; apply the same pattern to
apply_init_dtype_attrs_patch (use unique sentinel names and preserved originals
like original_save / original_init_dtype_attrs) to avoid stacking wrappers.

In `@src/axolotl/monkeypatch/moe_quant.py`:
- Around line 82-107: Compute and write the current runtime quantization state
into _moe_load_state before the early "already patched" return: determine mode
from cfg (same logic using getattr(cfg, "load_in_8bit", False) to select "8bit"
or "4bit"), set _moe_load_state["mode"] = mode and _moe_load_state["count"] = 0
(and if applicable pre-populate
_moe_load_state["quant_type"]/_moe_load_state["compress_statistics"] when mode
== "4bit") before the if _moe_load_state["patched"] check so later loads don't
see stale state from prior loads.

In `@src/axolotl/utils/schemas/config.py`:
- Around line 632-641: Extend the existing Pydantic validation that currently
checks adapter + load_in_4bit/load_in_8bit to also enforce that
quantize_moe_experts can only be true when the runtime backend is CUDA (reject
when backend is ROCm/other); specifically, in the same validator that references
quantize_moe_experts and the load_in_4bit/load_in_8bit flags, add a check that
the configured backend/device backend (or runtime torch backend detection)
indicates CUDA and raise a ValidationError if quantize_moe_experts is true but
the backend is not CUDA so the config fails validation early.

---

Outside diff comments:
In `@src/axolotl/core/trainers/base.py`:
- Around line 753-757: The save_pretrained call is passing is_main_process
unconditionally which breaks custom implementations (e.g., modeling_mamba's
save_pretrained) that don't accept that kwarg; update the code that calls
save_pretrained (the call where state_dict and is_main_process are passed) to
first inspect the save_pretrained signature (use inspect.signature on the
model's save_pretrained method) and only include is_main_process when the
parameter is accepted (otherwise call with only save_directory and state_dict);
ensure you reference the same save_pretrained method used in the trainer and use
self.accelerator.is_main_process for the flag.

---

Nitpick comments:
In `@examples/glm4.7-flash/qlora.yaml`:
- Around line 26-34: The config currently only adapts attention projection
modules (lora_target_modules: q_proj, v_proj, k_proj, o_proj) while the MoE
expert parameter targets (lora_target_parameters: mlp.experts.gate_up_proj,
mlp.experts.down_proj) are commented out; update the YAML to make intent
explicit by either uncommenting the lora_target_parameters entries (and enabling
lora_mlp_kernel if needed) when you intend to adapt experts, or add a one-line
inline comment above lora_target_modules explaining this example only adapts
attention projections and does not modify MoE expert parameters.

In `@src/axolotl/loaders/patch_manager.py`:
- Around line 380-387: Only apply the PEFT monkeypatch when adapters are
actually in use and ensure it's applied once: check the config flag that
indicates an adapter path is active (e.g., self.cfg.adapter_path or equivalent
adapter-enabled field) before calling patch_peft_target_parameters_matching(),
and guard the call with a one-time flag (module-level or attribute like
self._peft_patch_applied) so subsequent loads don't reapply the global patch;
keep patch_moe_quantization_on_load(self.cfg) behavior unchanged but move or
wrap patch_peft_target_parameters_matching() behind the adapter check and
one-time guard.