feat(kt): add --kt-skip-gpu-expert-cpu-copy to reclaim host RAM for GPU experts

[gvr13n]

Companion to the kt-kernel PR (kvcache-ai/ktransformers#2085); see kvcache-ai/ktransformers#2084 for the full rationale and numbers.

Adds the user-facing --kt-skip-gpu-expert-cpu-copy server arg and resolves it into an effective flag passed to KTMoEWrapper:

_mask_static = method == "MXFP4" or (method == "MXFP8" and not kt_enable_dynamic_expert_update)
skip = kt_skip_gpu_expert_cpu_copy and _mask_static

Only paths whose full-GPU prefill fallback is mask-aware may skip: MXFP4's _prepare_weight_mxfp4 → _prepare_weight_fp8 copies GPU experts device-to-device from original_layer and never reads the CPU copies; MXFP4's dynamic-update mask rewrite is unconditionally skipped, so its mask is static even with --kt-enable-dynamic-expert-update on. If the flag cannot be honored (e.g. MXFP8 + dynamic update), it warns once at layer 0 and falls back to stock behavior — safe by construction.

Measured on DeepSeek-V4-Flash (MXFP4, 100–110 GPU experts, RTX PRO 6000 + 123GB RAM): scheduler RSS ~172GB → ~99GB, decode swap io → 0, identical output.

Diff is pure additions (+29, 2 files: server_args.py, kt_ep_wrapper.py), rebased on current main.

[gemini-code-assist]

Code Review

This pull request introduces the --kt-skip-gpu-expert-cpu-copy option to skip allocating CPU copies of GPU-resident experts, reclaiming host RAM. The feedback suggests normalizing the method string to uppercase to ensure the case-insensitive comparison works correctly and prevents the option from being silently ignored.

Important

The consumer version of Gemini Code Assist on GitHub is being sunset. Starting June 18, 2026, new organization installations will be blocked, and all code review activity will officially cease on July 17, 2026.
For more details on the timeline and next steps, please review the Help Documentation.

[gemini-code-assist]

The comparison of _method with "MXFP4" and "MXFP8" is case-sensitive. If a user specifies --kt-method mxfp4 or --kt-method mxfp8 in lowercase, the condition _mask_static will evaluate to False, and the --kt-skip-gpu-expert-cpu-copy flag will be silently ignored (or a warning will be logged).

To ensure robustness and consistency with other parts of the codebase where self.kt_config.method is normalized using .upper(), we should normalize _method to uppercase.

Suggested change

	_method = self.kt_config.method
	_dyn = self.kt_config.kt_enable_dynamic_expert_update
	_mask_static = _method == "MXFP4" or (_method == "MXFP8" and not _dyn)
	_method = (self.kt_config.method or "").upper()
	_dyn = self.kt_config.kt_enable_dynamic_expert_update
	_mask_static = _method == "MXFP4" or (_method == "MXFP8" and not _dyn)

[gvr13n]

Fixed in d9a8b1c — --kt-method has no argparse choices constraint, so lowercase input was indeed possible and would silently disable the flag. The method string is now normalized with .upper() before the comparison.