[Sync][rebuild] FlashQLA/GDN (#1947) + param-buffer cpu-backup (#1952) + DistOptim step ckpt rider (#1945)

[aoshen02]

Summary

Rebuild-gated sync PR (touches docker/ + megatron.patch → needs image rebuild to validate). Three engine-agnostic slime PRs, mirror-slime:

• #1947 FlashQLA/GDN backend (--qwen-gdn-backend fla|flashqla): Dockerfile FlashQLA pip (arch-gated ARG INSTALL_FLASHQLA, x86=1/arm=0 — sm90/Hopper-only, verified gb200/Blackwell import-fails + downgrades tilelang) + qwen_gdn_backend.py + qwen3_5/qwen3_next selector + reloadable_process_group comm memory-check skip + tests + install doc.
• #1952 disable_param_buffers_cpu_backup (train-side tms, engine-agnostic): megatron.patch param-buffer cpu-backup + drop --disable-weights-backuper + arguments/placement wiring.
• #1945 DistributedOptimizer step ckpt rider (engine-agnostic): megatron.patch del tensors['step'] + step LocalNonpersistentObject + src_tensors.pop. (sglang v0.5.12 body = N/A; vime has no sglang image.)

Validation

Both images built (x86 FLASHQLA=1 + arm FLASHQLA=0), megatron.patch applies clean both arches. See #107.

🤖 Generated with Claude Code

[gemini-code-assist]

Code Review

This pull request introduces support for the FlashQLA backend as an optional GDN implementation for Qwen3.5 and Qwen3-Next models, including Dockerfile integration, documentation, command-line arguments, and tests. It also updates Megatron patches to support disabling parameter buffer CPU backups, refactors memory checks in reloadable process groups to skip certain communication operations, and refactors model offloading/switching logic. A review comment suggests improving the robustness of the version parsing utility in qwen_gdn_backend.py to handle non-numeric characters in version strings (such as pre-releases) using regular expressions.

[gemini-code-assist]

The _parse_version function uses a simple string split on . to parse version numbers. This can fail with a ValueError if any version component contains non-numeric characters (e.g., pre-releases or release candidates like 12.8rc1 or 2.8a0). Using a regular expression to extract digit sequences is much more robust and prevents potential runtime crashes.

Suggested change

	import torch
	def _parse_version(version):
	version = version.split("+", 1)[0]
	parts = version.split(".")
	major = int(parts[0])
	minor = int(parts[1]) if len(parts) > 1 else 0
	return major, minor
	import re
	import torch
	def _parse_version(version):
	digits = re.findall(r'\d+', version)
	major = int(digits[0]) if len(digits) > 0 else 0
	minor = int(digits[1]) if len(digits) > 1 else 0
	return major, minor

[CalvinXKY]

LGTM