qwen3 muon fork

[Shuwen-Fang]

Description

Start with a short description of what the PR does and how this is a change from
the past.

The rest of the description includes relevant details and context, examples:

• why is this change being made,
• the problem being solved and any relevant context,
• why this is a good solution,
• some information about the specific implementation,
• shortcomings of the solution and possible future improvements.

If the change fixes a bug or a Github issue, please include a link, e.g.,:
FIXES: b/123456
FIXES: #123456

You can also provide a comma-separated list. If you don't want to close a bug but
simply to reference it, use BUGS, e.g.:
BUGS: b/123456

Notice 1: Once all tests pass, the "pull ready" label will automatically be assigned.
This label is used for administrative purposes. Please do not add it manually.

Notice 2: For external contributions, our settings currently require an approval from a MaxText maintainer to trigger CI tests.

Tests

Please describe how you tested this change, and include any instructions and/or
commands to reproduce.

Checklist

Before submitting this PR, please make sure (put X in square brackets):

• I have performed a self-review of my code. For an optional AI review, add the gemini-review label.
• I have necessary comments in my code, particularly in hard-to-understand areas.
• I have run end-to-end tests tests and provided workload links above if applicable.
• I have made or will make corresponding changes to the doc if needed, including adding new documentation pages to the relevant Table of Contents (toctree directive) as explained in our documentation.

[gemini-code-assist]

Code Review

This pull request introduces a hybrid Gated Delta Net (GDN) implementation using Tokamax GDN v3 forward and custom VJP backward, refactors Qwen3-Next decoder blocks to support scanned and remainder layers, and forks the optax Muon optimizer to allow custom modifications. The review feedback highlights several critical issues: Flax Linen auto-naming of remainder layers will cause checkpoint conversion mismatches unless explicit names are provided; remainder layers are missing parameter hooks when scanning is enabled; the pure_jax_fused_conv1d_gdn function ignores the provided conv_state; and the non-mesh branch of Qwen3NextGatedDeltaNet discards the convolution bias by hardcoding it to None.

[gemini-code-assist]

When scan_layers is enabled and there are remainder layers, they are instantiated without an explicit name. Flax Linen will auto-name them (e.g., Qwen3NextDecoderLayerToLinen_0), which will mismatch the expected prefix params-decoder-layers_{layer_id} in param_mapping.py. Specifying name=f"layers_{layer_id}" ensures the parameter names are consistent and checkpoint conversion works correctly.

Suggested change

	layer = qwen3.Qwen3NextDecoderLayerToLinen(
	config=cfg,
	mesh=mesh,
	model_mode=model_mode,
	quant=self.quant,
	layer_idx=layer_id,
	)
	layer = qwen3.Qwen3NextDecoderLayerToLinen(
	config=cfg,
	mesh=mesh,
	model_mode=model_mode,
	quant=self.quant,
	layer_idx=layer_id,
	name=f"layers_{layer_id}",
	)

[gemini-code-assist]

When scan_layers is enabled, the hooks are only attached to the scanned blocks (local_layers and global_layer). If there are remainder layers, their parameters (such as attention query/key/value/out kernels) will not have the required hooks attached, leading to shape or value mismatches during checkpoint conversion. We should also attach hooks to the remainder layers.

Suggested change

	if scan_layers:
	_attach_block_hooks("params-decoder-scanned_blocks-local_layers", is_global=False)
	_attach_block_hooks("params-decoder-scanned_blocks-global_layer", is_global=True)
	if scan_layers:
	_attach_block_hooks("params-decoder-scanned_blocks-local_layers", is_global=False)
	_attach_block_hooks("params-decoder-scanned_blocks-global_layer", is_global=True)
	num_blocks = config.base_num_decoder_layers // config.inhomogeneous_layer_cycle_interval
	num_scanned = num_blocks * config.inhomogeneous_layer_cycle_interval
	num_remaining = config.base_num_decoder_layers % config.inhomogeneous_layer_cycle_interval
	if num_remaining > 0:
	for rem_idx in range(num_remaining):
	hf_layer_idx = num_scanned + rem_idx
	prefix = f"params-decoder-layers_{hf_layer_idx}"
	layer_in_block = rem_idx % config.inhomogeneous_layer_cycle_interval
	is_full_attention_layer = (layer_in_block + 1) % config.inhomogeneous_layer_cycle_interval == 0
	_attach_block_hooks(prefix, is_global=is_full_attention_layer)

[gemini-code-assist]

In pure_jax_fused_conv1d_gdn, conv_state is accepted as an argument but is completely ignored in the function body, which instead always pads with zeros. If conv_state is provided (e.g., during stateful chunk-by-chunk processing), it should be prepended to qkv along the sequence dimension to ensure correct causal convolution outputs across chunk boundaries.

    if conv_state is not None:
        conv_input = jnp.concatenate([conv_state, qkv], axis=1)
    else:
        conv_input = jnp.pad(qkv, ((0, 0), (conv_kernel_size - 1, 0), (0, 0)))

[gemini-code-assist]

In the non-mesh branch of Qwen3NextGatedDeltaNet, conv_bias is hardcoded to None. If the convolution layer actually has a bias, this will discard it and produce incorrect results. We should extract and pass the actual conv_bias if it exists.

Suggested change

	core_attn_out, (next_conv_state, next_recurrent_state) = hybrid_fused_conv1d_gdn(
	qkv=qkv,
	b=b,
	a=a,
	conv_weight=self.conv1d.kernel.value,
	conv_bias=None,
	conv_bias_arg = self.conv1d.bias.value if (hasattr(self.conv1d, "bias") and self.conv1d.bias is not None) else None
	core_attn_out, (next_conv_state, next_recurrent_state) = hybrid_fused_conv1d_gdn(
	qkv=qkv,
	b=b,
	a=a,
	conv_weight=self.conv1d.kernel.value,
	conv_bias=conv_bias_arg,