Load Qwen3.8 GGUFs with Embedded NextN/MTP Blocks

[jkyamog]

Qwen3.8-27B GGUFs can contain the normal target transformer layers followed by
embedded NextN/MTP predictor blocks. The current loader treated every GGUF
block as a target layer and rejected the model during startup when the total
block count was not divisible by Qwen's full-attention interval.

This change separates the raw GGUF block count from the target layer count:

block_count = 65
nextn_predict_layers = 1
target_layer_count = 64
full_attention_interval = 4

The target loader now validates the 64-layer trunk and ignores the trailing
embedded MTP block for ordinary target execution.

Reproduction

Model:

Qwen3.8-27B-Q6_K.gguf

Before this change, startup failed with:

target load: block_count=65 not divisible by full_attention_interval=4
[backend_factory] Qwen35Backend init failed
[server] backend creation failed

The GGUF metadata was verified directly:

general.architecture = qwen35
general.name = Qwen3.8-27B
qwen35.block_count = 65
qwen35.nextn_predict_layers = 1
qwen35.full_attention_interval = 4
blk.64.nextn.eh_proj.weight
blk.64.nextn.enorm.weight
blk.64.nextn.hnorm.weight
blk.64.nextn.shared_head_norm.weight

Implementation

• Adds a shared helper to derive the effective target layer count.
• Applies the embedded-block rule only to qwen35 and qwen35moe metadata.
• Rejects invalid metadata where nextn_predict_layers >= block_count.
• Verifies each ignored trailing block contains a blk.<index>.nextn. tensor.
• Excludes embedded NextN tensors from the ordinary target-layer load.
• Preserves raw block_count and nextn_predict_layers in model inspection.
• Reports both raw and effective counts in loader errors.

Non-Qwen architectures are not reinterpreted, even if they contain similarly
named metadata.

[jkyamog]

works with current 3.6 drafter, but its slow. I guess we need a 3.8 drafter.

[cubic-dev-ai]

All reported issues were addressed across 4 files

<sub>Reply with feedback, questions, or to request a fix.

Re-trigger cubic</sub>

[howard0su]

LGTM

[jkyamog]

updated to address issues reported by cubic-dev-ai