qwen3.6-27b-mtp no gains when MTP enabled

[gggiiia]

MTP Appears Non-Functional on Qwen3.6-27B-MTP (Performance Drops Instead of Improving)

Environment

• Model: Qwen3.6-27B-MTP-GGUF
• OS: Fedora 44
• Runtime: Vulkan (llama.cpp v2.20.1)
• GPU: AMD Radeon RX 7900 XTX

Issue

Enabling MTP for Qwen3.6-27B-MTP-GGUF does not appear to provide any multi-token prediction benefit.

Instead of increasing throughput, generation speed drops significantly, suggesting that MTP may not actually be functioning while still incurring additional overhead.

Results

Configuration	Generation Speed
MTP Disabled	~13 tokens/s
MTP Enabled	~13 tokens/s

Expected Behavior

When MTP is working correctly, generation speed should increase due to successful multi-token predictions.

For comparison, on the same system, Qwen3.6-35B-A3B-MTP-GGUF behaves as expected:

Model	MTP Disabled	MTP Enabled
Qwen3.6-35B-A3B-MTP-GGUF	~50 tokens/s	~110 tokens/s

Why This Looks Like an MTP Issue

The runtime and hardware are clearly capable of benefiting from MTP, as demonstrated by the 35B A3B model.

With Qwen3.6-27B-MTP-GGUF, enabling MTP appears to:

• Provide no observable multi-token prediction speedup.
• Reduce throughput by roughly 60%.
• Behave as though MTP overhead is present, but the MTP predictions themselves are not contributing to generation.

Reproduction

1. Load Qwen3.6-27B-MTP-GGUF.
2. Enable MTP in Advanced Settings.
3. Generate text and measure throughput.
4. Disable MTP and repeat.
5. Observe that throughput decreases from ~35 t/s to ~14 t/s when MTP is enabled.

Question

Is MTP currently expected to work with Qwen3.6-27B-MTP-GGUF under Vulkan and llama.cpp v2.20.1, or are there known limitations/issues affecting this model?