#2982 Qwen3.6-35B-A3B UD-Q4_K_XL: Low GPU utilization, limited generation speed, and MTP acceptance concerns

Description

Hi Strata team,

I'm trying to get the best performance out of **Qwen3.6-35B-A3B-UD-Q4_K_XL** on my system, but so far the results have been disappointing. I'd appreciate some guidance on the recommended configuration for this model, particularly regarding MoE expert offloading, GPU utilization, and MTP speculative decoding.

### Hardware and model

- **GPU:** NVIDIA RTX 5070 Ti, 16 GB VRAM
- **CPU:** AMD Ryzen 9 9950X, 16 cores
- **System RAM:** 62 GB
- **NVIDIA driver:** 617.42
- **Model:** Qwen3.6-35B-A3B-UD-Q4_K_XL.gguf
- **Quantization:** UD-Q4_K_XL
- **Context:** 65,536 tokens
- **Vision:** Enabled, using mmproj-F16.gguf

### Results with Strata

My current configuration uses:

- `--pool-workers 4`
- `--max-context 65536`
- `--prefill auto`
- `--mtp 2`
- `--vision`

The startup log reports:

- 4.40 GiB of model tensors on the GPU
- 42 of 256 experts per layer in VRAM
- Approximately 18.75 GiB of expert weights in pinned host memory
- CPU-side processing for the remaining experts
- 4.94 GiB of free VRAM after loading

In my testing, generation speed is approximately **60–77 tokens/s**, while GPU utilization is generally around **50–60%**.

I also tried increasing `--pool-workers` to 6 and `--mtp` to 4, but performance became worse, so I returned to 4 workers and MTP 2.

One thing that concerns me is that the logs report `MTP block loaded`, but Strata's session logs repeatedly show:

`drafts accepted 0 of 0`

I cannot tell whether MTP is actually participating in generation, whether these counters are expected for this model, or whether my configuration is missing something.

### Comparison with llama.cpp

I tested the same GGUF with a recent prebuilt CUDA-enabled llama.cpp build, using a 65,536-token context, CPU MoE offloading, and `--spec-type draft-mtp`.

The server reported:

- **Generation:** approximately 55.4 tokens/s
- **Prompt processing:** approximately 87 tokens/s
- **Draft acceptance:** 73.4% (1,082 accepted out of 1,474 generated draft tokens)
- **Mean accepted draft length:** 2.24

This confirms that speculative decoding can work with this model in llama.cpp, although its generation speed was still lower than my better Strata runs.

For comparison, my Qwen3.8 Flash-Next setup in Strata reaches approximately **100 tokens/s** on the same PC.

### Questions

1. Is my current Strata configuration appropriate for Qwen3.6-35B-A3B UD-Q4_K_XL on a 16 GB GPU?
2. Is there a recommended way to improve GPU utilization and reduce the CPU-side MoE bottleneck without increasing VRAM usage beyond what this GPU can handle?
3. Does `drafts accepted 0 of 0` indicate that MTP is not functioning, or is this expected behavior for Qwen3.6 in Strata?
4. Are there recommended MTP settings for this model? Increasing the value from 2 to 4 did not improve performance on my system.
5. Are there any known limitations or optimizations for Qwen3.6's hybrid MoE architecture, especially when most expert weights reside in pinned host memory?

I'm not looking to rebuild or compile Strata myself. I'd like to understand whether this performance is expected for my hardware or whether there is a known configuration or implementation issue that could improve it.

Thanks for your work on Strata. Any specific recommendations or clarification about the MTP logs and hybrid expert offloading would be appreciated.