Wire mlx-lm Vulkan fused kernels

[goniz]

Summary

Closes #51.

This PR wires the Vulkan fused-kernel work into the parent mlx-vulkan workspace:

• Adds the mlx-lm submodule at goniz/mlx-lm.
• Updates the mlx submodule to include Vulkan fused fast-kernel/API support from Add Vulkan fused fast kernel paths mlx#50.
• Updates the mlx-lm submodule to include the Vulkan gated-delta implementation from Add Vulkan gated delta kernel mlx-lm#1.
• Adjusts the workflow/dev script paths for the new mlx-lm submodule layout.

Related PRs:

• Add Vulkan fused fast kernel paths mlx#50
• Add Vulkan gated delta kernel mlx-lm#1

Validation

Run from this parent mlx-vulkan workspace:

• ./dev.sh test-cpp
  • 249 test cases passed
  • 3314 assertions passed
• ./dev.sh test-py
  • 699 passed
  • 13 skipped
• ./dev.sh generate
  • coherent output
• ./dev.sh generate --model mlx-community/Qwen3.6-35B-A3B-8bit
  • coherent output
• ./dev.sh run python scripts/model_generation_report.py --json-output /tmp/model_generation_report.json
  • all 10 model entries generated coherent output

Benchmarks

./dev.sh benchmark bf16

Running benchmark for model: mlx-community/Qwen3-0.6B-bf16
Running warmup..
Timing with prompt_tokens=4096, generation_tokens=128, batch_size=1.
Trial 1:  prompt_tps=2433.556, generation_tps=65.807, peak_memory=2.614, total_time=3.732
Trial 2:  prompt_tps=2419.242, generation_tps=65.866, peak_memory=2.614, total_time=3.746
Trial 3:  prompt_tps=2432.330, generation_tps=65.850, peak_memory=2.614, total_time=3.735
Trial 4:  prompt_tps=2425.489, generation_tps=65.864, peak_memory=2.615, total_time=3.740
Trial 5:  prompt_tps=2386.391, generation_tps=65.856, peak_memory=2.615, total_time=3.768
Averages: prompt_tps=2419.402, generation_tps=65.849, peak_memory=2.614

./dev.sh benchmark 8bit

Running benchmark for model: mlx-community/Qwen3-0.6B-8bit
Running warmup..
Timing with prompt_tokens=4096, generation_tokens=128, batch_size=1.
Trial 1:  prompt_tps=1349.699, generation_tps=88.834, peak_memory=2.055, total_time=4.586
Trial 2:  prompt_tps=1349.641, generation_tps=88.652, peak_memory=2.055, total_time=4.585
Trial 3:  prompt_tps=1345.601, generation_tps=88.767, peak_memory=2.056, total_time=4.589
Trial 4:  prompt_tps=1353.543, generation_tps=88.747, peak_memory=2.056, total_time=4.573
Trial 5:  prompt_tps=1343.560, generation_tps=88.527, peak_memory=2.056, total_time=4.600
Averages: prompt_tps=1348.409, generation_tps=88.705, peak_memory=2.056

Additional Qwen3.6 35B prefill check:

MLX_MPI_LIBNAME=/dev/null OMPI_MCA_accelerator=^rocm ./dev.sh run mlx_lm.benchmark --model mlx-community/Qwen3.6-35B-A3B-8bit -p 4096 -g 1 -n 3
Averages: prompt_tps=119.139, generation_tps=24232.064, peak_memory=40.349