Add preliminary MiniMax-M3 support

[jkyamog]

Adds preliminary text-only MiniMax-M3 support to ik_llama.cpp, based on the public llama.cpp MiniMax-M3 implementation shape and adapted to ik's graph/tensor-loader structure.

This is intended as a basic enablement PR: load GGUF MiniMax-M3, run server inference, and support hybrid CPU/GPU expert placement. It does not attempt vision/MTP support or graph/attention split support for MiniMax-M3.

Main Changes

• Adds LLM_ARCH_MINIMAX_M3 and architecture name mapping.
• Adds MiniMax-M3 tensor-name mapping.
• Adds MiniMax-M3 hparam loading for:
  • leading dense block count
  • expert FFN size
  • shared expert count
  • expert weight scale/norm/gating function
• Adds MiniMax-M3 tensor creation:
  • per-head Q/K norms
  • leading dense FFN layers
  • routed experts
  • shared experts
• Adds build_minimaxm3() graph:
  • RMS attention norm
  • Q/K per-head RMSNorm before partial RoPE
  • dense leading FFN via ggml_swiglu_oai
  • routed MoE via LLM_FFN_SWIGLU_OAI_MOE
  • shared expert FFN via ggml_swiglu_oai
• Updates fused MoE helper routing so LLM_FFN_SWIGLU_OAI_MOE uses ggml_moe_up_gate_ext() instead of the unsupported fused path variant.

Validation Performed

• CPU build of llama-server completed successfully.
• Docker CUDA image built successfully from the patched tree.
• MiniMax-M3 Q4 server loaded successfully with 128k context.
• Using Pinchbench tasks (0-17 task) to verify it works.

Runtime Notes

On a single 32GB V100 with MiniMax-M3 Q4 and CTX_SIZE=131072:

• All MoE on CPU works.
• Partial GPU MoE requires disabling current fused MoE CUDA paths:
  • --no-fused-moe
  • --no-fused-up-gate
  • --no-fused-mul-multiadd
• Without those flags, CUDA fused MoE hits an assertion because the current fused kernel path accepts SILU/RELU/GELU but not GGML_UNARY_OP_SWIGLU_OAI.
• Two GPU expert layers fit at 128k context on this V100, ending around 30.4GB used / 2.1GB free.

Known Limitations

• Text-only preliminary support.
• Split mode graph/attention is not implemented in the MiniMax-M3 graph builder.
• Fused CUDA MoE support for GGML_UNARY_OP_SWIGLU_OAI remains future work.

[jkyamog]

This is just based on ggml-org/llama.cpp#24523 and I just pointed codex is implement m3 support for ik_llama.cpp. Please kindly test, I haven't extensively tested this yet but so far its seems to work.

[ikawrakow]

I tested quickly, it seems to work. One should change the graph building to use standard ik_llama.cpp facilities, but I can do that after merging. Why is the PR marked ad draft?

[jkyamog]

I tested quickly, it seems to work. One should change the graph building to use standard ik_llama.cpp facilities, but I can do that after merging. Why is the PR marked ad draft?

Thanks for testing. I haven't really coded in c++ for over 3 decades, I can just vaguely understand c++ but ik_llama.cpp code base is much more advance than what I am comfortable understanding fully. Anyway I have done more testing a few improvements since:

• I added a working template, so Minimax 3 can use tools. After testing with openclaw I needed to get the tool call working.
• I got your comment above and fed it to codex, it has now refactored to use the ik_llama.cpp facilities which in my opinion is much better. Hopefully this saves some time.

I have now set this PR for review.

[ikawrakow]

Split mode graph does not work, but I'll fix it in a separate PR.