Support arbitrary expansion factor in Mamba2 (remove hardcoded 2 * n_embd assertion)

[limloop]

Problem

Currently, LLM_ARCH_MAMBA2 in llama.cpp has a hardcoded assertion:

GGML_ASSERT(2 * n_embd == d_inner);

This prevents loading any Mamba2 model where expand != 2 (e.g., expand=1.5, expand=1, etc.).

Proposed solution

Two small changes:

1. In llama-model.cpp (around line 4309):

- const int64_t d_in_proj = 2*d_inner + 2*n_group*d_state + n_head;
- // only an expansion factor of 2 is supported for now
- GGML_ASSERT(2 * n_embd == d_inner);
+ const int64_t conv_dim = d_inner + 2 * n_group * d_state;
+ const int64_t d_in_proj = d_inner + conv_dim + n_head;
+ 
+ if (2 * n_embd != d_inner) {
+     LLAMA_LOG_WARN("Mamba2: non-standard expansion factor (d_inner=%ld, n_embd=%ld). Continuing.\n", d_inner, 2 * n_embd);
+ }

2. In convert_hf_to_gguf.py:

Replace hardcoded 2 * self.d_model with actual expand value from the source model config.

Why this change?

• Models with expand=2 (default) continue to work identically
• Models with custom expansion factors (e.g., 1.5) become usable
• The change is minimal and backward compatible

Testing

Tested with:

• Standard Mamba2 (expand=2) — ✅ works
• Custom Mamba2 with expand=1, hidden_size=512, d_inner=512 — ✅ loads and runs

Discussion points

1. Is there any reason to keep the 2x restriction? (performance optimizations? numerical stability?)
2. Would a warning be acceptable instead of a hard assert?

[ggerganov]

I don't see a problem to support other factors. No need to keep the assert as it is or replace it with warning.

cc @compilade for extra opinion

[compilade]

Originally, I put the 2x expand assertion cause I thought it was mostly part of the architecture (since all Mamba models at the time had a 2x expand factor), but there's nothing which actually requires that specific factor (since n_embd and d_inner are stored separately).

So yes, the assertion can be removed.