fix(mllama): image inference crash on CUDA, eager-only paged flag, vision head config

[subin9]

Fixes three issues that together make Llama-3.2-11B-Vision (mllama / vllama) unusable for image inputs. Closes #2335.

1. Image inference crashes on CUDA (main fix)

On decode steps the mllama inputs processor still ran its image block, because Sequence::take_images() clones (rather than empties) the images when has_changed_prompt is false, so has_images() stays true after prefill. A decode input has no <|image|> token, so convert_sparse_cross_attention_mask_to_dense produced a cross-attention mask with a zero-sized max_num_images dimension, e.g. [1, 1, 0, 4] (0 elements). prepare_cross_attention_mask then calls .to_dtype(F32) on that tensor; launching a CUDA kernel over 0 elements returns CUDA_ERROR_INVALID_VALUE. On CPU it is a harmless no-op, which is why text generation (and CPU) worked.

Fix: gate the image block on is_prompt, matching Qwen2VLImageProcessor (let mut pixel_values = if is_prompt { pixel_values } else { None };). Cross-attention is skipped during decode anyway (mllama/text.rs: if cross_attn_states.is_none() { continue }), so passing image inputs on decode was both wasteful and, on CUDA, fatal.

2. supports_paged_attention advertised support the model can't do

VLlamaLoader::supports_paged_attention returned true, but MLlamaTextModel requires eager attention, so a default launch aborted with Error: Expected eager attention implementation rather than auto-disabling PagedAttention. Return false, matching the Qwen2-VL / Qwen2.5-VL loaders (which have the same eager requirement and correctly fall back).

3. Vision attention_heads config key ignored

MLlamaVisionConfig::num_attention_heads had #[serde(default = "d_attn_heads")] (=16) but no alias for HF's actual key attention_heads, so the vision head count from config.json was ignored and always 16. The stock 11B has 16 vision heads so it was masked; any mllama checkpoint with a different vision head count loaded with wrong dimensions or a weight shape mismatch. Added #[serde(alias = "attention_heads")].

Reproduction / verification

Reproduced and verified all three on an RTX 4070 Ti SUPER (CUDA 12.9) using a tiny same-architecture model, yujiepan/llama-3.2-vision-tiny-random (~22 MB), which hits the identical MLlamaModel forward path without the 20 GB download. Before: default launch errors; --paged-attn off + first image request -> CUDA_ERROR_INVALID_VALUE. After: model loads by default and small/large/repeat image requests plus text all return 200 with no CUDA errors. The stock meta-llama/Llama-3.2-11B-Vision-Instruct reproduces (1) and (2) identically.

[github-actions]

Code Metrics Report

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
 Language              Files        Lines         Code     Comments       Blanks
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
 C Header                 23         4454         3116          790          548
 CSS                       3          282          252            6           24
 CUDA                    119        23681        19217         1704         2760
 Dockerfile                1           35           19            9            7
 HTML                      2           27           27            0            0
 JavaScript                3          577          562           12            3
 Jinja2                    7          694          656            5           33
 JSON                     27        15864        15861            0            3
 Makefile                  1           18           16            0            2
 MDX                      36         5901            0         4327         1574
 Metal Shading Lan|       37        14287        11284         1136         1867
 PowerShell                1          656          570           31           55
 Python                  151        12284        10191          484         1609
 Shell                     3         1062          843          117          102
 Plain Text               53        10687            0         9209         1478
 TOML                     28         1368         1189           39          140
 TypeScript               11         1641         1404           66          171
 YAML                      3           25           23            2            0
─────────────────────────────────────────────────────────────────────────────────
 Jupyter Notebooks         3          122           83           23           16
 |- Markdown               1           60           30           22            8
 |- Python                 1          122          113            1            8
 (Total)                              304          226           46           32
─────────────────────────────────────────────────────────────────────────────────
 Markdown                273        11726            0         8736         2990
 |- BASH                  23          295          217           46           32
 |- Dockerfile             2            5            5            0            0
 |- JSON                   6          289          289            0            0
 |- PowerShell             1            1            1            0            0
 |- Python               135         7239         6021          310          908
 |- Rust                  62         3820         2838          388          594
 |- TOML                   7           77           65            0           12
 (Total)                            23452         9436         9480         4536
─────────────────────────────────────────────────────────────────────────────────
 Rust                    674       299702       267081         5872        26749
 |- Markdown             413         9761          452         8126         1183
 (Total)                           309463       267533        13998        27932
─────────────────────────────────────────────────────────────────────────────────
 Svelte                   19         1969         1826           51           92
 |- CSS                    1            4            4            0            0
 |- JavaScript            19          921          767           25          129
 (Total)                             2894         2597           76          221
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
 Total                  1478       429656       345022        41537        43097
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━