context : use n_embd_out for pooled embedding extraction

[extfs]

Summary

The MEAN/CLS/LAST pooling paths in encode() and decode() use n_embd (set to hparams.n_embd_inp()) to read from the pooled embedding tensor t_embd. For models with deepstack layers (qwen3vl), n_embd_inp() returns 16384 while the pooled tensor only has n_embd_out() = 4096 floats per sequence, causing:

GGML_ASSERT(offset + size <= ggml_nbytes(tensor) && "tensor read out of bounds")

The fix uses hparams.n_embd_out() instead, consistent with how POOLING_TYPE_NONE (line 1756) already handles it.

Affected models

Qwen3-VL-Embedding-8B (and any model with n_deepstack_layers > 0)

Test

Tested with Qwen3-VL-Embedding-8B-Q4_K_M.gguf, --embedding --pooling last:

• Before: crash on startup
• After: correct 4096-dim L2-normalized embeddings

No impact on models without deepstack (n_embd_inp == n_embd_out).

Use case

Qwen3-VL-Embedding can create embeddings for both images and text in the same vector space. A typical workflow is to index images using the full model on GPU, then run text-only retrieval queries against that index using a quantized GGUF on CPU. This fix makes the second step possible.

Disclosure

AI was used as a research/debugging aid to locate the root cause.