fix: align LlamaContextParams/LlamaModelParams with llama.cpp b6862 ABI

[Marvinthebored]

Problem

LlamaContextParams matches a much older llama.cpp than the bundled b6862 libraries. By coincidence both layouts are 120 bytes, so nothing crashes — but every field in the first 40 bytes (and the bool tail) silently lands on the wrong C struct member, on every platform and both FFI paths (purego struct registration on darwin, libffi descriptors elsewhere — ffi.go encodes the same stale layout).

Go field (old)	actual C field it landed on (b6862)
Seed	n_ctx
NCtx	n_batch
NBatch	n_ubatch
NUbatch	n_seq_max
NSeqMax	n_threads
NThreads	n_threads_batch
NThreadsBatch	rope_scaling_type
RopeScalingType	pooling_type
PoolingType	attention_type
AttentionType	flash_attn_type
(aligned again from rope_freq_base through abort_callback_data)
Logits	embeddings
Embeddings	offload_kqv
Offload_kqv	no_perf
FlashAttn	op_offload
NoPerf	swa_full

Observed symptoms (darwin-amd64, official b6862 dylib):

• cp.NUbatch = 512 → llama_init_from_model: failed to initialize the context: n_seq_max must be <= 256
• cp.Embeddings = 1 actually enabled offload_kqv; real embeddings stayed false → llama_get_embeddings* returns NULL
• Context_default_params() reported NCtx=2048 (that's n_batch's default read through the wrong offset)

Fix

• LlamaContextParams: drop Seed (removed from llama_context_params upstream long ago; LLAMA_DEFAULT_SEED kept for samplers), add FlashAttnType (LLAMA_FLASH_ATTN_TYPE_{AUTO,DISABLED,ENABLED}), replace the bool tail with b6862's {embeddings, offload_kqv, no_perf, op_offload, swa_full, kv_unified}.
• LlamaModelParams: add trailing NoHost (b6862).
• ffi.go: both libffi struct descriptors updated to the same layout.
• Hardcoded default fallbacks updated; new fields use llama.cpp's defaults.
• Examples: drop obsolete ctxParams.Logits = 1 (no context-level logits flag in b6862; per-token logits via batch.logits).
• ffi_test: assert on real fields instead of Seed (that assertion only passed because Seed was reading n_ctx).

Verification (darwin-amd64, official b6862 dylib)

With this patch, through the public API only:

defaults via gollama API: n_ctx=512 n_batch=2048 n_ubatch=512 n_seq_max=1 flash_attn_type=-1   ← correct llama.cpp defaults, correct fields
Init_from_model with mutated NCtx/NBatch/NUbatch/Embeddings/PoolingType → context created
Tokenize + Encode + Get_embeddings_ith → real 768-dim embedding (nomic-embed-text-v1.5 Q8_0), L2≈24.4

Note for darwin-amd64

Unrelated to this PR but adjacent: with the pinned purego v0.9.1, RegisterFunc(llama_init_from_model) panics with too many arguments on darwin-amd64 (the 120-byte by-value struct exceeds purego's stack-slot budget). Fixed upstream in purego via ebitengine/purego#425 (first available in v0.11.0-alpha.3). Happy to send that bump as a separate PR if wanted.

🤖 Generated with Claude Code