nx_f32_llama_stack.nx
buildroot/runtime/nx_f32_llama_stack.nx
about
nx_f32_llama_stack.nx -- multi-block transformer stack forward.
Calls nx_f32_llama_block_forward_v3 N times across the layers
array, then advances the KV cache ONCE per forward pass.
Double-buffered execution: alternates between buf_a and buf_b
to avoid allocating per-layer scratch. The final result lands
in x_out (which may NOT alias x_in).
Layers array: layers[i] is the i64 value of a *NxF32LlamaLayer.
Caller is responsible for allocating + populating each layer.
genealogy_id: standard_transformer_stack
lineage_id: substrate_f32_llama_stack_v1
dependencies 5 imports · 7 importers
imports: nx_syscalls.nxnx_tier.nxnx_f32.nxnx_f32_kv_cache.nxnx_f32_llama_block.nx
imported by: nx_f32_llama_stack_test.nxnx_f32_llm.nxnx_f32_llm_e2e_test.nxnx_f32_llm_load_test.nxnx_f32_llm_run_test.nxnx_f32_llm_run_v2_test.nxnx_f32_llm_test.nx
structs
| none |
consts
| 22 | const NX_F32_STK_OK: nx_int = 0 |
| 23 | const NX_F32_STK_ERR_BAD_DIM: nx_int = 1 |
| 24 | const NX_F32_STK_ERR_NULL: nx_int = 2 |
| 25 | const NX_F32_STK_ERR_LAYER: nx_int = 3 |
| 26 | const NX_F32_STK_N_VERDICTS: nx_int = 4 |
functions
| 28 | func nx_f32_stk_verdict_is_valid(v: nx_int) -> nx_int called by 1: main |
| 36 | func nx_f32_llama_stack_forward( called by 2: mainnx_f32_llm_forward calls 3: sys_mmapnx_f32_llama_block_forward_v3nx_f32_kv_cache_advance |