[ggerganov]

Overview

no_alloc is a load/context option, not a model hyperparameter. Remove llama_hparams::no_alloc and pass it explicitly instead:

• memory module constructors take a bool no_alloc parameter right after the model reference; wrappers forward it to their sub-caches
• llama_memory_params gains a no_alloc field so create_memory threads it without reading it from the model
• add llama_memory_bufs (llama-memory.h/.cpp): owns the {ggml context, backend buffer} pairs of a memory instance and centralizes what every leaf previously duplicated - buffer allocation, the no_alloc dummy-buffer path (size-0 buffer with the tensors attached so the scheduler won't allocate them), buffer clear, and memory_breakdown() (real buffer sizes, or simulated sizes via ggml_backend_alloc_ctx_tensors_from_buft_size when no_alloc). Used by llama_kv_cache, llama_memory_recurrent and llama_dsv4_comp_state
• llama_moe_cache takes no_alloc as a parameter and reuses llama_memory_bufs::alloc_buf for its own buffer layout
• llama_memory_get_buffers() returns an empty vector for a null memory (memory-less archs)

Resolves the TODO added in #31087.

Additional information

• test-llama-archs: 538 passed
• test-llama-context over all 135 dummy models generated by test-llama-archs -o: 135/135 passed
• a no_alloc=true sweep loading every dummy model with LLAMA_LOAD_MODE_NONE: all report a non-zero simulated memory breakdown and only null-base buffers

Requirements

• I have read and agree with the contributing guidelines
• AI usage disclosure: YES
  • Harness: dsh:llama.cpp
  • Hardware: 2x DGX Spark (RDMA, TP=2)
  • Model: ggml-org/GLM-5.3-Flash-GGUF
  • Stats: ◷ 2.7h · ⚙ 6.7m · ↑ 314k · ↓ 139k · ⇈ 123.9 tok/s · ⇊ 17.8 tok/s · ≡ 688 · ⌨ 43