dasLLAMA vulkan arc: decode-tail fusions, K-quant GEMV, memory-priority shield

CLAUDE.md

@@ -214,9 +214,9 @@ AST types (TypeDecl, Expression, Function, Structure, Enumeration, Variable, Mak
 
 Quick rules:
 - AST nodes have **unique ownership** — don't insert the same pointer into two parents; use `clone_type` / `clone_expression` / `clone_function` / `clone_variable` / `clone_structure` to duplicate.
-- AST pointers use plain `=` and pass-by-value. **No `var inscope`, no `<-`** for them. `var inscope` is only for the residual smart_ptr types.
+- AST pointers use plain `=` and pass-by-value. **No `var inscope`, no `<-`** for them. `var inscope` is for OWNED locals that should finalize at scope exit (smart_ptr types, plain arrays/structs) — never for gc_node AST pointers.
 - Tools that build AST at runtime (outside the compile pipeline) must wrap their scope in `ast_gc_guard() { ... }` from `daslib/ast`, or the leak detector reports `GC APP LEAK` at exit.
-- daslang has garbage collection, but plain `var arr : array<T>` does NOT finalize on scope exit. Either declare with `var inscope` (smart_ptr only), call `delete` explicitly, or move out via `<-`. Per-frame leaks in hot paths usually trace to a local `var arr` never deleted.
+- daslang has garbage collection, but plain `var arr : array<T>` does NOT finalize on scope exit. Either declare with `var inscope` (finalize-at-scope-exit — serves plain arrays/structs as well as the smart_ptr types; Boris-confirmed + compile-verified 2026-07-26), call `delete` explicitly, or move out via `<-`. Per-frame leaks in hot paths usually trace to a local `var arr` never deleted.
 
 Full migration table (when reading older docs that say `var inscope` or `<-` for AST types): **`skills/gc_migration.md`**.
 

modules/dasLLAMA/.das_module

@@ -5,7 +5,7 @@ require daslib/fio
 def initialize(project_path : string) {
     let dasllama_paths = [
         "dasllama_env", "dasllama_par", "dasllama_parity", "dasllama_tune", "dasllama_math", "dasllama_math_default", "dasllama_math_aarch64_neon", "dasllama_math_gen", "dasllama_math_metal", "dasllama_math_vulkan", "dasllama_math_accelerate", "dasllama_kernel_access", "dasllama_vulkan_lens", "dasllama_metal_lens", "dasllama_metal_common", "dasllama_metal_kernels", "dasllama_metal_prefill", "dasllama_metal_llama", "dasllama_gemm_gen", "dasllama_gemm_register", "dasllama_gemm_schema", "dasllama_quant", "dasllama_gguf",
-        "dasllama_unicode", "dasllama_tokenizer", "dasllama_bpe", "dasllama_common", "dasllama_prefix", "dasllama_audio", "dasllama_audio_io", "dasllama_whisper", "dasllama_qwen3a", "dasllama_parakeet", "dasllama_canary", "dasllama_gemma4a", "dasllama_vad", "dasllama_asr",
+        "dasllama_unicode", "dasllama_tokenizer", "dasllama_bpe", "dasllama_common", "dasllama_layout", "dasllama_prefix", "dasllama_audio", "dasllama_audio_io", "dasllama_whisper", "dasllama_qwen3a", "dasllama_parakeet", "dasllama_canary", "dasllama_gemma4a", "dasllama_vad", "dasllama_asr",
         "dasllama_arch_llama", "dasllama_arch_qwen2", "dasllama_arch_qwen3", "dasllama_arch_phi3", "dasllama_arch_gemma2", "dasllama_arch_gemma3", "dasllama_arch_gemma4", "dasllama_arch_qwen2moe", "dasllama_arch_qwen3moe", "dasllama_arch_qwen35", "dasllama_arch_gptoss", "dasllama_arch_mistral3", "dasllama_arch_glm4moe", "dasllama_image",
         "dasllama_transformer", "dasllama_chat", "dasllama"
     ]

modules/dasLLAMA/ENVIRONMENT.md

@@ -44,6 +44,8 @@ Read by the inference engine itself, so these affect any program that loads a mo
 | `DASLLAMA_GPU_COMBINE` | flag | on | Device-side routed MoE combine; 0 falls back to the host combine. |
 | `DASLLAMA_GPU_HEAT` | number | 0 | Expert heat threshold: hold the N hottest experts resident regardless of layer placement. |
 | `DASLLAMA_GPU_PROF` | flag | off | Report lifetime GPU queue submissions (real commands plus staging round-trips). |
+| `DASLLAMA_CONV_PROF` | flag | off | Bucket gguf -> image conversion time by kind over the weight walk; one clock pair per tensor. |
+| `DASLLAMA_ALLOW_INTERP_LOAD` | flag | off | Permit a big gguf load without -jit; the transforms run interpreted, so expect minutes per GB. |
 
 ## Metal backend
 
@@ -90,6 +92,10 @@ Vulkan GPU backend. Present only where the dasVulkan package is installed.
 |---|---|---|---|
 | `DASLLAMA_COOPMAT` | number | auto | Cooperative-matrix mode; the flash-attention twin needs it even when the GEMM runs sdot4. |
 | `DASLLAMA_MM_SMALL` | number | 32 | Small-batch tier: 32 = sdot4 (default, beats both coopmat tiles below the crossover), 64 = coopmat M, 128 = always-L. |
+| `DASLLAMA_MM_SMALLD` | number | 64 | Small-d cutoff routing narrow roles (k/v) to the small tier; widening measured worse, so this is an instrument. |
+| `DASLLAMA_VK_FUSE` | flag | on | Fused decode tail (add+rms+requant, qk-norm+rope); 0 pins the split dispatches for a same-build A/B. |
+| `DASLLAMA_VK_XFERQ` | flag | on | Stream expert uploads on the dedicated transfer queue, overlapped via a timeline semaphore; 0 keeps the single-queue rail. |
+| `DASLLAMA_VK_MEMPRIO` | flag | on | Tag allocations high-priority (VK_EXT_memory_priority) so the driver demotes desktop memory, not ours. |
 | `DASLLAMA_VK_FA` | flag | on | Vulkan flash-attention kernel; 0 falls back to the chunked path. |
 | `DASLLAMA_VK_REBAR` | flag | on | Use a ReBAR device-local host-visible heap when one larger than 1GB is present. |
 | `DASLLAMA_VK_HAZARD_PARANOID` | flag | off | Barrier at every dispatch (correctness bisect). |
@@ -134,6 +140,8 @@ Apple Accelerate / AMX float lane. `DASLLAMA_ACCEL` arms the whole group.
 
 | Variable | Type | Default | Effect |
 |---|---|---|---|
+| `PROBE_PATH` | path | D:/_wcliff.bin | Scratch file the write-cliff probe writes; put it on the drive under test. |
+| `PROBE_GB` | number | 50 | Gigabytes the write-cliff probe writes before reporting. |
 | `DASLLAMA_BENCH_MODEL` | text | tinyllama | Model name for the isolated GEMM bench. |
 | `DASLLAMA_BENCH_NTOK` | number | unset | Token count for the isolated GEMM bench. |
 | `DASLLAMA_BENCH_SKIP_ROWMAJOR` | flag | off | Skip the row-major arm of the isolated GEMM bench. |

modules/dasLLAMA/benchmarks/decode_step_trace.das

@@ -4,6 +4,7 @@ options _jit_fast_math = true   // ggml-parity FP laxity — the CPU prefill/sam
 
 require dasllama/dasllama_transformer   // nolint:STYLE029 — umbrella fires each arch [init] registration; requiring dasllama_common directly would drop them
 require dasllama/dasllama_math   // setup_dasllama_jobque_ + pin_kernel_backend + parallel_argmax
+require dasllama/dasllama_layout   // convert_model_to_metal_blob — the blob-flavor transform
 require ?das_metal dasllama/dasllama_metal_llama   // nolint:STYLE030 — GPU driver half, Apple builds only
 require daslib/jobque_boost
 require daslib/clargs