GaijinEntertainment · borisbat · Jul 29, 2026 · Jul 27, 2026 · Jul 28, 2026 · Jul 28, 2026
diff --git a/doc/reflections/das2rst.das b/doc/reflections/das2rst.das
@@ -216,6 +216,7 @@ def document_module_fio(_root : string) {
     var mod = [get_module("fio_core"), find_module("fio")]
     var groups <- array<DocGroup>(
         hide_group(group_by_regex("Internal builtin functions", mod, %regex~builtin%%)),
+        group_by_regex("Direct IO and mapping advice", mod, %regex~(dwrite_open|dwrite_append|dwrite_band|dwrite_commit|dwrite_stat|dwrite_close|prefetch_map)$%%),
         group_by_regex("File manipulation", mod, %regex~(f|long_f|stat$|getchar|remove|rename|copy_file|file_size|equivalent|is_symlink|set_mtime)%%),
         group_by_regex("Path manipulation", mod, %regex~(base_name|dir_name|get_full_file_name|extension|stem|replace_extension|path_join|normalize|to_generic_path|is_absolute|relative|relative_result|parent)$%%),
         group_by_regex("Directory manipulation", mod, %regex~(dir|dir_rec|mkdir|mkdir_rec|mkdir_result|rmdir|rmdir_rec|rmdir_rec_result|rmdir_result|chdir|getcwd)$%%),

diff --git a/doc/source/stdlib/handmade/function-fio-dwrite_append-0x6658e660a37b258d.rst b/doc/source/stdlib/handmade/function-fio-dwrite_append-0x6658e660a37b258d.rst
@@ -0,0 +1 @@
+Append ``bytes`` from ``data`` at the writer's current end. Any pointer, any size — the handle re-blocks into aligned bands internally, and a band-multiple, band-aligned source is written directly with no copy. Returns false when a write fails (disk full, in practice); a failed writer stays failed, so checking ``dwrite_close`` once is enough.
diff --git a/doc/source/stdlib/handmade/function-fio-dwrite_band-0x160228ade96e97be.rst b/doc/source/stdlib/handmade/function-fio-dwrite_band-0x160228ade96e97be.rst
@@ -0,0 +1 @@
+Hand out the writer's own staging band so a producer can build bytes in place instead of building a buffer and appending a copy. ``avail`` receives the free space left in the current band; fill up to that many bytes and account for them with ``dwrite_commit``. Returns null on a dead writer.
diff --git a/doc/source/stdlib/handmade/function-fio-dwrite_close-0xca8d8214309caf7.rst b/doc/source/stdlib/handmade/function-fio-dwrite_close-0xca8d8214309caf7.rst
@@ -0,0 +1 @@
+Flush the writer's tail (sector-padded on cache-bypassing platforms), truncate the file to exactly what was appended, and free the handle. Returns false if anything failed anywhere along the writer's life — this is the one completion verdict a caller needs to check.
diff --git a/doc/source/stdlib/handmade/function-fio-dwrite_commit-0x122d2094a7b99d77.rst b/doc/source/stdlib/handmade/function-fio-dwrite_commit-0x122d2094a7b99d77.rst
@@ -0,0 +1 @@
+Account ``bytes`` written into the pointer ``dwrite_band`` handed out; a full band flushes to the device automatically. Returns false when the commit overruns the band's remaining space (a producer bug) or the writer already failed.
diff --git a/doc/source/stdlib/handmade/function-fio-dwrite_open-0x46f123a0dc4b0ce0.rst b/doc/source/stdlib/handmade/function-fio-dwrite_open-0x46f123a0dc4b0ce0.rst
@@ -0,0 +1 @@
+Open ``path`` for cache-bypassing sequential append (``FILE_FLAG_NO_BUFFERING`` on Windows, ``F_NOCACHE`` on macOS, ``posix_fallocate`` plus ``fadvise`` elsewhere). ``total_bytes`` preallocates the file — writes must then be strictly ascending, which is what avoids NTFS valid-data-length zero-fill; the file is truncated back to what was actually appended at close. ``band_bytes`` sizes the internal sector-aligned bounce buffer (16 MB when in doubt). Returns null on failure. Pair with ``dwrite_append`` or the ``dwrite_band``/``dwrite_commit`` producer surface, and always finish with ``dwrite_close``.
diff --git a/doc/source/stdlib/handmade/function-fio-dwrite_stat-0xbf270ab2cb2df9d6.rst b/doc/source/stdlib/handmade/function-fio-dwrite_stat-0xbf270ab2cb2df9d6.rst
@@ -0,0 +1 @@
+Where the writer spent its time so far: ``which`` 0 = staging nanoseconds (memcpy into the bounce band), 1 = syscall nanoseconds, 2 = bytes written directly from caller memory, 3 = bytes that went through the bounce band. Call before ``dwrite_close`` — the handle is gone after.
diff --git a/doc/source/stdlib/handmade/function-fio-fmap_open_rw-0x2abb9e9d0dfc252e.rst b/doc/source/stdlib/handmade/function-fio-fmap_open_rw-0x2abb9e9d0dfc252e.rst
@@ -0,0 +1 @@
+The writable twin of ``fmap_open``: memory-maps the file at ``path`` as a SHARED read-write view and returns the mapping's base pointer, writing the mapped byte count through ``size``. Writes through the mapping go back to the file itself via the page cache — this is a real shared mapping, not copy-on-write, so no commit charge is taken for the view. The mapping stays valid until passed to ``fmap_close``. Returns null when the file cannot be opened for writing, is empty, or cannot be mapped.
diff --git a/doc/source/stdlib/handmade/function-fio-prefetch_map-0xc63f5d4e91b2e653.rst b/doc/source/stdlib/handmade/function-fio-prefetch_map-0xc63f5d4e91b2e653.rst
@@ -0,0 +1 @@
+Advisory readahead over a mapped range: asks the OS to fault ``bytes`` starting at ``base`` into the page cache ahead of use (``PrefetchVirtualMemory`` on Windows, ``madvise(MADV_WILLNEED)`` on POSIX). Use it on an ``fmap``/``fmap_open`` mapping before a parallel pass over the data — on-demand faults taken inside worker lanes serialize them. Returns false when the OS declines; the call is purely advisory, so reads work either way.
diff --git a/include/daScript/simulate/aot_builtin_fio.h b/include/daScript/simulate/aot_builtin_fio.h
@@ -88,7 +88,15 @@ namespace das {
     DAS_API vec4f builtin_load ( Context & context, SimNode_CallBase *, vec4f * args );
     DAS_API void builtin_map_file ( const FILE* _f, const TBlock<void, TTemporary<TArray<uint8_t>>>& blk, Context*, LineInfoArg * at );
     DAS_API void * builtin_fmap_open ( const char * name, uint64_t * size, Context * context, LineInfoArg * at );
+    DAS_API void * builtin_fmap_open_rw ( const char * name, uint64_t * size, Context * context, LineInfoArg * at );
     DAS_API void builtin_fmap_close ( void * data, uint64_t size, Context * context, LineInfoArg * at );
+    DAS_API void * builtin_dwrite_open ( const char * name, uint64_t total_bytes, uint64_t band_bytes, Context * context, LineInfoArg * at );
+    DAS_API bool builtin_dwrite_append ( void * h, void * data, uint64_t bytes, Context * context, LineInfoArg * at );
+    DAS_API void * builtin_dwrite_band ( void * h, uint64_t * avail, Context * context, LineInfoArg * at );
+    DAS_API bool builtin_dwrite_commit ( void * h, uint64_t bytes, Context * context, LineInfoArg * at );
+    DAS_API uint64_t builtin_dwrite_stat ( void * h, int32_t which, Context * context, LineInfoArg * at );
+    DAS_API bool builtin_dwrite_close ( void * h, Context * context, LineInfoArg * at );
+    DAS_API bool builtin_prefetch_map ( void * base, uint64_t bytes, Context * context, LineInfoArg * at );
     DAS_API char * builtin_dirname ( const char * name, Context * context, LineInfoArg * at );
     DAS_API char * builtin_basename ( const char * name, Context * context, LineInfoArg * at );
     DAS_API bool builtin_fstat ( const FILE * f, FStat & fs, Context * context, LineInfoArg * at );

diff --git a/modules/dasLLAMA/.das_module b/modules/dasLLAMA/.das_module
@@ -4,7 +4,7 @@ require daslib/fio
 [export]
 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_env", "dasllama_config", "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_shapes", "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_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"

diff --git a/modules/dasLLAMA/ENVIRONMENT.md b/modules/dasLLAMA/ENVIRONMENT.md
@@ -46,6 +46,7 @@ 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_PREFETCH` | flag | on | Advisory source-mapping readahead at gguf load (cold-conversion fix); =0 restores on-demand faulting. |
 
 ## Metal backend
 
@@ -95,6 +96,8 @@ Vulkan GPU backend. Present only where the dasVulkan package is installed.
 | `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_IMPORT` | flag | on | Stream mirrors import the mapped .dlim (VK_EXT_external_memory_host) instead of pinned copies; =0 restores the copy path. |
+| `DASLLAMA_TRIM` | flag | off | Serve from P3-trimmed vulkan images (big CPU weight families dropped; folded into the flavor identity). |
 | `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. |
@@ -187,6 +190,7 @@ Apple Accelerate / AMX float lane. `DASLLAMA_ACCEL` arms the whole group.
 | `DASLLAMA_WHISPER_DIR` | path | unset | Directory of whisper models for the audio tests. |
 | `DASLLAMA_CORPUS_DIR` | path | unset | Directory of audio corpus files for the transcription tests. |
 | `TMPDIR` | path | /tmp | Scratch directory for test artifacts; set by the OS on macOS. |
+| `TEMP` | path | unset | Windows scratch-directory fallback when TMPDIR is unset (the test runner's log dir). |
 
 ## Examples and tutorials
Original file line number	Diff line number	Diff line change
		@@ -0,0 +1 @@
		Append ``bytes`` from ``data`` at the writer's current end. Any pointer, any size — the handle re-blocks into aligned bands internally, and a band-multiple, band-aligned source is written directly with no copy. Returns false when a write fails (disk full, in practice); a failed writer stays failed, so checking ``dwrite_close`` once is enough.
Original file line number	Diff line number	Diff line change
		@@ -0,0 +1 @@
		Hand out the writer's own staging band so a producer can build bytes in place instead of building a buffer and appending a copy. ``avail`` receives the free space left in the current band; fill up to that many bytes and account for them with ``dwrite_commit``. Returns null on a dead writer.
Original file line number	Diff line number	Diff line change
		@@ -0,0 +1 @@
		Flush the writer's tail (sector-padded on cache-bypassing platforms), truncate the file to exactly what was appended, and free the handle. Returns false if anything failed anywhere along the writer's life — this is the one completion verdict a caller needs to check.
Original file line number	Diff line number	Diff line change
		@@ -0,0 +1 @@
		Account ``bytes`` written into the pointer ``dwrite_band`` handed out; a full band flushes to the device automatically. Returns false when the commit overruns the band's remaining space (a producer bug) or the writer already failed.
Original file line number	Diff line number	Diff line change
		@@ -0,0 +1 @@
		Open ``path`` for cache-bypassing sequential append (``FILE_FLAG_NO_BUFFERING`` on Windows, ``F_NOCACHE`` on macOS, ``posix_fallocate`` plus ``fadvise`` elsewhere). ``total_bytes`` preallocates the file — writes must then be strictly ascending, which is what avoids NTFS valid-data-length zero-fill; the file is truncated back to what was actually appended at close. ``band_bytes`` sizes the internal sector-aligned bounce buffer (16 MB when in doubt). Returns null on failure. Pair with ``dwrite_append`` or the ``dwrite_band``/``dwrite_commit`` producer surface, and always finish with ``dwrite_close``.
Original file line number	Diff line number	Diff line change
		@@ -0,0 +1 @@
		Where the writer spent its time so far: ``which`` 0 = staging nanoseconds (memcpy into the bounce band), 1 = syscall nanoseconds, 2 = bytes written directly from caller memory, 3 = bytes that went through the bounce band. Call before ``dwrite_close`` — the handle is gone after.
Original file line number	Diff line number	Diff line change
		@@ -0,0 +1 @@
		The writable twin of ``fmap_open``: memory-maps the file at ``path`` as a SHARED read-write view and returns the mapping's base pointer, writing the mapped byte count through ``size``. Writes through the mapping go back to the file itself via the page cache — this is a real shared mapping, not copy-on-write, so no commit charge is taken for the view. The mapping stays valid until passed to ``fmap_close``. Returns null when the file cannot be opened for writing, is empty, or cannot be mapped.
Original file line number	Diff line number	Diff line change
		@@ -0,0 +1 @@
		Advisory readahead over a mapped range: asks the OS to fault ``bytes`` starting at ``base`` into the page cache ahead of use (``PrefetchVirtualMemory`` on Windows, ``madvise(MADV_WILLNEED)`` on POSIX). Use it on an ``fmap``/``fmap_open`` mapping before a parallel pass over the data — on-demand faults taken inside worker lanes serialize them. Returns false when the OS declines; the call is purely advisory, so reads work either way.