Run frontier MoE models on hardware you already own — pure C, zero deps, experts streamed from disk. Tiny engine, immense model. 🐦
C
Merge branch 'dev' into build/makefile-autodeps
Kenneth-Javier Ortigozad00ad2fe9d46
diff --cc c/Makefileindex 033316c9,a86af234..21c5988f--- a/c/Makefile+++ b/c/Makefile@@@ -1256,9 -1232,32 +1256,29 @@@ tests/test_qi_gemm$(EXE): tests/test_qi# DeepSeek V4.1 Flash: one file, like every other portable engine. The fp4 experts# stream from the official checkpoint and quant.h's mxfp4 kernel reads them as they# are, so there is no conversion target here to go with it.-deepseek_v41$(EXE): deepseek_v41.c matmul_f32.h sse41_kernels.h cli_args.h st.h json.h tok.h tok_unicode.h tok_unicode_deepseek.h compat.h omp_tune.h kv_prefix.h pin_pool.h quant.h fp8_format.h idot.h \- sparse_attn.h hyper_connections.h serve_codec.h serve_poll.h .build-config+deepseek_v41$(EXE): deepseek_v41.c .build-config$(CC) $(CFLAGS) deepseek_v41.c -o deepseek_v41$(EXE) $(LDFLAGS)+ # Xiaomi MiMo-V2.6 (Flash 309B, Pro 1.02T): one file plus its ViT header, reading the+ # release as it is -- MXFP4 experts streamed per layer, FP8/BF16 dense resident.+ ifneq ($(EXE),)+ .PHONY: mimo+ mimo: mimo$(EXE)+ endif-mimo$(EXE): mimo.c mimo_vision.h sse41_kernels.h cli_args.h compat.h json.h st.h quant.h idot.h fp8_format.h \- omp_tune.h kv_prefix.h tok.h tok_unicode.h tok_unicode_o200k.h tok_unicode_deepseek.h serve_codec.h \- decode_batch.h serve_poll.h stop_ids.h .build-config++mimo$(EXE): mimo.c .build-config+ $(CC) $(CFLAGS) mimo.c -o mimo$(EXE) $(LDFLAGS)++ # Tiny MiMo fixture + vendor oracle, the Kimi K3 pattern: the generator is numpy (and+ # Pillow for the test picture) and deterministic; ref.json is committed and comes from+ # Xiaomi's own modeling code (tools/make_mimo_ref.py, maintainer-run, SHA-256-pinned),+ # so CI needs neither torch nor network. The fixture carries the vision tower too.+ .PHONY: mimo-tiny-generate mimo-tiny-check+ mimo-tiny-generate:+ $(PYTHON) tools/make_mimo_tiny.py --output ./mimo_tiny --force --vision++ mimo-tiny-check: mimo-tiny-generate mimo$(EXE)+ $(PYTHON) tests/mimo_tiny_harness.py --binary ./mimo$(EXE) --fixture ./mimo_tiny+# Qwen3.8-Flash-Next text-only sibling. qwen38.c owns the direct checkpoint# loader and SERVE=1 protocol. With CUDA=1 it links the same expert tier as# qwen36 (fp8 streaming mode: hot experts get VRAM copies, the RAM LRU stays);@@@ -1486,54 -1474,53 +1506,57 @@@ tests/test_xdna_failure$(EXE): tests/te## M-list defaults to 1,32,64 (the M64 bucket); pass 65,130,256 to qualify# the M256 bucket through the same owner.-tests/xdna_physical_probe$(EXE): tests/xdna_physical_probe.c colibri.c backend_xdna.c backend_xdna.h st.h uring.h json.h tok.h tok_unicode.h tok_unicode_deepseek.h compat.h grammar.h tier.h quant.h sample.h kv_persist.h telemetry.h route_trace.h $(VK_OBJ)- $(CC) $(CFLAGS) -DCOLI_XDNA $< backend_xdna.c $(VK_OBJ) -o $@ $(LDFLAGS)+tests/xdna_physical_probe$(EXE): tests/xdna_physical_probe.c colibri.c tests/backend_xdna_lane.o $(VK_OBJ)+ $(CC) $(CFLAGS) -DCOLI_XDNA $< tests/backend_xdna_lane.o $(VK_OBJ) -o $@ $(LDFLAGS)-tests/test_json$(EXE): tests/test_json.c json.h+tests/test_json$(EXE): tests/test_json.c$(CC) $(CFLAGS) $< -o $@ $(LDFLAGS)-tests/test_tok_o200k$(EXE): tests/test_tok_o200k.c tok.h tok_unicode.h tok_unicode_deepseek.h tok_unicode_o200k.h json.h+tests/test_tok_o200k$(EXE): tests/test_tok_o200k.c$(CC) $(CFLAGS) $< -o $@ $(LDFLAGS)-tests/test_tok_gpt2$(EXE): tests/test_tok_gpt2.c tok.h tok_unicode.h tok_unicode_deepseek.h tok_unicode_o200k.h json.h+tests/test_tok_gpt2$(EXE): tests/test_tok_gpt2.c$(CC) $(CFLAGS) $< -o $@ $(LDFLAGS)-tests/test_tok_deepseek$(EXE): tests/test_tok_deepseek.c tok.h tok_unicode.h tok_unicode_deepseek.h tok_unicode_o200k.h json.h+tests/test_tok_deepseek$(EXE): tests/test_tok_deepseek.c$(CC) $(CFLAGS) $< -o $@ $(LDFLAGS)-tests/test_k3_ram_budget$(EXE): tests/test_k3_ram_budget.c sse41_kernels.h kimi_k3.c st.h tok.h quant.h fp8_format.h idot.h compat.h omp_tune.h route_trace.h kv_prefix.h serve_codec.h $(VK_OBJ)+tests/test_k3_ram_budget$(EXE): tests/test_k3_ram_budget.c kimi_k3.c $(VK_OBJ)$(CC) $(CFLAGS) $< $(VK_OBJ) -o $@ $(LDFLAGS)-tests/test_k3_mmap$(EXE): tests/test_k3_mmap.c sse41_kernels.h kimi_k3.c st.h tok.h quant.h fp8_format.h idot.h compat.h omp_tune.h route_trace.h kv_prefix.h serve_codec.h $(VK_OBJ)+tests/test_k3_mmap$(EXE): tests/test_k3_mmap.c kimi_k3.c $(VK_OBJ)$(CC) $(NOCUDA_CFLAGS) $< $(VK_OBJ) -o $@ $(NOCUDA_LDFLAGS)-tests/test_tok_kimi_tiny$(EXE): tests/test_tok_kimi_tiny.c tok.h tok_unicode.h tok_unicode_deepseek.h tok_unicode_o200k.h json.h+tests/test_tok_kimi_tiny$(EXE): tests/test_tok_kimi_tiny.c$(CC) $(CFLAGS) $< -o $@ $(LDFLAGS)-tests/test_st_pread$(EXE): tests/test_st_pread.c st.h json.h compat.h+tests/test_st_pread$(EXE): tests/test_st_pread.c$(CC) $(CFLAGS) -DST_PREAD_CHUNK=7 $< -o $@ $(LDFLAGS)-tests/test_st_slice$(EXE): tests/test_st_slice.c st.h json.h compat.h+tests/test_st_slice$(EXE): tests/test_st_slice.c$(CC) $(CFLAGS) $< -o $@ $(LDFLAGS)-tests/test_qwen38_tokenizer$(EXE): tests/test_qwen38_tokenizer.c sse41_kernels.h qwen38.c qwen38_core.h kv_prefix.h qwen38_nfc.h qwen38_nfc_tables.h edge_runtime.c edge_runtime.h edge_adapters.h edge_adapter_internal.h st.h json.h compat.h quant.h fp8_format.h idot.h route_trace.h tok_unicode.h tok_unicode_deepseek.h tok_unicode_o200k.h- $(CC) $(NOCUDA_CFLAGS) $< edge_runtime.c -o $@ $(NOCUDA_LDFLAGS)+tests/test_qwen38_tokenizer$(EXE): tests/test_qwen38_tokenizer.c qwen38.c tests/edge_runtime_nocuda.o+ $(CC) $(NOCUDA_CFLAGS) $< tests/edge_runtime_nocuda.o -o $@ $(NOCUDA_LDFLAGS)+# edge_runtime.c as its own object, compiled with the flags this test gave it on+# the same command line before (#1741): one unit per command keeps each .d whole.+tests/edge_runtime_nocuda.o: edge_runtime.c .build-config+ $(CC) $(NOCUDA_CFLAGS) -c edge_runtime.c -o $@-tests/test_qwen38_config$(EXE): tests/test_qwen38_config.c sse41_kernels.h qwen38.c qwen38_core.h kv_prefix.h qwen38_nfc.h qwen38_nfc_tables.h st.h json.h compat.h quant.h fp8_format.h idot.h route_trace.h tok_unicode.h tok_unicode_deepseek.h tok_unicode_o200k.h+tests/test_qwen38_config$(EXE): tests/test_qwen38_config.c qwen38.c$(CC) $(NOCUDA_CFLAGS) $< -o $@ $(NOCUDA_LDFLAGS)-tests/test_dsv41_serve_budget$(EXE): tests/test_dsv41_serve_budget.c sse41_kernels.h deepseek_v41.c cli_args.h st.h json.h tok.h tok_unicode.h tok_unicode_deepseek.h compat.h omp_tune.h kv_prefix.h pin_pool.h quant.h fp8_format.h idot.h sparse_attn.h hyper_connections.h serve_codec.h serve_poll.h+tests/test_dsv41_serve_budget$(EXE): tests/test_dsv41_serve_budget.c deepseek_v41.c$(CC) $(NOCUDA_CFLAGS) $< -o $@ $(NOCUDA_LDFLAGS)-tests/test_dsv41_untrusted_load$(EXE): tests/test_dsv41_untrusted_load.c sse41_kernels.h deepseek_v41.c cli_args.h st.h json.h tok.h tok_unicode.h tok_unicode_deepseek.h compat.h omp_tune.h kv_prefix.h pin_pool.h quant.h fp8_format.h idot.h sparse_attn.h hyper_connections.h serve_codec.h serve_poll.h++tests/test_dsv41_untrusted_load$(EXE): tests/test_dsv41_untrusted_load.c deepseek_v41.c+ $(CC) $(NOCUDA_CFLAGS) $< -o $@ $(NOCUDA_LDFLAGS)+-tests/test_qwen38_idot$(EXE): tests/test_qwen38_idot.c sse41_kernels.h qwen38.c qwen38_core.h kv_prefix.h qwen38_nfc.h qwen38_nfc_tables.h st.h json.h compat.h quant.h fp8_format.h idot.h route_trace.h tok_unicode.h tok_unicode_deepseek.h tok_unicode_o200k.h+tests/test_qwen38_idot$(EXE): tests/test_qwen38_idot.c qwen38.c$(CC) $(NOCUDA_CFLAGS) $< -o $@ $(NOCUDA_LDFLAGS)-tests/test_qwen38_serve_framing$(EXE): tests/test_qwen38_serve_framing.c sse41_kernels.h qwen38.c qwen38_core.h kv_prefix.h qwen38_nfc.h qwen38_nfc_tables.h st.h json.h compat.h quant.h fp8_format.h idot.h route_trace.h serve_codec.h tok_unicode.h tok_unicode_deepseek.h tok_unicode_o200k.h+tests/test_qwen38_serve_framing$(EXE): tests/test_qwen38_serve_framing.c qwen38.c$(CC) $(NOCUDA_CFLAGS) $< -o $@ $(NOCUDA_LDFLAGS)-tests/test_qwen38_vision$(EXE): tests/test_qwen38_vision.c qwen38_vision.h st.h json.h compat.h+tests/test_qwen38_vision$(EXE): tests/test_qwen38_vision.c$(CC) $(CFLAGS) -Wno-unused-function $< -o $@ $(LDFLAGS)# Genera la fixture della torre e la confronta con l'oracolo upstream. Non serve