Skip to content

Run frontier MoE models on hardware you already own — pure C, zero deps, experts streamed from disk. Tiny engine, immense model. 🐦

C

Merge dev: keep dev's build-flags stamp (#1707) and serve_budget.h (#1712) next to this branch's header

JustVugg77e6a0deca55
diff --cc c/Makefile
index 749737ca,cc740262..f56c08d7
--- a/c/Makefile
+++ b/c/Makefile
@@@ -1249,10 -1250,10 +1257,10 @@@ NOCUDA_LDFLAGS = $(filter-out -lcudart
# GLM-5.3-Flash: routed experts stream from the int4-gs64 container.
# METAL=1 accelerates resident matrices and routed MoE; CPU remains fallback.
-glm53$(EXE): glm53.c decode_batch.h pin_pool.h cli_args.h st.h json.h stop_ids.h tok.h tok_unicode.h tok_unicode_o200k.h compat.h omp_tune.h serve_poll.h route_trace.h quant.h idot.h hyper_connections.h delta_attention.h sparse_index.h vision_tower.h backend_metal.h backend_vulkan.h edge_adapter_internal.h edge_adapters.h edge_runtime.h segment_adapter_internal.h segment_adapters.h segment_runtime.h $(METAL_OBJ) $(VK_OBJ) $(VK_SPV)
+glm53$(EXE): glm53.c decode_batch.h pin_pool.h cli_args.h st.h json.h stop_ids.h tok.h tok_unicode.h tok_unicode_o200k.h compat.h omp_tune.h serve_poll.h route_trace.h quant.h fp8_format.h idot.h hyper_connections.h delta_attention.h sparse_index.h vision_tower.h backend_metal.h backend_vulkan.h edge_adapter_internal.h edge_adapters.h edge_runtime.h segment_adapter_internal.h segment_adapters.h segment_runtime.h $(METAL_OBJ) $(VK_OBJ) $(VK_SPV)
$(CC) $(CFLAGS) glm53.c $(METAL_OBJ) $(VK_OBJ) -o glm53$(EXE) $(LDFLAGS)
- kimi_k3$(EXE): kimi_k3.c cli_args.h st.h json.h tok.h tok_unicode.h tok_unicode_o200k.h compat.h quant.h fp8_format.h idot.h omp_tune.h route_trace.h kv_prefix.h pin_pool.h serve_codec.h backend_cuda.h backend_metal.h backend_vulkan.h edge_adapters.h edge_runtime.h edge_tok_internal.h hybrid_split.h segment_adapter_internal.h segment_adapters.h segment_runtime.h $(CUDA_OBJ) $(VK_OBJ) $(VK_SPV) $(METAL_OBJ)
-kimi_k3$(EXE): kimi_k3.c cli_args.h st.h json.h tok.h tok_unicode.h tok_unicode_o200k.h compat.h quant.h idot.h omp_tune.h route_trace.h kv_prefix.h pin_pool.h serve_codec.h serve_budget.h backend_cuda.h backend_metal.h backend_vulkan.h edge_adapters.h edge_runtime.h edge_tok_internal.h hybrid_split.h segment_adapter_internal.h segment_adapters.h segment_runtime.h $(CUDA_OBJ) $(VK_OBJ) $(VK_SPV) $(METAL_OBJ)
++kimi_k3$(EXE): kimi_k3.c serve_budget.h cli_args.h st.h json.h tok.h tok_unicode.h tok_unicode_o200k.h compat.h quant.h fp8_format.h idot.h omp_tune.h route_trace.h kv_prefix.h pin_pool.h serve_codec.h backend_cuda.h backend_metal.h backend_vulkan.h edge_adapters.h edge_runtime.h edge_tok_internal.h hybrid_split.h segment_adapter_internal.h segment_adapters.h segment_runtime.h $(CUDA_OBJ) $(VK_OBJ) $(VK_SPV) $(METAL_OBJ)
$(CC) $(CFLAGS) kimi_k3.c $(CUDA_OBJ) $(VK_OBJ) $(METAL_OBJ) -o kimi_k3$(EXE) $(LDFLAGS)
# Use a baseline that matches the compiler target. macOS already targets a
@@@ -1448,13 -1452,13 +1459,13 @@@ tests/bench_inkling_shared_batch$(EXE)
tests/test_inkling_cache_index$(EXE): tests/test_inkling_cache_index.c inkling.c st.h json.h tok.h tok_unicode.h tok_unicode_o200k.h compat.h omp_tune.h route_trace.h kv_prefix.h serve_codec.h $(INK_CUDA_OBJ) $(METAL_OBJ)
$(CC) $(CFLAGS) $< $(INK_CUDA_OBJ) $(METAL_OBJ) -o $@ $(LDFLAGS)
- tests/test_kimi_serve_framing$(EXE): tests/test_kimi_serve_framing.c kimi_k3.c kv_prefix.h serve_codec.h st.h json.h tok.h tok_unicode.h tok_unicode_o200k.h compat.h quant.h fp8_format.h idot.h omp_tune.h route_trace.h
-tests/test_kimi_serve_framing$(EXE): tests/test_kimi_serve_framing.c kimi_k3.c kv_prefix.h serve_codec.h serve_budget.h st.h json.h tok.h tok_unicode.h tok_unicode_o200k.h compat.h quant.h idot.h omp_tune.h route_trace.h
++tests/test_kimi_serve_framing$(EXE): tests/test_kimi_serve_framing.c serve_budget.h kimi_k3.c kv_prefix.h serve_codec.h st.h json.h tok.h tok_unicode.h tok_unicode_o200k.h compat.h quant.h fp8_format.h idot.h omp_tune.h route_trace.h
$(CC) $(NOCUDA_CFLAGS) $< $(VK_OBJ) -o $@ $(NOCUDA_LDFLAGS)
-tests/test_kimi_cache_index$(EXE): tests/test_kimi_cache_index.c kimi_k3.c kv_prefix.h serve_codec.h st.h json.h tok.h tok_unicode.h tok_unicode_o200k.h compat.h quant.h idot.h omp_tune.h route_trace.h
+tests/test_kimi_cache_index$(EXE): tests/test_kimi_cache_index.c kimi_k3.c kv_prefix.h serve_codec.h st.h json.h tok.h tok_unicode.h tok_unicode_o200k.h compat.h quant.h fp8_format.h idot.h omp_tune.h route_trace.h
$(CC) $(NOCUDA_CFLAGS) $< $(VK_OBJ) -o $@ $(NOCUDA_LDFLAGS)
-tests/test_k3_chat_tools$(EXE): tests/test_k3_chat_tools.c kimi_k3.c kv_prefix.h serve_codec.h st.h json.h tok.h tok_unicode.h tok_unicode_o200k.h compat.h quant.h idot.h omp_tune.h route_trace.h
+tests/test_k3_chat_tools$(EXE): tests/test_k3_chat_tools.c kimi_k3.c kv_prefix.h serve_codec.h st.h json.h tok.h tok_unicode.h tok_unicode_o200k.h compat.h quant.h fp8_format.h idot.h omp_tune.h route_trace.h
$(CC) $(NOCUDA_CFLAGS) $< $(VK_OBJ) -o $@ $(NOCUDA_LDFLAGS)
# olmoe's matmul_q, not colibri's: compares the IDOT path against FP32
diff --cc c/Makefile.deepseek-v4
index 5426cecd,7986c748..97df35f7
--- a/c/Makefile.deepseek-v4
+++ b/c/Makefile.deepseek-v4
@@@ -246,20 -264,20 +264,20 @@@ $(REGISTRY_OBJ): expert_store_registry.
$(CC) $(CFLAGS) -c expert_store_registry.c -o $@
$(TARGET_OBJS) $(TEST_UNIT_OBJS): %.o: deepseek_v4.c deepseek_v4.h \
- deepseek_v4_internal.h deepseek_v4_dspark.inc st.h json.h compat.h tensor.h quant.h \
+ deepseek_v4_internal.h deepseek_v4_dspark.inc st.h json.h compat.h tensor.h quant.h fp8_format.h \
route_trace.h \
native_quant.h native_quant_batch.h native_quant_dual.h \
- native_quant_fp4_rows16.h expert_store_registry.h
+ native_quant_fp4_rows16.h expert_store_registry.h $(V4_FLAGS_STAMP)
$(CC) $(CFLAGS) -D$* -c deepseek_v4.c -o $@
$(V4_HOT_TEST_OBJ): deepseek_v4.c deepseek_v4.h deepseek_v4_internal.h \
- st.h json.h compat.h tensor.h quant.h route_trace.h \
+ st.h json.h compat.h tensor.h quant.h fp8_format.h route_trace.h \
- native_quant.h native_quant_fp4_rows16.h expert_store_registry.h
+ native_quant.h native_quant_fp4_rows16.h expert_store_registry.h $(V4_FLAGS_STAMP)
$(CC) $(CFLAGS) -DCOLI_V4_TEST_HOOKS \
-DCOLI_V4_UNIT_EXPERT_STORE_HOT_ROWS16 -c deepseek_v4.c -o $@
$(V4_BATCH_TEST_OBJ): deepseek_v4.c deepseek_v4.h deepseek_v4_internal.h \
- tensor.h quant.h fp8_format.h native_quant.h native_quant_batch.h
- tensor.h quant.h native_quant.h native_quant_batch.h $(V4_FLAGS_STAMP)
++ tensor.h quant.h fp8_format.h native_quant.h native_quant_batch.h $(V4_FLAGS_STAMP)
$(CC) $(CFLAGS) -DCOLI_V4_TEST_HOOKS \
-DCOLI_V4_UNIT_NATIVE_QUANT_BATCH -c deepseek_v4.c -o $@