Run frontier MoE models on hardware you already own — pure C, zero deps, experts streamed from disk. Tiny engine, immense model. 🐦
C
Merge remote-tracking branch 'origin/dev' into feat/qwen-image
JustVuggce87ff11622e
diff --cc c/Makefileindex 9ae57efd,bfa43574..681158e4--- a/c/Makefile+++ b/c/Makefile@@@ -1185,21 -1195,9 +1195,21 @@@ qwen36: qwen36$(EXEendif# Rebuild when CUDA_DLL changes; otherwise an existing CPU-only executable can# be reported as up to date despite selecting the Windows CUDA DLL tier.- qwen36$(EXE): qwen36.c sse41_kernels.h decode_batch.h serve_poll.h cli_args.h qwen36_tier.h expert_ffn.h idot.h st.h json.h compat.h omp_tune.h kv_prefix.h pin_pool.h edge_adapter_internal.h edge_adapters.h edge_runtime.h segment_adapter_internal.h segment_adapters.h segment_runtime.h gsgemv.h qgemv.h $(QWEN36_TIER_SRC) $(CUDA_OBJ) .build-config+ qwen36$(EXE): qwen36.c qwen38_vision.h sse41_kernels.h decode_batch.h serve_poll.h cli_args.h qwen36_tier.h expert_ffn.h idot.h st.h json.h compat.h omp_tune.h kv_prefix.h pin_pool.h edge_adapter_internal.h edge_adapters.h edge_runtime.h segment_adapter_internal.h segment_adapters.h segment_runtime.h gsgemv.h qgemv.h $(QWEN36_TIER_SRC) $(CUDA_OBJ) .build-config$(CC) $(QWEN36_CFLAGS) qwen36.c $(QWEN36_TIER_SRC) $(CUDA_OBJ) -o qwen36$(EXE) $(QWEN36_LDFLAGS)+# Qwen-Image-2.1 text-to-image (docs/qwen-image.md). It reads the diffusers+# checkpoint as it is: no conversion step. qi_gemm.h is its one matrix kernel.+ifneq ($(EXE),)+.PHONY: qwenimage+qwenimage: qwenimage$(EXE)+endif+qwenimage$(EXE): qwenimage.c qi_gemm.h qwenimage_vae.h qwenimage_preview.h st.h json.h tok.h tok_unicode.h tok_unicode_o200k.h tok_unicode_deepseek.h qwen38_nfc.h qwen38_nfc_tables.h serve_poll.h compat.h .build-config+ $(CC) $(CFLAGS) qwenimage.c -o qwenimage$(EXE) $(LDFLAGS)++test_qi_gemm$(EXE): tests/test_qi_gemm.c qi_gemm.h+ $(CC) $(CFLAGS) tests/test_qi_gemm.c -o test_qi_gemm$(EXE) $(LDFLAGS)+# 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.diff --cc c/coliindex 57ce3f77,a91d23f1..03b00f20--- a/c/coli+++ b/c/coli@@@ -2437,28 -1867,12 +2439,28 @@@ def cmd_chat(a)base=getattr(a,"attach",None) or "http://127.0.0.1:8000"mid=server_probe(base, getattr(a,"api_key",None))if mid:- banner(f"chat · {mid} · attached")- chat_attached(a, base, mid); return- if getattr(a,"attach",None):+ entry=server_model_entry(base, getattr(a,"api_key",None))+ serves_images="image_generation" in (entry.get("capabilities") or [])+ wants_images=bool(a.model) and is_image_model(a.model)+ # A server of the other kind is not the model asked for: an image+ # --model must not end up chatting with a text server (or the other+ # way round) just because something listens on the port. An+ # explicit --attach still means "that server, whatever it is".+ if getattr(a,"attach",None) or not a.model or serves_images==wants_images:+ if serves_images:+ banner(f"images · {mid} · attached")+ image_repl(a, RemoteImageBackend(base, mid, getattr(a,"api_key",None),+ entry.get("image")), mid)+ print(f" {C.dim}goodbye: the engine keeps running for the next session{C.r}")+ return+ banner(f"chat · {mid} · attached")+ chat_attached(a, base, mid); return+ print(f" {C.dim}{base} serves {mid}, not an "+ f"{'image' if wants_images else 'text'} model: starting a private engine{C.r}")+ elif getattr(a,"attach",None):sys.exit(f"--attach: no coli serve answering at {base} (start one with: coli serve --model <dir>)")need_model(a.model)- family=resolve_model(a.model).descriptor+ resolved=resolve_model(a.model); family=resolved.descriptorarch=family.idif not family.has_gateway_adapter:sys.exit(f"{family.display_name}: gateway adapter is not wired")diff --cc c/openai_server.pyindex b9710c7e,a7ccf782..703c44ff--- a/c/openai_server.py+++ b/c/openai_server.py@@@ -22,9 -22,9 +22,10 @@@ import uuiimport v4_dsml # vendored DeepSeek V4 DSML reference primitivesimport v41_dsml # ...and V4.1's, whose tag names differ by a space+import image_engine # the qwenimage serve protocol, PNG and request rulesfrom family_registry import (FamilyConfigError, UnknownFamilyError, family_by_id,family_ids, resolve_model)+ from family_registry import default_model_id as registry_default_model_idfrom http.server import BaseHTTPRequestHandler, ThreadingHTTPServerfrom pathlib import Pathfrom urllib.parse import unquote, urlsplit