release: merge dev → main (proposed v1.2.0)

.dockerignore

@@ -0,0 +1,20 @@
+# The Docker build only needs c/ (engine sources + launcher). Everything else
+# is either huge, host-specific, or irrelevant to the image.
+.git
+web/node_modules
+web/dist
+site
+docs
+**/*.o
+**/*.exe
+c/colibri
+c/glm
+c/olmoe
+c/glm_tiny
+**/.coli_usage
+**/.coli_kv*
+**/__pycache__
+*.safetensors
+*.log
+CODEX-BRIEF.md
+wt-inc4

.github/workflows/ci.yml

@@ -12,8 +12,8 @@ jobs:
     runs-on: ubuntu-latest
     steps:
       - uses: actions/checkout@v4
-      - name: Build colibri
-        run: cd c && make colibri
+      - name: Build colibri + inkling
+        run: cd c && make colibri inkling
       - name: C test suite
         run: cd c && make test-c
 
@@ -40,6 +40,38 @@ jobs:
           nvcc -O2 -std=c++17 -arch=sm_80 -c backend_cuda.cu -o /dev/null \
             -Xcompiler=-Wall,-Wextra
           echo "CUDA syntax check passed"
+      - name: Compile backend_cuda_ink.cu (syntax only, no GPU)
+        run: |
+          cd c
+          # inkling's own CUDA backend (bf16 residents + int4 expert GEMM); same
+          # no-pipeline rule as above so a compile error can actually fail the job.
+          nvcc -O2 -std=c++17 -arch=sm_80 -c backend_cuda_ink.cu -o /dev/null \
+            -Xcompiler=-Wall,-Wextra
+          echo "inkling CUDA syntax check passed"
+
+  inkling-oracle:
+    name: Inkling oracle (token-exact vs transformers)
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@v4
+      - uses: actions/setup-python@v5
+        with:
+          python-version: '3.12'
+          cache: pip
+          cache-dependency-path: c/tools/oracle-requirements.txt
+      - name: Install torch (CPU) + transformers
+        run: pip install -r c/tools/oracle-requirements.txt
+      - name: Build inkling
+        run: cd c && make inkling
+      - name: Tiny-model fixture + token-exact oracle (f32, bit-exact experts)
+        run: |
+          cd c
+          # Random-init InklingForCausalLM via HF transformers, then the C engine
+          # reproduces it token-for-token. bits=0 keeps experts f32 (bit-exact
+          # forward); the greedy pass exits non-zero on any mismatch, so this
+          # gate fails the moment a shared-code change regresses the engine.
+          python3 tools/make_tiny_inkling.py tiny_inkling
+          SNAP=tiny_inkling ./inkling 8 0 tiny_inkling/ref_inkling.json
 
   engine-hip-syntax:
     name: HIP syntax check
@@ -58,8 +90,23 @@ jobs:
           # link of the test binary below covers the single-source claim, and
           # the engine link is exercised on real ROCm hardware (GPU_BACKENDS.md).
           docker run --rm -v "$PWD":/src -w /src rocm/dev-ubuntu-24.04:6.2 bash -c '
-            apt-get update && apt-get install -y --no-install-recommends make gcc g++ &&
-            make -C c gpu-compile HIP=1 HIP_ARCH=gfx1100'
+            apt-get update && apt-get install -y --no-install-recommends make gcc g++ rocwmma-dev binutils &&
+            make -C c gpu-compile HIP=1 HIP_ARCH=gfx1100 &&
+            # The engine link is skipped here (see above), so assert the property
+            # that link needs instead: backend_cuda.o must be position-independent.
+            # hipcc emits a non-PIC/PIE object unless told otherwise, while gcc
+            # links -pie by default on the distros users build on, so absolute
+            # 32-bit relocations here mean `make HIP=1 colibri` dies with
+            #   R_X86_64_32 ... can not be used when making a PIE object
+            # That is exactly what happened before -fPIE was added to HIPCCFLAGS:
+            # this job stayed green over a HIP build that could not link at all,
+            # because gpu-compile links the TEST binary with the hipcc driver
+            # (PIC-agnostic) and never runs the gcc engine link. One readelf keeps
+            # that regression from silently coming back.
+            if readelf -r c/backend_cuda.o | grep -qE "R_X86_64_32S?"; then
+              echo "backend_cuda.o has absolute 32-bit relocations: the PIE link of colibri will fail (HIPCCFLAGS needs -fPIE or -fPIC)" >&2
+              exit 1
+            fi'
           echo "HIP syntax check passed"
 
   windows-cuda-build:

.github/workflows/release.yml

@@ -55,6 +55,26 @@ jobs:
           make colibri ${{ matrix.make_args }}
           ls -lh colibri${{ matrix.ext }}
 
+      - uses: actions/setup-node@v4
+        with:
+          node-version: '20'
+          cache: npm
+          cache-dependency-path: web/package-lock.json
+
+      # The dashboard is what README shows and what serves the Expert Atlas, but it
+      # was never packaged: a release user has no web/ to build from, so `coli web`
+      # pointed them at a directory the archive does not contain. Built here rather
+      # than committed, so dist/ stays a build artifact.
+      # shell: bash — on Windows the job default is msys2, but setup-node installs
+      # the native node/npm, which msys2's minimal PATH does not see.
+      - name: Build web dashboard
+        shell: bash
+        run: |
+          cd web
+          npm ci
+          npm run build
+          ls -lh dist/index.html dist/experts.json
+
       - name: Package
         run: |
           TAG=${GITHUB_REF#refs/tags/}
@@ -71,6 +91,10 @@ jobs:
           cp c/resource_plan.py dist/
           cp c/doctor.py dist/
           cp LICENSE dist/
+          # web/dist sits NEXT TO coli in the archive; both coli and openai_server.py
+          # probe that layout as well as the source checkout's one-level-up form.
+          mkdir -p dist/web
+          cp -r web/dist dist/web/dist
           cd dist
           if [ "${{ matrix.ext }}" = ".exe" ]; then
             7z a colibri-${TAG}-${{ matrix.name }}.zip *
@@ -100,6 +124,12 @@ jobs:
           # (The v1.1.0 tag build failed here when the banner still said "colibrì".)
           echo "$out" | grep -q "colibri" || { echo "FAIL: coli did not run"; exit 1; }
           echo "OK: coli locates the engine in the published archive"
+          # Same reasoning for the dashboard: assert the files are there AND that the
+          # server resolves them, since a right-looking archive with an unresolvable
+          # path is exactly the failure this whole step exists to catch.
+          test -f web/dist/index.html   || { echo "FAIL: dashboard missing from archive"; exit 1; }
+          test -f web/dist/experts.json || { echo "FAIL: expert atlas missing from archive"; exit 1; }
+          python3 -c "import sys; sys.path.insert(0, '.'); from openai_server import APIHandler as A; d = A.WEB_DIST; assert (d / 'index.html').is_file(), 'server resolved WEB_DIST to %s, which has no index.html' % d; assert (d / 'experts.json').is_file(), 'no expert atlas under %s' % d; print('OK: the packaged server resolves the dashboard at', d)"
 
       # Only the archives -- never the loose files. `dist/colibri-*.*` used to work by accident
       # (the engine was versioned, so it did not match); now that the engine is plainly named,

.gitignore

@@ -15,6 +15,7 @@ c/colibri.exe
 c/glm
 c/glm.exe
 c/olmoe
+c/inkling
 c/olmoe.exe
 c/iobench
 c/iobench.exe
@@ -68,3 +69,4 @@ olmoe_merged/
 olmoe_i4/
 c/olmoe_merged/
 c/olmoe_i4/
+c/tiny_inkling/

README.md

@@ -263,6 +263,7 @@ the full 756 GB on disk at once:
 COLI_MODEL=/nvme/glm52_i4 ./coli chat     # RAM budget, cache and MTP auto-detected
 COLI_MODEL=/nvme/glm52_i4 ./coli plan     # inspect the planned VRAM/RAM/disk placement
 COLI_MODEL=/nvme/glm52_i4 ./coli doctor   # read-only readiness check
+COLI_MODEL=/nvme/glm52_i4 ./coli doctor --deep  # strict tensors/shards/index/mirror preflight
 ./coli web  --model /nvme/glm52_i4        # API + web dashboard on one port
 ./coli serve --model /nvme/glm52_i4       # OpenAI-compatible API only
 ```

c/.gitignore

@@ -5,3 +5,6 @@ olmoe_i4/
 .build-config
 tests/test_st_mirror
 tests/test_st_pread
+tools/libiq3.so
+tools/libiq3.dylib
+tools/iq3.dll

c/Makefile

@@ -183,7 +183,7 @@ HIP       ?= 0
 ROCM_HOME ?= /opt/rocm
 HIPCC     ?= $(ROCM_HOME)/bin/hipcc
 HIP_ARCH  ?= native
-HIPCCFLAGS ?= -O3 -std=c++17 -x hip --offload-arch=$(HIP_ARCH) -Wall -Wextra
+HIPCCFLAGS ?= -O3 -std=c++17 -x hip --offload-arch=$(HIP_ARCH) -Wall -Wextra -fPIE
 GPUCC      = $(NVCC)
 GPUFLAGS   = $(NVCCFLAGS)
 GPUCC_NAME = nvcc (set CUDA_HOME or NVCC)
@@ -199,7 +199,7 @@ else
 PYTHON    ?= python3
 endif
 CUDA_OBJ  =
-TEST_BINS = tests/test_json$(EXE) tests/test_st$(EXE) tests/test_st_mirror$(EXE) tests/test_st_pread$(EXE) tests/test_tier$(EXE) tests/test_grammar$(EXE) tests/test_schema_gbnf$(EXE) tests/test_decode_batch$(EXE) tests/test_idot$(EXE) tests/test_i4_grouped$(EXE) tests/test_stops$(EXE) tests/test_topp$(EXE) tests/test_sample_nan$(EXE) tests/test_temp_env$(EXE) tests/test_tok_o200k$(EXE) tests/test_kv_alloc$(EXE) tests/test_int3$(EXE) tests/test_int3_load$(EXE) tests/test_i4_acc512$(EXE) tests/test_compat_direct$(EXE) tests/test_dsa_select$(EXE) tests/test_logit_nan$(EXE) tests/test_pipe_block$(EXE) tests/test_e8_kernel$(EXE)
+TEST_BINS = tests/test_json$(EXE) tests/test_st$(EXE) tests/test_st_mirror$(EXE) tests/test_st_pread$(EXE) tests/test_tier$(EXE) tests/test_grammar$(EXE) tests/test_schema_gbnf$(EXE) tests/test_decode_batch$(EXE) tests/test_idot$(EXE) tests/test_i4_grouped$(EXE) tests/test_stops$(EXE) tests/test_topp$(EXE) tests/test_sample_nan$(EXE) tests/test_temp_env$(EXE) tests/test_tok_o200k$(EXE) tests/test_kv_alloc$(EXE) tests/test_int3$(EXE) tests/test_int3_load$(EXE) tests/test_i4_acc512$(EXE) tests/test_compat_direct$(EXE) tests/test_dsa_select$(EXE) tests/test_logit_nan$(EXE) tests/test_router_nan$(EXE) tests/test_pipe_block$(EXE) tests/test_e8_kernel$(EXE)
 ifneq (,$(LINUX))
 TEST_BINS += tests/test_uring$(EXE)
 endif
@@ -227,6 +227,7 @@ endif
 CFLAGS   += -DCOLI_CUDA
 LDFLAGS  += -L$(CUDA_HOME)/lib64 -Wl,-rpath,$(CUDA_HOME)/lib64 -lcudart -lstdc++
 CUDA_OBJ  = backend_cuda.o
+INK_CUDA_OBJ = backend_cuda_ink.o
 endif
 
 ifeq ($(HIP),1)
@@ -300,6 +301,10 @@ cuda-dll: backend_cuda.cu backend_cuda.h
 		-L"$(CUDA_HOME)/lib/x64" -lcudart \
 		backend_cuda.cu -o coli_cuda.dll
 
+backend_cuda_ink.o: backend_cuda_ink.cu backend_cuda_ink.h .build-config
+	@command -v "$(NVCC)" >/dev/null 2>&1 || { echo "nvcc not found: set CUDA_HOME or NVCC" >&2; exit 1; }
+	"$(NVCC)" $(NVCCFLAGS) -c backend_cuda_ink.cu -o $@
+
 backend_cuda.o: backend_cuda.cu backend_cuda.h backend_gpu_compat.h .build-config
 	@command -v "$(GPUCC)" >/dev/null 2>&1 || { echo "$(GPUCC_NAME) not found" >&2; exit 1; }
 	"$(GPUCC)" $(GPUFLAGS) -c backend_cuda.cu -o $@
@@ -311,6 +316,27 @@ metal-test: tests/test_backend_metal.mm backend_metal.mm backend_metal.h
 	$(METALXX) tests/test_backend_metal.mm backend_metal.mm -framework Metal -framework Foundation -o backend_metal_test
 	./backend_metal_test
 
+# Standalone large-batch GEMM correctness sweep — reproduces the long-context prefill
+# corruption in seconds (no model). OMP-parallel CPU reference.
+gemm-test: tests/test_gemm_largebatch.mm backend_metal.mm backend_metal.h
+	$(METALXX) $(OMPC) tests/test_gemm_largebatch.mm backend_metal.mm $(OMPL) -framework Metal -framework Foundation -o gemm_largebatch_test
+	./gemm_largebatch_test
+
+# fmt=6 (E8/IQ3) encoder for tools/convert_fp8_to_int4.py. OPTIONAL: iq3_pack.py
+# falls back to its numpy path when the library is absent, producing the same
+# bytes ~13x slower. Uses the same ARCH as the engine, so an AVX2 host gets the
+# SIMD search and anything else compiles the scalar fallback.
+ifneq (,$(DARWIN))
+IQ3LIB = tools/libiq3.dylib
+else ifneq ($(IS_WIN),)
+IQ3LIB = tools/iq3.dll
+else
+IQ3LIB = tools/libiq3.so
+endif
+iq3: $(IQ3LIB)
+$(IQ3LIB): tools/iq3_encode.c
+	$(CC) $(CFLAGS) -fPIC -shared $< -o $@ $(LDFLAGS)
+
 cuda-test: backend_cuda.cu backend_cuda.h backend_gpu_compat.h tests/test_backend_cuda.cu
 	@command -v "$(GPUCC)" >/dev/null 2>&1 || { echo "$(GPUCC_NAME) not found" >&2; exit 1; }
 	"$(GPUCC)" $(GPUFLAGS) backend_cuda.cu tests/test_backend_cuda.cu -o backend_cuda_test$(EXE)
@@ -333,6 +359,9 @@ cuda-bench: backend_cuda.cu backend_cuda.h backend_gpu_compat.h tests/bench_tens
 olmoe$(EXE): olmoe.c st.h json.h compat.h
 	$(CC) $(CFLAGS) olmoe.c -o olmoe$(EXE) $(LDFLAGS)
 
+inkling$(EXE): inkling.c st.h json.h compat.h $(INK_CUDA_OBJ)
+	$(CC) $(CFLAGS) inkling.c $(INK_CUDA_OBJ) -o inkling$(EXE) $(LDFLAGS)
+
 # Use a baseline that matches the compiler target. macOS already targets a
 # portable baseline when ARCH is empty; forcing the x86 value there breaks
 # Apple Silicon. Unknown targets use native rather than an invalid x86 flag.
@@ -421,6 +450,9 @@ tests/test_int3_load$(EXE): tests/test_int3_load.c colibri.c st.h uring.h json.h
 tests/test_logit_nan$(EXE): tests/test_logit_nan.c colibri.c st.h uring.h json.h tok.h tok_unicode.h compat.h grammar.h tier.h
 	$(CC) $(CFLAGS) $< -o $@ $(LDFLAGS)
 
+tests/test_router_nan$(EXE): tests/test_router_nan.c colibri.c st.h uring.h json.h tok.h tok_unicode.h compat.h grammar.h tier.h
+	$(CC) $(CFLAGS) $< -o $@ $(LDFLAGS)
+
 tests/test_i4_acc512$(EXE): tests/test_i4_acc512.c
 	$(CC) $(CFLAGS) $< -o $@ $(LDFLAGS)
 
@@ -495,8 +527,17 @@ install: colibri$(EXE) olmoe$(EXE)
 	$(INSTALL) -m 755 coli $(DESTDIR)$(BINDIR)/coli
 	$(INSTALL) -m 755 colibri$(EXE) $(DESTDIR)$(LIBEXECDIR)/colibri$(EXE)
 	$(INSTALL) -m 755 olmoe$(EXE) $(DESTDIR)$(LIBEXECDIR)/olmoe$(EXE)
-	$(INSTALL) -m 644 resource_plan.py doctor.py openai_server.py $(DESTDIR)$(LIBEXECDIR)/
+	$(INSTALL) -m 644 resource_plan.py doctor.py openai_server.py version.py $(DESTDIR)$(LIBEXECDIR)/
 	$(INSTALL) -m 644 tools/*.py $(DESTDIR)$(LIBEXECDIR)/tools/
+	@# The dashboard is an optional build artifact (cd web && npm run build), so install
+	@# it only when it exists. It goes NEXT TO openai_server.py, which probes ./web/dist.
+	@if [ -f ../web/dist/index.html ]; then \
+	  $(INSTALL) -d $(DESTDIR)$(LIBEXECDIR)/web/dist; \
+	  cp -R ../web/dist/. $(DESTDIR)$(LIBEXECDIR)/web/dist/; \
+	  echo "installed the web dashboard"; \
+	else \
+	  echo "web/dist absent -- skipping the dashboard (build it with: cd web && npm run build)"; \
+	fi
 
 uninstall:
 	rm -f $(DESTDIR)$(BINDIR)/coli
@@ -507,4 +548,4 @@ clean:
 
 bench: iobench$(EXE)
 	@if [ -n "$(ARGS)" ]; then ./iobench$(EXE) $(ARGS); else echo "built iobench$(EXE) — run: ./iobench$(EXE) <file> <MB> <iters> <threads> <direct 0|1>"; fi
-.PHONY: all colibri glm cuda-test hip-test gpu-compile cuda-bench cuda-dll portable test-c test-python test check clean install uninstall bench
+.PHONY: all colibri glm iq3 cuda-test hip-test gpu-compile cuda-bench cuda-dll portable test-c test-python test check clean install uninstall bench