Skip to content
Merged
Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
76 commits
Select commit Hold shift + click to select a range
e25c6de
Add a registry for the 214 environment variables, and check against it
ZacharyZcR Aug 3, 2026
50528cd
Register the six variables dev grew since this branch opened
ZacharyZcR Aug 5, 2026
1133431
Sixteen more, from one day of dev
ZacharyZcR Aug 7, 2026
6522d24
Register environment variables added on dev
ZacharyZcR Aug 12, 2026
8cf9fad
fix(env): resync registry after dev rebase
ZacharyZcR Aug 26, 2026
77408c2
fix(env): cover GLM-5.3 and Qwen3.8
ZacharyZcR Aug 31, 2026
f301bb7
Merge dev and refresh environment registry for current engines
ZacharyZcR Sep 22, 2026
9ad07b2
fix(env): cover engine prefixes and rebuild registry consumers
ZacharyZcR Sep 22, 2026
20951f1
fix(env): scan literal calls across lines without comment false posit…
ZacharyZcR Sep 22, 2026
14d0d73
build: track environment registry in DeepSeek V4 units
ZacharyZcR Sep 22, 2026
061fe47
Honor disabled environment dump flags across engines
ZacharyZcR Sep 22, 2026
8ed0780
Merge latest dev and refresh environment registry
ZacharyZcR Sep 23, 2026
d9af40f
Merge dev and refresh environment registry coverage
ZacharyZcR Sep 27, 2026
a0a16da
同步上游并补齐图像引擎环境变量注册表
ZacharyZcR Sep 30, 2026
4d848a9
fix(api): select the full GLM-5.3 prompt contract from checkpoint met…
rudycelekli Nov 2, 2026
38875c9
Merge upstream/dev into feat/env-registry
ZacharyZcR Nov 2, 2026
499ad84
env: cover mimo, laya and gliner_decide, and what dev read since October
ZacharyZcR Nov 2, 2026
29f1a0f
fix(responses): cancel upstream work when clients disconnect before o…
rudycelekli Nov 2, 2026
62c2c29
fix(responses): preserve clients reading after a write half-close
rudycelekli Nov 2, 2026
b2ab632
fix(ui): cancel confirmed proxy resets and close owned responses
rudycelekli Nov 2, 2026
892b8cb
fix(install): ship chat UI and IQ3 runtime data
benmaster82 Nov 5, 2026
30e4419
test(install): harden UI startup readiness on CI
benmaster82 Nov 5, 2026
d3ce1be
Merge dev and resolve install file list
benmaster82 Nov 5, 2026
b39605b
Update
Marouane-elfadli Nov 5, 2026
9c4f731
update
Marouane-elfadli Nov 5, 2026
0413ed4
update
Marouane-elfadli Nov 5, 2026
63d3bed
Add files via upload
Marouane-elfadli Nov 5, 2026
b39db13
update
Marouane-elfadli Nov 5, 2026
9cfa502
Vulkan: trunk-first VRAM placement for deepseek_v41 (ReBAR)
Nov 6, 2026
05e2c6d
dsv41: the trunk-first VRAM placement experiment record
Nov 6, 2026
d7f3289
Merge remote-tracking branch 'upstream/dev' into repair/pr800
ZacharyZcR Nov 6, 2026
e636f53
Register new upstream engine environment variables
ZacharyZcR Nov 6, 2026
5c45ad5
dsv4_cuda_init: a refused device leaves no stale CUDA error
bobsummerwill Oct 27, 2026
60986bb
native_quant.h: the GGUF helpers are C only
bobsummerwill Oct 27, 2026
d2f33fd
feat(dsv4): distribute CUDA layers and expert caches across devices
ZacharyZcR Sep 27, 2026
ab4064c
feat(dsv4): resident routed experts with a budgeted auto mode
bobsummerwill Nov 4, 2026
a264bfb
feat(dsv4): route and run resident MoE on the device
bobsummerwill Nov 4, 2026
b88831b
perf(dsv4): keep the indexer resident and decode FP4 from a table
bobsummerwill Nov 4, 2026
191c65a
feat(dsv4): run the compressor and indexer preparation on CUDA
bobsummerwill Nov 4, 2026
ae8efd4
perf(dsv4): keep decode output projection intermediates on CUDA
bobsummerwill Nov 4, 2026
c88a7dd
perf(dsv4): exact and batched vocabulary head on CUDA
bobsummerwill Nov 4, 2026
7f7424a
perf(dsv4): keep decode lookups in VRAM and log per-turn stats
bobsummerwill Nov 4, 2026
4891a34
fix(dsv4): per-cache mirror guard, reserve and pin cap
bobsummerwill Nov 4, 2026
decd940
deepseek-v4: peer mirror pool on a second GPU (DSV4_CUDA_PEER_DEVICE)
bobsummerwill Oct 15, 2026
ea6d87f
deepseek-v4: peer compute on the second GPU (DSV4_CUDA_PEER_COMPUTE)
bobsummerwill Oct 15, 2026
a702573
fix(dsv4): latch a CUDA fault off and name decode failures
bobsummerwill Nov 4, 2026
83c2a28
test: stub dsv4_cuda_proj_batch_release for current dev
bobsummerwill Nov 6, 2026
eada1b7
test: stub dsv4_cuda_proj_batch_release for current dev
bobsummerwill Nov 6, 2026
7571363
test: stub dsv4_cuda_proj_batch_release for current dev
bobsummerwill Nov 6, 2026
74bfa09
dsv4: release every touched device on a failed init
bobsummerwill Nov 6, 2026
034e9a3
build: let -MMD own the gpu-devices test header
bobsummerwill Nov 6, 2026
3ecd67e
build: let -MMD own the gpu-devices test header
bobsummerwill Nov 6, 2026
1c3b4ca
build: let -MMD own the gpu-devices test header
bobsummerwill Nov 6, 2026
1d152be
build: drop the hand-listed bank-pair test header
bobsummerwill Nov 6, 2026
e893e54
build: drop the hand-listed bank-pair test header
bobsummerwill Nov 6, 2026
3297adb
build: drop the hand-listed bank-pair test header
bobsummerwill Nov 6, 2026
c34d3a9
dsv4: export dsv4_cuda_wo_decode from the backend
bobsummerwill Nov 6, 2026
96ddc4d
Vulkan: a tier pool's budget is page-granular, so a three-expert tier…
jameswitney Nov 6, 2026
d43520b
Merge origin/dev into perf/vram-trunk-first: the CUDA chain and V4.1'…
jameswitney Nov 8, 2026
2027439
Merge origin/dev into perf/vram-trunk-first: Vulkan's buffer memory t…
jameswitney Nov 8, 2026
bb38833
cpu: the grid IQ types' dot product with the signs on the activation
JustVugg Nov 8, 2026
85168da
Merge pull request #2610 from benmaster82/fix/install-runtime-assets
JustVugg Nov 8, 2026
a1af656
Merge pull request #2617 from Marouane-elfadli/fix/libgomp-static-msy…
JustVugg Nov 8, 2026
b426a19
Merge pull request #2552 from rudycelekli/fix/glm53-full-chat-templat…
JustVugg Nov 8, 2026
8263c80
Merge pull request #2558 from rudycelekli/fix/responses-prefill-disco…
JustVugg Nov 8, 2026
9ffab23
Merge pull request #2664 from bobsummerwill/split-20261107/dsv4-build…
JustVugg Nov 8, 2026
7d85d12
Merge pull request #2665 from bobsummerwill/split-20261107/dsv4-multi…
JustVugg Nov 8, 2026
579e4a3
Merge pull request #2666 from bobsummerwill/split-20261107/dsv4-resident
JustVugg Nov 8, 2026
9e72397
Merge pull request #2667 from bobsummerwill/split-20261107/dsv4-cuda-…
JustVugg Nov 8, 2026
28d0988
Merge pull request #2669 from bobsummerwill/split-20261107/dsv4-peer-…
JustVugg Nov 8, 2026
86eae5a
Merge pull request #2637 from jameswitney/perf/vram-trunk-first
JustVugg Nov 8, 2026
bc7a8f6
Merge pull request #800 from ZacharyZcR/feat/env-registry
JustVugg Nov 8, 2026
e5dedfa
env registry: the variables dev and this batch read since #800 was wr…
JustVugg Nov 8, 2026
50a73e1
deepseek_v4: the CUDA tier's unit includes the device-list parser
JustVugg Nov 8, 2026
5cb097f
env registry: no line for what coli serve or Windows set on every engine
JustVugg Nov 8, 2026
6b09123
ci: the MSYS2 installs without the libgomp package CI's MSYS2 does no…
JustVugg Nov 8, 2026
File filter

Filter by extension

Filter by extension


Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
12 changes: 12 additions & 0 deletions .github/workflows/ci.yml
Original file line number Diff line number Diff line change
Expand Up @@ -12,6 +12,18 @@ permissions:
contents: read

jobs:
# The environment-variable registry (c/coli_env.h) against the sources. No
# build, no model, no GPU -- seconds, so it reports before anything else.
env-registry:
name: Env registry in sync
runs-on: ubuntu-latest
steps:
- uses: actions/checkout@v4
- uses: actions/setup-python@v5
with:
python-version: '3.12'
- run: make -C c check-env

engine:
name: Engine (Linux, CPU)
runs-on: ubuntu-latest
Expand Down
24 changes: 24 additions & 0 deletions CHANGELOG.md
Original file line number Diff line number Diff line change
Expand Up @@ -44,6 +44,16 @@ Format follows [Keep a Changelog](https://keepachangelog.com/).

### Performance

- **CPU: the grid IQ types' dot product with the signs on the activation.** IQ2_XXS, IQ2_XS,
IQ2_S, IQ3_XXS and IQ3_S applied their sign bits to the decoded weights, and the integer dot
product took them off again to multiply by the activation. The grid rows now go in as they are
and the signs onto the activation: the same two operands for every product, so the same bits
(`tests/test_gq_signs` against the kernels it replaced). The indices of a sub-block come from
one load, and IQ1_M's decoder reads its scales and index bits a word at a time. One Zen 4
core, x86-64-v3, a 27B gate matrix: IQ2_XXS 3.58 -> 4.37 GB/s, IQ3_XXS 3.45 -> 5.22, IQ2_XS
3.08 -> 3.53, IQ1_M 1.16 -> 1.90; sixteen threads IQ2_XXS 20.7 -> 25.5. Qwen3.8-27B UD-IQ2_XXS
on the CPU, the same greedy answers: a Ryzen 7 8700GE prompt / decode 2.7 / 2.4 -> 2.8 / 2.8
tok/s, a Core i5-13500 1.9 / 1.9 -> 2.0 / 2.2.
- **Qwen3.8-27B on Vulkan: its attention on the matrix units.** `chain_attn_flash` takes 64 query
rows a workgroup of one kv head's heads, so it ran where the heads a kv head divide 64; the 27B's
six took the blocked shader, and its decode, which follows the prompt's path, read every
Expand Down Expand Up @@ -552,6 +562,11 @@ Format follows [Keep a Changelog](https://keepachangelog.com/).
vectorizes. Built for Sandy Bridge with glibc's software `fmaf`, Qwen3.8 IQ2_XS: the recurrence
208 -> 52 ms a forward, the same perplexity (2.2402 nats a token over 258 tokens); the planar
expert kernel's f32 mode (`QWEN_EXPERT_ACT=f32`) on Qwen3.6 0.59 -> 2.9 tok/s.
- `make install` ships the chat UI and `coli ui` finds it in libexec. The
Models page resolves the installed support modules and the launcher in bin,
so it can import GGUFs and start a gateway outside a source checkout.
- `make install` includes `tools/iq3xxs_grid.json`, required by the IQ3/E8
codec when a conversion uses the published lattice table.
- **Qwen3.8: memory that grew with every answer** (#2592). The expert loader reads a batch of
missing experts in parallel (on by default) with a team as large as the batch, and libgomp
ends the threads a smaller team leaves out and starts new ones for a larger. Each read
Expand Down Expand Up @@ -612,6 +627,15 @@ Format follows [Keep a Changelog](https://keepachangelog.com/).

- Qwen3.8-Flash-Next's long-context table measured again on 2.1.0: 65,536 tokens 208 / 52.2
tok/s, 131,072 190 / 51.9, 262,144 151 / 42.9 (prompt / decode).
### Fixed

- **#2612**: on MSYS2 UCRT64 with gcc 16, the static OpenMP library (`libgomp.a`)
now comes from its own package, so a fresh install failed at the link with
`cannot find -lgomp`. `make` warns up front and names
`mingw-w64-ucrt-x86_64-libgomp`; `setup.sh` probes the way the Makefile links
(`-static`) and names the right package; `coli setup` treats a gcc without
`libgomp.a` as "build packages missing"; the Windows and quickstart docs and the
CI/release MSYS2 install lists now include the package.

## [2.1.0] - 2026-10-31

Expand Down
4 changes: 4 additions & 0 deletions c/.gitignore
Original file line number Diff line number Diff line change
Expand Up @@ -18,6 +18,7 @@ build/
COLI_V4_UNIT_*.o
expert_store_registry.o
deepseek_v4
tools/bench_dsv4_head_batch
test_expert_store_registry
tests/test_st_mirror
tests/test_st_pread
Expand Down Expand Up @@ -81,3 +82,6 @@ tests/xdna_fail_fixtures/
tests/xdna_physical_probe
tests/xdna_fake_helper*.so
libcoli_xdna.so
shaders/chain_dnconv_rows.spv
shaders/qmatmul_bd.spv
shaders/qmatmul_gate_up_bd.spv
59 changes: 53 additions & 6 deletions c/Makefile
Original file line number Diff line number Diff line change
Expand Up @@ -142,6 +142,20 @@ CFLAGS = -D_FILE_OFFSET_BITS=64 -O3 -march=$(ARCH) -fopenmp -Wall -Wextra -Wno-
# (warns), so link psapi explicitly or the build fails undefined-reference on
# modern GCC (e.g. 16.x under UCRT). Harmless where the pragma also resolves it.
LDFLAGS = -lm -fopenmp -static -lpsapi $(EXTRA_LDFLAGS)
# The -static link needs libgomp.a. Since gcc 16 MSYS2 ships it in its own package
# (mingw-w64-ucrt-x86_64-libgomp) instead of pulling it in with gcc, so a fresh
# install has -fopenmp but no archive, and the link stops with a bare
# "cannot find -lgomp" (#1900, #2612). Ask the compiler where the archive is: when
# it is missing, -print-file-name just echoes "libgomp.a" back. Warn up front and
# name the package; the build still goes on, so goals that never link OpenMP are
# not blocked. LIBGOMP_CHECK=0 silences it.
ifneq ($(LIBGOMP_CHECK),0)
ifneq ($(filter-out clean%,$(or $(MAKECMDGOALS),all)),)
ifeq ($(shell $(DETECT_CC) -print-file-name=libgomp.a 2>/dev/null),libgomp.a)
$(warning $(DETECT_CC) has no libgomp.a, so the -static -fopenmp link will fail with "cannot find -lgomp". In the "MSYS2 UCRT64" shell run: pacman -S --needed mingw-w64-ucrt-x86_64-libgomp (docs/windows.md))
endif
endif
endif
ifeq ($(LTO),1)
CFLAGS += -flto
LDFLAGS += -flto
Expand Down Expand Up @@ -567,6 +581,8 @@ TEST_BINS += tests/test_gsgemv_sse41$(EXE) tests/test_qgemv_sse41$(EXE) \
tests/test_idot_ssse3$(EXE) tests/test_int_kernel_exact_ssse3$(EXE)
endif
ifeq ($(COLI_V4_SUPPORTED),1)
TEST_BINS += tests/test_v4_gpu_devices$(EXE)
TEST_BINS += tests/test_v4_hybrid_policy$(EXE) tests/test_k3_fill_budget$(EXE) tests/test_v4_bank_pair$(EXE)
TEST_BINS += tests/test_deepseek_v4$(EXE) tests/test_v4_ownership$(EXE) \
tests/test_v4_serve_framing$(EXE) \
tests/test_segment_adapters_registration$(EXE) \
Expand Down Expand Up @@ -1109,7 +1125,7 @@ cu_tier.o: cu_tier.c vk_tier.c vk_tier.h cu_tier.h cu_tier_ops.h .build-config
$(CC) $(CFLAGS) -c cu_tier.c -o $@


backend_cuda_dsv4.o: backend_cuda_dsv4.cu backend_cuda_dsv4.h .build-config $(DSV4_DEEPGEMM_DEP)
backend_cuda_dsv4.o: backend_cuda_dsv4.cu backend_cuda_dsv4.h backend_cuda_dsv4_quant.cuh .build-config $(DSV4_DEEPGEMM_DEP)
@command -v "$(NVCC)" >/dev/null 2>&1 || { echo "nvcc not found: set CUDA_HOME or NVCC" >&2; exit 1; }
"$(NVCC)" $(NVCCFLAGS) -c backend_cuda_dsv4.cu -o $@

Expand Down Expand Up @@ -1144,7 +1160,7 @@ else
DSV4_DLL_TC =
DSV4_DLL_LIBS = -lcublasLt -lcudart -lcuda
endif
cuda-dsv4-dll: backend_cuda_dsv4.cu backend_cuda_dsv4.h dsv4.def
cuda-dsv4-dll: backend_cuda_dsv4.cu backend_cuda_dsv4.h backend_cuda_dsv4_quant.cuh dsv4.def
@command -v "$(NVCC)" >/dev/null 2>&1 || { echo "nvcc not found: set CUDA_HOME or NVCC" >&2; exit 1; }
@command -v cl >/dev/null 2>&1 || { echo "cl.exe (MSVC) not in PATH — run vcvars64.bat first" >&2; exit 1; }
@cl 2>&1 | grep -qE "x64|AMD64" || { echo "cl.exe is the 32-bit (x86) compiler: nvcc then fails inside cuda_fp16.hpp with 'asm operand type size(8)'. Open 'x64 Native Tools Command Prompt for VS 2022' (or run vcvars64.bat), not the generic Developer Command Prompt (#1405)" >&2; exit 1; }
Expand All @@ -1157,7 +1173,7 @@ cuda-dsv4-dll: backend_cuda_dsv4.cu backend_cuda_dsv4.h dsv4.def
# checkout into $(DEEPGEMM_HOME), see the DEEPGEMM block above); this target
# fetches them on first use.
DEEPGEMM_INC = -I"$(DEEPGEMM_HOME)/deep_gemm/include" -I"$(DEEPGEMM_HOME)/cutlass/include" -I"$(DEEPGEMM_HOME)/third-party/cutlass/include"
cuda-dsv4-dg-dll: backend_cuda_dsv4.cu backend_cuda_dsv4.h dsv4.def $(DEEPGEMM_STAMP)
cuda-dsv4-dg-dll: backend_cuda_dsv4.cu backend_cuda_dsv4.h backend_cuda_dsv4_quant.cuh dsv4.def $(DEEPGEMM_STAMP)
@command -v "$(NVCC)" >/dev/null 2>&1 || { echo "nvcc not found: set CUDA_HOME or NVCC" >&2; exit 1; }
@command -v cl >/dev/null 2>&1 || { echo "cl.exe (MSVC) not in PATH — run vcvars64.bat first" >&2; exit 1; }
@cl 2>&1 | grep -qE "x64|AMD64" || { echo "cl.exe is the 32-bit (x86) compiler: nvcc then fails inside cuda_fp16.hpp with 'asm operand type size(8)'. Open 'x64 Native Tools Command Prompt for VS 2022' (or run vcvars64.bat), not the generic Developer Command Prompt (#1405)" >&2; exit 1; }
Expand Down Expand Up @@ -1346,8 +1362,18 @@ cuda-test: tests/test_mxfp4_expert_cuda.cu backend_cuda.cu backend_cuda.h fp8_fo
# against the CPU reference in each test, then the generic build's kernel
# variants (expert-major vs per-route MoE, warp vs block dense batch) against
# each other bit for bit. Needs a GPU, so CI runs none of it; it does not
# build these tests either, only backend_cuda_dsv4.cu itself.
dsv4-cuda-test: tests/test_dsv4_dense_batch_cuda.c tests/test_dsv4_attention_batch_cuda.c tests/test_dsv4_moe_batch_cuda.c tests/test_dsv4_generic_bitwise_cuda.c tests/test_dsv4_bank_bytes_cuda.c backend_cuda_dsv4.h $(DSV4_CUDA_OBJ)
# build these tests either, only backend_cuda_dsv4.cu itself. First, the
# two-GPU start-up (peer probe, init rollback; it includes backend_cuda_dsv4.cu
# with shims, so it links what the backend needs -- cuBLASLt for the default
# tensor-core path, NCCL with NCCL=1 -- plus the driver API it asks whether a
# context exists), then the decode MoE with experts on a peer GPU (peer
# compute and in-place reads vs one GPU, and an injected peer failure; skips
# without two peer-capable GPUs).
dsv4-cuda-test: tests/test_dsv4_dense_batch_cuda.c tests/test_dsv4_attention_batch_cuda.c tests/test_dsv4_moe_batch_cuda.c tests/test_dsv4_generic_bitwise_cuda.c tests/test_dsv4_bank_bytes_cuda.c tests/test_dsv4_peer_init_cuda.cu tests/test_dsv4_peer_moe_cuda.c backend_cuda_dsv4.cu backend_cuda_dsv4.h $(DSV4_CUDA_OBJ)
"$(NVCC)" $(NVCCFLAGS) tests/test_dsv4_peer_init_cuda.cu -o dsv4_peer_init_cuda_test$(EXE) -lcuda -lcublasLt $(if $(filter 1,$(NCCL)),-lnccl)
./dsv4_peer_init_cuda_test$(EXE)
$(CC) $(CFLAGS) tests/test_dsv4_peer_moe_cuda.c $(DSV4_CUDA_OBJ) -o dsv4_peer_moe_cuda_test$(EXE) $(LDFLAGS)
./dsv4_peer_moe_cuda_test$(EXE)
$(CC) $(CFLAGS) tests/test_dsv4_bank_bytes_cuda.c $(DSV4_CUDA_OBJ) -o dsv4_bank_bytes_cuda_test$(EXE) $(LDFLAGS)
./dsv4_bank_bytes_cuda_test$(EXE)
$(CC) $(CFLAGS) tests/test_dsv4_dense_batch_cuda.c $(DSV4_CUDA_OBJ) -o dsv4_cuda_test$(EXE) $(LDFLAGS)
Expand Down Expand Up @@ -2352,6 +2378,9 @@ tests/test_gq_dot$(EXE): tests/test_gq_dot.c
# no-op on a build without AVX2. `tests/test_gq_r16 bench S` times them.
tests/test_gq_r16$(EXE): tests/test_gq_r16.c
$(CC) $(CFLAGS) $< -o $@ $(LDFLAGS)
# The grid types' dot product with the signs on the activation against the one it replaced
tests/test_gq_signs$(EXE): tests/test_gq_signs.c
$(CC) $(CFLAGS) $< -o $@ $(LDFLAGS)
# A matrix's file rows eight at a time (gq_rows_f8) against their scalar reference and the R16
# forms' kernels, to the bit; a no-op on a build without AVX2. `tests/test_rows_f8 bench` times them.
tests/test_rows_f8$(EXE): tests/test_rows_f8.c
Expand Down Expand Up @@ -3279,6 +3308,9 @@ tests/test_v4_hybrid_policy$(EXE): tests/test_v4_hybrid_policy.c
tests/test_k3_fill_budget$(EXE): tests/test_k3_fill_budget.c
$(CC) $(CFLAGS) $< -o $@ $(LDFLAGS)

tests/test_v4_gpu_devices$(EXE): tests/test_v4_gpu_devices.c
$(CC) $(NOCUDA_CFLAGS) $< -o $@ $(NOCUDA_LDFLAGS)

tests/test_v4_bank_pair$(EXE): tests/test_v4_bank_pair.c
$(CC) $(CFLAGS) $< -o $@ $(LDFLAGS)

Expand Down Expand Up @@ -3368,6 +3400,9 @@ tests/test_moe_gs_guard$(EXE): tests/test_moe_gs_guard.c colibri.c $(VK_OBJ)
tests/test_omp_tune$(EXE): tests/test_omp_tune.c
$(CC) $(CFLAGS) $< -o $@ $(LDFLAGS)

tests/test_env_registry$(EXE): tests/test_env_registry.c
$(CC) $(NOCUDA_CFLAGS) $< -o $@ $(NOCUDA_LDFLAGS)

tests/test_compat_env$(EXE): tests/test_compat_env.c
$(CC) $(CFLAGS) $< -o $@ $(LDFLAGS)

Expand Down Expand Up @@ -3528,8 +3563,18 @@ test-asan:
@$(ASAN_ENV) $(MAKE) test-c EXTRA_CFLAGS="$(ASAN_CFLAGS)" EXTRA_LDFLAGS="$(ASAN_LDFLAGS)"
@echo "test-asan: suite clean under ASan + UBSan"

# --- environment-variable registry --------------------------------------
# c/coli_env.h is the single list of every variable the engines read. This
# target asserts it still matches the sources: a getenv() with no row, a row
# nothing reads, an unsorted table (coli_env_find binary-searches it) or a
# duplicate name all fail here. Needs no build, no model and no GPU.
check-env:
@$(PYTHON) tools/check_env_registry.py
.PHONY: check-env

# Local validation: one portable CPU build and dependency-free tests.
check:
$(MAKE) check-env
$(MAKE) clean
$(MAKE) portable
$(MAKE) test
Expand Down Expand Up @@ -3560,7 +3605,9 @@ install: colibri-tui$(EXE) colibri$(EXE) glm53$(EXE) inkling$(EXE) kimi_k3$(EXE)
openai_server.py cluster.py v4_dsml.py v41_dsml.py version.py \
image_engine.py term_image.py setup_flow.py setup_hw.py setup_catalog.py \
setup_download.py mcp_server.py power.py $(DESTDIR)$(LIBEXECDIR)/
$(INSTALL) -m 644 tools/*.py $(DESTDIR)$(LIBEXECDIR)/tools/
$(INSTALL) -m 644 tools/*.py tools/iq3xxs_grid.json $(DESTDIR)$(LIBEXECDIR)/tools/
$(INSTALL) -d $(DESTDIR)$(LIBEXECDIR)/ui
$(INSTALL) -m 644 ../ui/*.py ../ui/index.html $(DESTDIR)$(LIBEXECDIR)/ui/
@# 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 \
Expand Down
Loading
Loading