| |
| |
| """colab_compile.py — Colab T4 专用 llama.cpp CUDA 编译链 (sm_75) |
| 编译 -> 打包 llama_cpp_cuda_colab_t4.tar.gz -> 第一时间推仓 (tchbcb/samai-9b) |
| 幂等旗标: /content/k8b/CC_* |
| 日志: /content/k8b/colab_compile.log""" |
| import os, subprocess, sys, time, traceback |
|
|
| K8B = "/content/k8b" |
| BIN = "/content/lcbin" |
| LCSRC = "/content/lcbuild/llama.cpp" |
| LOG = K8B + "/colab_compile.log" |
| os.makedirs(K8B, exist_ok=True) |
| log = open(LOG, "a", buffering=1) |
|
|
| def P(m): |
| log.write("[%s] %s\n" % (time.strftime("%m-%d %H:%M:%S"), m)) |
|
|
| def sh(c, t=7200): |
| p = subprocess.run(c, shell=True, capture_output=True, text=True, timeout=t, errors="replace") |
| return ((p.stdout or "") + (p.stderr or ""))[-1500:] |
|
|
| def flag(n, c="1"): |
| open(K8B + "/" + n, "w").write(str(c)[:400]) |
|
|
| def have(n): |
| return os.path.exists(K8B + "/" + n) |
|
|
| def hf_token(): |
| return open("/root/.cache/huggingface/token").read().strip() |
|
|
| def main(): |
| t0 = time.time() |
| P("==== colab compile start ====") |
| try: |
| |
| if not have("CC_CLONE"): |
| if not os.path.exists(LCSRC + "/CMakeLists.txt"): |
| r = sh("rm -rf /content/lcbuild && mkdir -p /content/lcbuild && " |
| "git clone --depth 1 https://github.com/ggml-org/llama.cpp %s" % LCSRC, 900) |
| if not os.path.exists(LCSRC + "/CMakeLists.txt"): |
| raise RuntimeError("clone failed: " + r[-300:]) |
| flag("CC_CLONE", sh("cd %s && git rev-parse --short HEAD" % LCSRC, 15)) |
| P("clone ok") |
| |
| if not have("CC_BUILD"): |
| r = sh("cd /content/lcbuild && cmake -S %s -B %s/build -DGGML_CUDA=ON " |
| "-DCMAKE_CUDA_ARCHITECTURES=75 -DLLAMA_CURL=OFF -DCMAKE_BUILD_TYPE=Release " |
| "> %s/cmake.log 2>&1; echo RC=$?; tail -3 %s/cmake.log" % (LCSRC, LCSRC, LCSRC, LCSRC), 600) |
| P("cmake: " + r[-200:]) |
| if "RC=0" not in r: |
| raise RuntimeError("cmake failed") |
| r = sh("cd %s/build && cmake --build . --config Release -j2 > %s/build.log 2>&1; echo RC=$?" % (LCSRC, LCSRC), 7200) |
| P("build: " + r[-200:]) |
| if "RC=0" not in r: |
| raise RuntimeError("build failed: " + sh("tail -20 %s/build.log" % LCSRC)[-600:]) |
| flag("CC_BUILD") |
| P("build ok (%.0fs)" % (time.time() - t0)) |
| |
| if not have("CC_FLAT"): |
| os.makedirs(BIN, exist_ok=True) |
| sh("cp -a %s/build/bin/* %s/ 2>/dev/null" % (LCSRC, BIN), 300) |
| sh("find %s/build -name '*.so*' -exec cp -a {} %s/ \\; 2>/dev/null" % (LCSRC, BIN), 300) |
| need = ["llama-server", "llama-quantize", "llama-cli"] |
| miss = [b for b in need if not os.path.exists(BIN + "/" + b)] |
| if miss: |
| raise RuntimeError("binaries missing: %s" % miss) |
| r = sh("ls -la %s | head -20" % BIN, 15) |
| P("flat ok:\n" + r) |
| flag("CC_FLAT") |
| |
| if not have("CC_PUSH"): |
| tar = K8B + "/llama_cpp_cuda_colab_t4.tar.gz" |
| sh("tar czf %s -C %s . > /dev/null 2>&1" % (tar, BIN), 600) |
| if os.path.getsize(tar) < 3e7: |
| raise RuntimeError("tar too small: %d" % os.path.getsize(tar)) |
| from huggingface_hub import HfApi |
| api = HfApi(token=hf_token()) |
| api.upload_file(path_or_fileobj=tar, path_in_repo="llama_cpp_cuda_colab_t4.tar.gz", |
| repo_id="tchbcb/samai-9b", repo_type="model", |
| commit_message="colab T4 (sm_75) CUDA build, first-push") |
| flag("CC_PUSH", "%.0fs" % (time.time() - t0)) |
| P("CUDA tar pushed (%.0f MB, %.0fs)" % (os.path.getsize(tar) / 1e6, time.time() - t0)) |
| P("==== COMPILE_ALL_DONE in %.0fs ====" % (time.time() - t0)) |
| flag("COMPILE_ALL_DONE", "%.0fs" % (time.time() - t0)) |
| except Exception as e: |
| traceback.print_exc(file=log) |
| flag("COMPILE_FAIL", repr(e)[:200]) |
| P("==== COMPILE_FAIL: %s ====" % repr(e)[:200]) |
|
|
| if __name__ == "__main__": |
| main() |
|
|
