artifacts/m15_scripts/colab_deploy2.py

#!/usr/bin/env python3
# -*- coding: utf-8 -*-
"""colab_deploy2.py — Colab T4 单卡 12G RAM 现实路径: 底模量化 + LoRA 运行时挂载
血统: W_eff = quant(B) + scale*Δ10 (Δ10=续训链完整累积 delta, 相对原始底模)
流程: E1 convert底模->f16 -> E2 surgery -> E3 双量化 -> E4 lora->gguf -> E5 推仓 -> E6 serve
RAM 全程 <3G (convert/quantize 均 mmap), 绕开 cgroup OOM
幂等旗标: /content/k8b/D2_* ; 日志: /content/k8b/colab_deploy2.log"""
import os, subprocess, sys, time, traceback

K8B = "/content/k8b"
K27 = "/content/k27"
BASE = K27 + "/Qwen3.5-9B"
LC = "/content/lcbuild/llama.cpp"
BIN = "/content/lcbin"
F16 = K8B + "/base_f16.gguf"
Q4 = K8B + "/base_q4_k_m.gguf"
Q48 = K8B + "/base_mixbit_4x8.gguf"
ADP = K8B + "/m12_adapters/m13_r10"
LORA = K8B + "/m13_r10_lora.gguf"
MMPROJ = K8B + "/mmproj_m11.gguf"
LOG = K8B + "/colab_deploy2.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=14400):
    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()

LD = "LD_LIBRARY_PATH=/content/lcbin:/usr/lib64-nvidia:/usr/local/nvidia/lib64:/usr/lib/x86_64-linux-gnu"

def main():
    t0 = time.time()
    P("==== colab deploy2 (lora-mount path) start ====")
    try:
        # E1: convert 底模 -> f16 GGUF (mmap, RAM 友好)
        if not have("D2_CONVERT"):
            if not os.path.exists(LC + "/convert_hf_to_gguf.py"):
                raise RuntimeError("llama.cpp src missing")
            r = sh("cd %s && NO_LOCAL_GGUF=1 PYTHONPATH=%s/gguf-py python3 convert_hf_to_gguf.py "
                   "%s --outfile %s --outtype f16 > %s/convert2.log 2>&1; echo RC=$?; tail -2 %s/convert2.log"
                   % (LC, LC, BASE, F16, K8B, K8B), 14400)
            P("convert: " + r[-250:])
            if not os.path.exists(F16) or os.path.getsize(F16) < 15e9:
                raise RuntimeError("convert failed")  # [FIX] 只信文件, RC 被 tqdm stderr 淹没 (坑#69)
            flag("D2_CONVERT")
            P("convert ok (%.1fG, %.0fs)" % (os.path.getsize(F16) / 1e9, time.time() - t0))
            # 删底模 HF 缓存 (symlink BASE 指向它, convert 后无用)
            if os.path.islink(BASE):
                tgt = os.path.realpath(BASE)
                os.remove(BASE)
                sh("rm -rf %s/hf/models--Qwen--Qwen3.5-9B %s" % (K27, tgt), 900)
                P("base cache freed")
        # E2: surgery v2 (bc->32, nx->0)
        if not have("D2_SURGERY"):
            r = subprocess.run([sys.executable, K8B + "/r12_surgery.py", F16],
                               capture_output=True, text=True, timeout=2400,
                               env=dict(os.environ, PYTHONPATH=LC + "/gguf-py"))
            outp = r.stdout + r.stderr
            P("surgery: %s" % outp[-300:])
            if '"R12_SURGERY_V2_DONE": true' not in outp:
                raise RuntimeError("surgery failed")
            flag("D2_SURGERY")
        # E3: 双量化 (等编译产物)
        if not os.path.exists(BIN + "/llama-quantize"):
            P("llama-quantize 未就绪, 等编译 (build pct: %s)" %
              sh("grep -oE '\\[ *[0-9]+%\\]' /content/lcbuild/llama.cpp/build.log 2>/dev/null | tail -1", 15))
            for i in range(360):
                time.sleep(30)
                if os.path.exists(BIN + "/llama-quantize"):
                    break
                if have("COMPILE_FAIL"):
                    raise RuntimeError("compile failed while waiting")
            if not os.path.exists(BIN + "/llama-quantize"):
                raise RuntimeError("quantize binary timeout")
        if not have("D2_QUANT4"):
            cmd = "CUDA_VISIBLE_DEVICES= " + LD + " %s/llama-quantize %s %s q4_k_m" % (BIN, F16, Q4)
            sh(cmd + " > %s/qtz4.log 2>&1" % K8B, 14400)
            if not os.path.exists(Q4) or os.path.getsize(Q4) < 4e9:
                raise RuntimeError("quant4 failed: " + sh("tail -6 %s/qtz4.log" % K8B)[-300:])
            flag("D2_QUANT4")
            P("quant4 ok (%.1fG)" % (os.path.getsize(Q4) / 1e9))
        if not have("D2_QUANT48"):
            ATTN = ("--tensor-type attn_q=q8_0 --tensor-type attn_k=q8_0 "
                    "--tensor-type attn_v=q8_0 --tensor-type attn_output=q8_0")
            cmd = ("CUDA_VISIBLE_DEVICES= " + LD + " %s/llama-quantize --token-embedding-type q8_0 "
                   "--output-tensor-type q8_0 %s %s %s q4_k_m" % (BIN, ATTN, F16, Q48))
            sh(cmd + " > %s/qtz48.log 2>&1" % K8B, 14400)
            if not os.path.exists(Q48) or os.path.getsize(Q48) < 4e9:
                raise RuntimeError("quant48 failed: " + sh("tail -6 %s/qtz48.log" % K8B)[-300:])
            flag("D2_QUANT48")
            P("quant48 ok (%.1fG)" % (os.path.getsize(Q48) / 1e9))
            sh("rm -f %s" % F16, 30)  # 释放 18.5G
            P("f16 freed")
        # E4: adapter -> lora.gguf (PEFT 242M, RAM 小)
        if not have("D2_LORA"):
            r = sh("cd %s && PYTHONPATH=%s/gguf-py python3 convert_lora_to_gguf.py "
                   "%s --outfile %s > %s/lora_conv.log 2>&1; echo RC=$?; tail -3 %s/lora_conv.log"
                   % (LC, LC, ADP, LORA, K8B, K8B), 3600)
            P("lora conv: " + r[-300:])
            if not os.path.exists(LORA) or os.path.getsize(LORA) < 1e8:  # [FIX] file-only check
                raise RuntimeError("lora convert failed")
            flag("D2_LORA")
            P("lora.gguf ok (%.0fM)" % (os.path.getsize(LORA) / 1e6))
        # E5: 推仓 (第一时间: lora 先行, 再两份 base 盘 + sha)
        if not have("D2_PUSH"):
            import hashlib
            from huggingface_hub import HfApi
            api = HfApi(token=hf_token())
            items = [(LORA, "m13_r10_lora.gguf: R12 adapter as gguf-lora (run-time mount)"),
                     (Q4, "base_q4_k_m.gguf: pure 4bit base for lora mount (colab T4)"),
                     (Q48, "base_mixbit_4x8.gguf: 4x8 base for lora mount (colab T4)")]
            for f, msg in items:
                sha = hashlib.sha256(open(f, "rb").read()).hexdigest()
                name = os.path.basename(f)
                api.upload_file(path_or_fileobj=f, path_in_repo=name,
                                repo_id="tchbcb/samai-9b", repo_type="model", commit_message=msg)
                open(K8B + "/" + name + ".sha256", "w").write(sha + "  " + name + "\n")
                api.upload_file(path_or_fileobj=K8B + "/" + name + ".sha256",
                                path_in_repo=name + ".sha256", repo_id="tchbcb/samai-9b", repo_type="model")
                P("pushed %s sha=%s" % (name, sha[:12]))
            flag("D2_PUSH")
        # E6: serve (Q4_K_M + LoRA 挂载, :8080, key=1234)
        if not have("D2_SERVE"):
            sh("pkill -f '[l]lama-server' 2>/dev/null; sleep 1", 20)
            cmd = (LD + " nohup %s/llama-server -m %s --lora %s --mmproj %s "
                   "--host 127.0.0.1 --port 8080 --api-key 1234 -c 32768 -ngl 99 -fa on "
                   "--parallel 2 > %s/llama_server.log 2>&1 &" % (BIN, Q4, LORA, MMPROJ, K8B))
            sh(cmd, 30)
            time.sleep(15)
            r = sh("curl -s -o /dev/null -w '%{http_code}' --max-time 5 -H 'Authorization: Bearer 1234' "
                   "http://127.0.0.1:8080/health; echo; nvidia-smi --query-gpu=memory.used --format=csv,noheader", 40)
            P("serve check: " + r)
            if "200" not in r:
                P("srv log: " + sh("tail -12 %s/llama_server.log" % K8B))
                raise RuntimeError("llama-server not healthy")
            flag("D2_SERVE")
            P("SERVE OK :8080 key=1234 (Q4 + lora mount)")
        P("==== DEPLOY2_ALL_DONE in %.0fs ====" % (time.time() - t0))
        flag("D2_ALL_DONE", "%.0fs" % (time.time() - t0))
    except Exception as e:
        traceback.print_exc(file=log)
        flag("D2_FAIL", repr(e)[:200])
        P("==== D2_FAIL: %s ====" % repr(e)[:200])

if __name__ == "__main__":
    main()