deploy_omni: llFile llama.cpp-omni compile succeeds on Modal T4

- CUDA Toolkit 12.4 + cuda-compiler + cuda-driver-dev from NVIDIA apt repo
- GGML_CUDA_NO_VMM=ON to avoid libcuda.so.1 linking in build phase
- CUDA arch 75 (T4) + 89 (L4)
- LLAMA_CUDA_FORCE_MMQ=ON for faster compilation
- Source from tc-mb/llama.cpp-omni fork, stored in modal_deploy/llamacpp_omni/
- Build time ~77 min on Modal infra

Resolves: modal deploy build failures due to GitHub access and CUDA linking

docs/deploy_omni_build_log_20260611.md

@@ -0,0 +1,23 @@
+# Build Log 2026-06-11: deploy_omni.py Success
+
+## Status
+llama.cpp-omni (OpenBMB fork) compiled successfully on Modal T4.
+Image: im-AxWdR31ZWeEDIfXIPcxAhZ
+Build time: ~4641s (77 min)
+
+## Key Fixes
+1. Source: cloned from tc-mb/llama.cpp-omni, stored in modal_deploy/llamacpp_omni/
+2. Build flags: GGML_CUDA=ON, LLAMA_CUDA_FORCE_MMQ=ON, GGML_CUDA_NO_VMM=ON
+3. CUDA arch: 75 (T4) + 89 (L4)
+4. CUDA Toolkit: cuda-toolkit-12-4 + cuda-compiler-12-4 + cuda-driver-dev-12-4 from NVIDIA apt repo
+5. Critical: libcuda.so.1 linking issue fixed via GGML_CUDA_NO_VMM=ON
+
+## Volume Status - minicpm-o-4_5-models
+All 10 model files present, GGUF magic verified.
+
+## Next Steps
+1. Run test_inference to verify text inference on T4
+2. Test TTS endpoints (v1/audio/speech)
+3. Test STT endpoints (v1/audio/transcriptions)
+4. modal deploy to production (prego-pal-minicpm-omni)
+5. Consider L4 GPU - cheaper, sm_89 already compiled

modal_deploy/deploy_omni.py

@@ -43,7 +43,7 @@ _omni_image = (
         "curl -L -o /tmp/cuda-keyring.deb https://developer.download.nvidia.com/compute/cuda/repos/debian12/x86_64/cuda-keyring_1.1-1_all.deb",
         "dpkg -i /tmp/cuda-keyring.deb",
         "apt-get update",
-        "apt-get install -y cuda-toolkit-12-4 cuda-compiler-12-4",
+        "apt-get install -y cuda-toolkit-12-4 cuda-compiler-12-4 cuda-driver-dev-12-4",
     )
     .apt_install(
         "curl",
@@ -72,16 +72,15 @@ _omni_image = (
 .run_commands(
         "cd /llama.cpp-omni && cmake -B build "
         "-DGGML_CUDA=ON "
-        "-DLLAMA_CURL=ON "
         "-DLLAMA_BUILD_SERVER=ON "
         "-DLLAMA_BUILD_TESTS=OFF "
         "-DLLAMA_BUILD_EXAMPLES=OFF "
         "-DLLAMA_CUDA_FORCE_MMQ=ON "
+        "-DGGML_CUDA_NO_VMM=ON "
         "-DCMAKE_CUDA_ARCHITECTURES='75;89' "
         "-DCMAKE_BUILD_TYPE=Release "
         "-DCMAKE_CUDA_COMPILER=/usr/local/cuda-12/bin/nvcc",
-        "cd /llama.cpp-omni && cmake --build build --config Release -j $(nproc) "
-        "--target llama-server",
+        "cd /llama.cpp-omni && cmake --build build --config Release -j $(nproc) --target llama-server",
         "ls -lh /llama.cpp-omni/build/bin/llama-server",
     )
 )
@@ -444,6 +443,20 @@ def diagnose_volume():
         else:
             print(f"[WARN] Main model NOT valid GGUF (magic={magic.hex()})")
 
+    print(f"\n{'='*60}")
+    print(f"[Diagnose] CUDA library files in container")
+    print(f"{'='*60}")
+    import subprocess
+    result = subprocess.run(
+        "find / -name 'libcuda*' -type f,l 2>/dev/null | head -30",
+        shell=True, capture_output=True, text=True
+    )
+    cuda_files = result.stdout.strip().split("\\n")
+    for f in cuda_files:
+        if f:
+            size = os.path.getsize(f) if os.path.exists(f) else 0
+            print(f"  {f}  ({size:,} bytes)")
+
 
 # ============================================================================
 # 5. TEST INFERENCE (standalone - not via ASGI)