Shang Chieh Tseng d948926581 Fix Tesla K80 CUBLAS compatibility with two-tier fallback strategy ...

This commit implements comprehensive Tesla K80 (Kepler, compute 3.7)
compatibility for batched matrix multiplication operations.

**Problem:**
Modern CUBLAS functions fail on Tesla K80 with CUBLAS_STATUS_ARCH_MISMATCH:
1. CUBLAS_GEMM_DEFAULT_TENSOR_OP requires Tensor Cores (Volta+ only)
2. cublasGemmStridedBatchedEx/cublasGemmBatchedEx have architectural
   requirements beyond algorithm selection

**Solution - Two-Tier Fallback:**

Tier 1: Algorithm Selection
- Volta+ (cc >= 7.0): CUBLAS_GEMM_DEFAULT_TENSOR_OP
- Pre-Volta (cc < 7.0): CUBLAS_GEMM_DEFAULT

Tier 2: Function Selection
- Volta+ or non-FP32: Use *Ex variants (flexible precision)
- Kepler/Maxwell/Pascal with FP32: Use legacy type-specific functions
  (cublasSgemmStridedBatched, cublasSgemmBatched)

**Changes:**

CUDA Implementation:
- ml/backend/ggml/ggml/src/ggml-cuda/ggml-cuda.cu
  * ggml_cuda_op_mul_mat_cublas: Algorithm selection for non-batched ops
  * ggml_cuda_mul_mat_batched_cublas_impl: Two-tier fallback for batched ops
  * Added GGML_CUDA_DEBUG environment variable for conditional debug logging
  * Comprehensive function documentation explaining fallback strategy

Documentation:
- CLAUDE.md
  * Added Tesla K80 CUBLAS Compatibility section
  * Documented GGML_CUDA_DEBUG environment variable
  * Enhanced "Running Ollama" section with log capture examples
  * Updated Files Modified list

Code Comments:
- Added detailed comments throughout CUDA code explaining:
  * Why TENSOR_OP fails on pre-Volta GPUs
  * Why *Ex functions require architectural support
  * Compute capability checks and fallback logic
  * Debug logging usage

**Testing:**
All models verified working on Tesla K80:
- ✅ gemma3:4b
- ✅ gpt-oss
- ✅ deepseek-r1

Debug flag tested in both enabled and disabled states.

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude <noreply@anthropic.com>

2025-11-05 23:52:45 +08:00

28 KiB

Raw Blame History

View Raw