JIT compilation #170
Labels
Comments
|
For reference the following CI command on main will compile in 142minutes on a 32 core Zen3 4GHZ server container. Without Changing TORCH_CUDA_ARCH_LIST doesn't have much impact on speed. |
Merged
yzh119
added a commit
that referenced
this issue
Oct 7, 2024
This PR implements the JIT compilation (#170 ) of flashinfer, after this PR, flashinfer will compile kernels just-in-time for different input data types and shapes, and cached the kernels at the disk, instead of pre-compile a set of kernels in the wheel. # Motivation The pip wheel size is exploding as we add support to more data types, more head dimensions, more attention variants and more kernel implementation. Pre-compile everything is not sustainable, and impedes development speed. This PR refactors the codebase to use torch's [JIT Compiling Extensions](https://pytorch.org/tutorials/advanced/cpp_extension.html#jit-compiling-extensions) feature instead of pre-compile kernels in the wheel. ## Attention Variants We learned from [FlexAttention](https://pytorch.org/blog/flexattention/) and describes every attention variant as a template class, each instance of the struct can carry some closure variable defined in local memory or shared memory, below are two examples (logits soft cap and alibi attention, the programming interface is tentative and will be updated as we improve the programmability of the JIT template): ```cuda template <typename ParamsT> struct LogitsSoftCap { using DTypeQ = typename ParamsT::DTypeQ; using DTypeKV = typename ParamsT::DTypeKV; using DTypeO = typename ParamsT::DTypeO; uint32_t qo_len, kv_len; uint32_t window_left; __device__ __host__ LogitsSoftCap(const ParamsT& params, uint32_t batch_idx, uint8_t* smem_ptr) { qo_len = params.get_qo_len(batch_idx); kv_len = params.get_kv_len(batch_idx); window_left = kv_len; } template <typename T> __device__ __forceinline__ T QueryTransform(const ParamsT& params, T q) { return float(q) * params.sm_scale * math::ptx_rcp(params.logits_soft_cap); } template <typename T> __device__ __forceinline__ T LogitsTransform(const ParamsT& params, T logits, uint32_t batch_idx, uint32_t qo_idx, uint32_t kv_idx, uint32_t qo_head_idx, uint32_t kv_head_idx) { return params.logits_soft_cap * math::log2e * float(math::tanh(logits)); } __device__ __forceinline__ bool LogitsMask(const ParamsT& params, uint32_t batch_idx, uint32_t qo_idx, uint32_t kv_idx, uint32_t qo_head_idx, uint32_t kv_head_idx) { return true; } }; template <typename ParamsT> struct ALIBIAttention { using DTypeQ = typename ParamsT::DTypeQ; using DTypeKV = typename ParamsT::DTypeKV; using DTypeO = typename ParamsT::DTypeO; using IdType = typename ParamsT::IdType; uint32_t qo_len, kv_len; uint32_t window_left; __device__ __host__ ALIBIAttention(const ParamsT& params, uint32_t batch_idx, uint8_t* smem_ptr) { qo_len = params.get_qo_len(batch_idx); kv_len = params.get_kv_len(batch_idx); window_left = kv_len; } template <typename T> __device__ __forceinline__ T QueryTransform(const ParamsT& params, T q) { return float(q) * params.sm_scale * math::log2e; } template <typename T> __device__ __forceinline__ T LogitsTransform(const ParamsT& params, T logits, uint32_t batch_idx, uint32_t qo_idx, uint32_t kv_idx, uint32_t qo_head_idx, uint32_t kv_head_idx) { return logits + params.alibi_slopes[qo_head_idx] * float(int(kv_idx) - int(qo_idx)); } __device__ __forceinline__ bool LogitsMask(const ParamsT& params, uint32_t batch_idx, uint32_t qo_idx, uint32_t kv_idx, uint32_t qo_head_idx, uint32_t kv_head_idx) { return true; } }; ``` User can customize their own `ParamsT` class and variants class to define their own attention variants, we hope such refactor will make the codebase more concise and extensive. # Roadmap After this PR, we will add support for: 1. PyPI wheels #153 2. fp8 tensor cores attention: #502 3. different head dimensions: #142 #454 #455 4. flashattention3 #369 5. multi-head latency attention #237 6. Generate ParamsT and Attention variants description from python dsl The development of this features have been blocked by the limitation of wheel size (binary size >= 2GB will trigger some linking issues), I hope this PR will make development easier in the future.
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
As the combination of shapes and configurations increases, our pip wheel size grows and the compilation time becomes long.
PyTorch supports Just-In-Time compilation of extensions:
https://pytorch.org/tutorials/advanced/cpp_extension.html#jit-compiling-extensions, which makes it possible to only compile kernels corresponding to certain configurations/shapes, thus reducing both the wheel size and the development overhead on the codebase.
We can release a
flashinfer_jitwheel where all kernels are compiled with JIT.The text was updated successfully, but these errors were encountered: