xformers 0.0.23.dev651__tar.gz → 0.0.23.dev657__tar.gz

{xformers-0.0.23.dev651/xformers.egg-info → xformers-0.0.23.dev657}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: xformers
-Version: 0.0.23.dev651
+Version: 0.0.23.dev657
 Summary: XFormers: A collection of composable Transformer building blocks.
 Home-page: https://facebookresearch.github.io/xformers/
 Author: Facebook AI Research

{xformers-0.0.23.dev651 → xformers-0.0.23.dev657}/setup.py

@@ -118,9 +118,9 @@ def get_cuda_version(cuda_dir) -> int:
 
 
 def get_flash_attention_extensions(cuda_version: int, extra_compile_args):
-    # XXX: Not supported on windows yet
+    # XXX: Not supported on windows for cuda<12
     # https://github.com/Dao-AILab/flash-attention/issues/345
-    if platform.system() != "Linux":
+    if platform.system() != "Linux" and cuda_version < 1200:
         return []
     # Figure out default archs to target
     DEFAULT_ARCHS_LIST = ""

{xformers-0.0.23.dev651 → xformers-0.0.23.dev657}/third_party/flash-attention/csrc/flash_attn/flash_api.cpp

@@ -1037,13 +1037,14 @@ mha_varlen_bwd(const at::Tensor &dout,  // total_q x num_heads, x head_size
 
 std::vector<at::Tensor>
 mha_fwd_kvcache(at::Tensor &q,                 // batch_size x seqlen_q x num_heads x head_size
-                const at::Tensor &kcache,            // batch_size x seqlen_k x num_heads_k x head_size
-                const at::Tensor &vcache,            // batch_size x seqlen_k x num_heads_k x head_size
+                const at::Tensor &kcache,            // batch_size_c x seqlen_k x num_heads_k x head_size
+                const at::Tensor &vcache,            // batch_size_c x seqlen_k x num_heads_k x head_size
                 c10::optional<const at::Tensor> &k_, // batch_size x seqlen_knew x num_heads_k x head_size
                 c10::optional<const at::Tensor> &v_, // batch_size x seqlen_knew x num_heads_k x head_size
                 c10::optional<const at::Tensor> &seqlens_k_, // batch_size
                 c10::optional<const at::Tensor> &rotary_cos_, // seqlen_ro x (rotary_dim / 2)
                 c10::optional<const at::Tensor> &rotary_sin_, // seqlen_ro x (rotary_dim / 2)
+                c10::optional<const at::Tensor> &cache_batch_idx_, // indices to index into the KV cache
                 c10::optional<at::Tensor> &out_,             // batch_size x seqlen_q x num_heads x head_size
                 const float softmax_scale,
                 bool is_causal,
@@ -1084,6 +1085,7 @@ mha_fwd_kvcache(at::Tensor &q,                 // batch_size x seqlen_q x num_he
     const int head_size_og = sizes[3];
     const int seqlen_k = kcache.size(1);
     const int num_heads_k = kcache.size(2);
+    const int batch_size_c = kcache.size(0);
     TORCH_CHECK(batch_size > 0, "batch size must be postive");
     TORCH_CHECK(head_size_og <= 256, "FlashAttention forward only supports head dimension at most 256");
     TORCH_CHECK(num_heads % num_heads_k == 0, "Number of heads in key/value must divide number of heads in query");
@@ -1102,8 +1104,8 @@ mha_fwd_kvcache(at::Tensor &q,                 // batch_size x seqlen_q x num_he
     }
 
     CHECK_SHAPE(q, batch_size, seqlen_q, num_heads, head_size_og);
-    CHECK_SHAPE(kcache, batch_size, seqlen_k, num_heads_k, head_size_og);
-    CHECK_SHAPE(vcache, batch_size, seqlen_k, num_heads_k, head_size_og);
+    CHECK_SHAPE(kcache, batch_size_c, seqlen_k, num_heads_k, head_size_og);
+    CHECK_SHAPE(vcache, batch_size_c, seqlen_k, num_heads_k, head_size_og);
 
     at::Tensor q_padded, kcache_padded, vcache_padded;
     if (head_size_og % 8 != 0) {
@@ -1229,6 +1231,13 @@ mha_fwd_kvcache(at::Tensor &q,                 // batch_size x seqlen_q x num_he
         params.rotary_dim = 0;
     }
 
+    if (cache_batch_idx_.has_value()) {
+        auto cache_batch_idx = cache_batch_idx_.value();
+        CHECK_DEVICE(cache_batch_idx);
+        CHECK_CONTIGUOUS(cache_batch_idx);
+        TORCH_CHECK(cache_batch_idx.scalar_type() == torch::kInt32, "cache_batch_idx must have dtype int32");
+        params.cache_batch_idx = reinterpret_cast<int *>(cache_batch_idx.data_ptr());
+    }
     // This needs to match with run_mha_fwd_splitkv_dispatch
     const int block_n = head_size <= 64 ? 256 : (head_size <= 128 ? 128 : 64);
     const int num_n_blocks = (seqlen_k + block_n - 1) / block_n;
@@ -1248,8 +1257,8 @@ mha_fwd_kvcache(at::Tensor &q,                 // batch_size x seqlen_q x num_he
     }
 
     auto stream = at::cuda::getCurrentCUDAStream().stream();
-    // Only split kernel supports appending to KV cache
-    run_mha_fwd(params, stream, /*force_split_kernel=*/k_.has_value());
+    // Only split kernel supports appending to KV cache, or indexing to the cache with cache_batch_idx
+    run_mha_fwd(params, stream, /*force_split_kernel=*/k_.has_value() || cache_batch_idx_.has_value());
 
     if (head_size_og % 8 != 0) {
         out = out.index({"...", torch::indexing::Slice(torch::indexing::None, head_size_og)});

{xformers-0.0.23.dev651 → xformers-0.0.23.dev657}/third_party/flash-attention/csrc/flash_attn/src/flash.h

@@ -95,6 +95,9 @@ struct Flash_fwd_params : public Qkv_params {
     void * __restrict__ rotary_cos_ptr;
     void * __restrict__ rotary_sin_ptr;
 
+    // The indices to index into the KV cache.
+    int *__restrict__ cache_batch_idx;
+
     // The dropout probability (probability of keeping an activation).
     float p_dropout;
     // uint32_t p_dropout_in_uint;

{xformers-0.0.23.dev651 → xformers-0.0.23.dev657}/third_party/flash-attention/csrc/flash_attn/src/flash_bwd_launch_template.h

@@ -137,7 +137,7 @@ void run_flash_bwd(Flash_bwd_params &params, cudaStream_t stream, const bool con
 
 template<typename T>
 void run_mha_bwd_hdim32(Flash_bwd_params &params, cudaStream_t stream, const bool configure) {
-    constexpr int Headdim = 32;
+    constexpr static int Headdim = 32;
     int device;
     cudaGetDevice(&device);
     int max_smem_per_block;
@@ -158,7 +158,7 @@ void run_mha_bwd_hdim32(Flash_bwd_params &params, cudaStream_t stream, const boo
 
 template<typename T>
 void run_mha_bwd_hdim64(Flash_bwd_params &params, cudaStream_t stream, const bool configure) {
-    constexpr int Headdim = 64;
+    constexpr static int Headdim = 64;
     int device;
     cudaGetDevice(&device);
     int max_smem_per_block;
@@ -201,7 +201,7 @@ void run_mha_bwd_hdim64(Flash_bwd_params &params, cudaStream_t stream, const boo
 
 template<typename T>
 void run_mha_bwd_hdim96(Flash_bwd_params &params, cudaStream_t stream, const bool configure) {
-    constexpr int Headdim = 96;
+    constexpr static int Headdim = 96;
     int device;
     cudaGetDevice(&device);
     int max_smem_per_block;
@@ -228,7 +228,7 @@ void run_mha_bwd_hdim96(Flash_bwd_params &params, cudaStream_t stream, const boo
 
 template<typename T>
 void run_mha_bwd_hdim128(Flash_bwd_params &params, cudaStream_t stream, const bool configure) {
-    constexpr int Headdim = 128;
+    constexpr static int Headdim = 128;
     int device;
     cudaGetDevice(&device);
     int max_smem_per_block;
@@ -264,7 +264,7 @@ void run_mha_bwd_hdim128(Flash_bwd_params &params, cudaStream_t stream, const bo
 
 template<typename T>
 void run_mha_bwd_hdim160(Flash_bwd_params &params, cudaStream_t stream, const bool configure) {
-    constexpr int Headdim = 160;
+    constexpr static int Headdim = 160;
     int device;
     cudaGetDevice(&device);
     int max_smem_per_block;
@@ -281,7 +281,7 @@ void run_mha_bwd_hdim160(Flash_bwd_params &params, cudaStream_t stream, const bo
 
 template<typename T>
 void run_mha_bwd_hdim192(Flash_bwd_params &params, cudaStream_t stream, const bool configure) {
-    constexpr int Headdim = 192;
+    constexpr static int Headdim = 192;
     int device;
     cudaGetDevice(&device);
     int max_smem_per_block;
@@ -298,7 +298,7 @@ void run_mha_bwd_hdim192(Flash_bwd_params &params, cudaStream_t stream, const bo
 
 template<typename T>
 void run_mha_bwd_hdim224(Flash_bwd_params &params, cudaStream_t stream, const bool configure) {
-    constexpr int Headdim = 224;
+    constexpr static int Headdim = 224;
     BOOL_SWITCH(params.p_dropout < 1.f, Is_dropout, [&] {
         run_flash_bwd<Flash_bwd_kernel_traits<Headdim, 64, 64, 8, 4, 4, 4, false, false, T>, Is_dropout>(params, stream, configure);
     });
@@ -306,7 +306,7 @@ void run_mha_bwd_hdim224(Flash_bwd_params &params, cudaStream_t stream, const bo
 
 template<typename T>
 void run_mha_bwd_hdim256(Flash_bwd_params &params, cudaStream_t stream, const bool configure) {
-    constexpr int Headdim = 256;
+    constexpr static int Headdim = 256;
     int device;
     cudaGetDevice(&device);
     int max_smem_per_block;

{xformers-0.0.23.dev651 → xformers-0.0.23.dev657}/third_party/flash-attention/csrc/flash_attn/src/flash_fwd_kernel.h

@@ -668,9 +668,10 @@ inline __device__ void compute_attn_1rowblock_splitkv(const Params &params, cons
     const index_t row_offset_q = binfo.q_offset(params.q_batch_stride, params.q_row_stride, bidb)
         + m_block * kBlockM * params.q_row_stride + bidh * params.q_head_stride;
     // We move K and V to the last block.
-    const index_t row_offset_k = binfo.k_offset(params.k_batch_stride, params.k_row_stride, bidb)
+    const int bidb_cache = params.cache_batch_idx == nullptr ? bidb : params.cache_batch_idx[bidb];
+    const index_t row_offset_k = binfo.k_offset(params.k_batch_stride, params.k_row_stride, bidb_cache)
         + (n_block_max - 1) * kBlockN * params.k_row_stride + (bidh / params.h_h_k_ratio) * params.k_head_stride;
-    const index_t row_offset_v = binfo.k_offset(params.v_batch_stride, params.v_row_stride, bidb)
+    const index_t row_offset_v = binfo.k_offset(params.v_batch_stride, params.v_row_stride, bidb_cache)
         + (n_block_max - 1) * kBlockN * params.v_row_stride + (bidh / params.h_h_k_ratio) * params.v_head_stride;
 
     Tensor gQ = make_tensor(make_gmem_ptr(reinterpret_cast<Element *>(params.q_ptr) + row_offset_q),

{xformers-0.0.23.dev651 → xformers-0.0.23.dev657}/third_party/flash-attention/csrc/flash_attn/src/flash_fwd_launch_template.h

@@ -104,7 +104,7 @@ void run_flash_splitkv_fwd(Flash_fwd_params &params, cudaStream_t stream) {
         // We want kBlockM to be as small as possible for more parallelism.
         // With 128 threads we can load 512 elements at a time, so if headdim is divisible by 128, kBlockM = 4.
         // If headdim is divisible by 64, then we set kBlockM = 8, etc.
-        constexpr int kBlockM = Kernel_traits::kHeadDim % 128 == 0 ? 4 : (Kernel_traits::kHeadDim % 64 == 0 ? 8 : 16);
+        constexpr static int kBlockM = Kernel_traits::kHeadDim % 128 == 0 ? 4 : (Kernel_traits::kHeadDim % 64 == 0 ? 8 : 16);
         dim3 grid_combine((params.b * params.h * params.seqlen_q + kBlockM - 1) / kBlockM);
         BOOL_SWITCH(is_even_K, IsEvenKConst, [&] {
             if (params.num_splits <= 2) {
@@ -129,17 +129,17 @@ void run_flash_splitkv_fwd(Flash_fwd_params &params, cudaStream_t stream) {
 
 template<typename T, int Headdim>
 void run_mha_fwd_splitkv_dispatch(Flash_fwd_params &params, cudaStream_t stream) {
-    constexpr int kBlockM = 64;  // Fixed for all head dimensions
+    constexpr static int kBlockM = 64;  // Fixed for all head dimensions
     // TD [2023-08-28]: nvcc segfaults for headdim 96 with block size 64 x 256,
     // and for headdim 192 with block size 64 x 128.
     // Also for headdim 160 with block size 64 x 128 after the rotary addition.
-    constexpr int kBlockN = Headdim <= 64 ? 256 : (Headdim <= 128 ? 128 : 64);
+    constexpr static int kBlockN = Headdim <= 64 ? 256 : (Headdim <= 128 ? 128 : 64);
     run_flash_splitkv_fwd<Flash_fwd_kernel_traits<Headdim, kBlockM, kBlockN, 4, false, false, T>>(params, stream);
 }
 
 template<typename T>
 void run_mha_fwd_hdim32(Flash_fwd_params &params, cudaStream_t stream) {
-    constexpr int Headdim = 32;
+    constexpr static int Headdim = 32;
     BOOL_SWITCH(params.p_dropout < 1.f, Is_dropout, [&] {
         BOOL_SWITCH(params.is_causal, Is_causal, [&] {
             run_flash_fwd<Flash_fwd_kernel_traits<Headdim, 128, 128, 4, false, false, T>, Is_dropout, Is_causal>(params, stream);
@@ -149,7 +149,7 @@ void run_mha_fwd_hdim32(Flash_fwd_params &params, cudaStream_t stream) {
 
 template<typename T>
 void run_mha_fwd_hdim64(Flash_fwd_params &params, cudaStream_t stream) {
-    constexpr int Headdim = 64;
+    constexpr static int Headdim = 64;
     BOOL_SWITCH(params.p_dropout < 1.f, Is_dropout, [&] {
         BOOL_SWITCH(params.is_causal, Is_causal, [&] {
             if constexpr(!Is_dropout) {
@@ -171,7 +171,7 @@ void run_mha_fwd_hdim64(Flash_fwd_params &params, cudaStream_t stream) {
 
 template<typename T>
 void run_mha_fwd_hdim96(Flash_fwd_params &params, cudaStream_t stream) {
-    constexpr int Headdim = 96;
+    constexpr static int Headdim = 96;
     auto dprops = at::cuda::getCurrentDeviceProperties();
     bool is_sm8x = dprops->major == 8 && dprops->minor > 0;
     BOOL_SWITCH(params.p_dropout < 1.f, Is_dropout, [&] {
@@ -197,7 +197,7 @@ void run_mha_fwd_hdim96(Flash_fwd_params &params, cudaStream_t stream) {
 
 template<typename T>
 void run_mha_fwd_hdim128(Flash_fwd_params &params, cudaStream_t stream) {
-    constexpr int Headdim = 128;
+    constexpr static int Headdim = 128;
     auto dprops = at::cuda::getCurrentDeviceProperties();
     bool is_sm8x = dprops->major == 8 && dprops->minor > 0;
     BOOL_SWITCH(params.p_dropout < 1.f, Is_dropout, [&] {
@@ -234,7 +234,7 @@ void run_mha_fwd_hdim128(Flash_fwd_params &params, cudaStream_t stream) {
 
 template<typename T>
 void run_mha_fwd_hdim160(Flash_fwd_params &params, cudaStream_t stream) {
-    constexpr int Headdim = 160;
+    constexpr static int Headdim = 160;
     auto dprops = at::cuda::getCurrentDeviceProperties();
     bool is_sm8x = dprops->major == 8 && dprops->minor > 0;
     BOOL_SWITCH(params.p_dropout < 1.f, Is_dropout, [&] {
@@ -264,7 +264,7 @@ void run_mha_fwd_hdim160(Flash_fwd_params &params, cudaStream_t stream) {
 
 template<typename T>
 void run_mha_fwd_hdim192(Flash_fwd_params &params, cudaStream_t stream) {
-    constexpr int Headdim = 192;
+    constexpr static int Headdim = 192;
     BOOL_SWITCH(params.p_dropout < 1.f, Is_dropout, [&] {
         BOOL_SWITCH(params.is_causal, Is_causal, [&] {
             if constexpr(!Is_dropout) {
@@ -283,7 +283,7 @@ void run_mha_fwd_hdim192(Flash_fwd_params &params, cudaStream_t stream) {
 
 template<typename T>
 void run_mha_fwd_hdim224(Flash_fwd_params &params, cudaStream_t stream) {
-    constexpr int Headdim = 224;
+    constexpr static int Headdim = 224;
     int device;
     cudaGetDevice(&device);
     int max_smem_per_block;
@@ -309,7 +309,7 @@ void run_mha_fwd_hdim224(Flash_fwd_params &params, cudaStream_t stream) {
 
 template<typename T>
 void run_mha_fwd_hdim256(Flash_fwd_params &params, cudaStream_t stream) {
-    constexpr int Headdim = 256;
+    constexpr static int Headdim = 256;
     int device;
     cudaGetDevice(&device);
     int max_smem_per_sm, max_smem_per_block;

{xformers-0.0.23.dev651/xformers/_flash_attn → xformers-0.0.23.dev657/third_party/flash-attention/flash_attn}/__init__.py

@@ -1,4 +1,4 @@
-__version__ = "2.3.0"
+__version__ = "2.3.2"
 
 from flash_attn.flash_attn_interface import (
     flash_attn_func,

{xformers-0.0.23.dev651 → xformers-0.0.23.dev657}/third_party/flash-attention/flash_attn/flash_attn_interface.py

@@ -928,6 +928,7 @@ def flash_attn_with_kvcache(
     rotary_cos=None,
     rotary_sin=None,
     cache_seqlens: Optional[Union[(int, torch.Tensor)]] = None,
+    cache_batch_idx: Optional[torch.Tensor] = None,
     softmax_scale=None,
     causal=False,
     window_size=(-1, -1),  # -1 means infinite context window
@@ -978,8 +979,8 @@ def flash_attn_with_kvcache(
 
     Arguments:
         q: (batch_size, seqlen, nheads, headdim)
-        k_cache: (batch_size, seqlen_cache, nheads_k, headdim)
-        v_cache: (batch_size, seqlen_cache, nheads_k, headdim)
+        k_cache: (batch_size_cache, seqlen_cache, nheads_k, headdim)
+        v_cache: (batch_size_cache, seqlen_cache, nheads_k, headdim)
         k [optional]: (batch_size, seqlen_new, nheads_k, headdim). If not None, we concatenate
             k with k_cache, starting at the indices specified by cache_seqlens.
         v [optional]: (batch_size, seqlen_new, nheads_k, headdim). Similar to k.
@@ -988,6 +989,10 @@ def flash_attn_with_kvcache(
         rotary_sin [optional]: (seqlen_ro, rotary_dim / 2). Similar to rotary_cos.
         cache_seqlens: int, or (batch_size,), dtype torch.int32. The sequence lengths of the
             KV cache.
+        cache_batch_idx: (batch_size,), dtype torch.int32. The indices used to index into the KV cache.
+            If None, we assume that the batch indices are [0, 1, 2, ..., batch_size - 1].
+            If the indices are not distinct, and k and v are provided, the values updated in the cache
+                 might come from any of the duplicate indices.
         softmax_scale: float. The scaling of QK^T before applying softmax.
             Default to 1 / sqrt(headdim).
         causal: bool. Whether to apply causal attention mask (e.g., for auto-regressive modeling).
@@ -1014,6 +1019,8 @@ def flash_attn_with_kvcache(
         cache_seqlens = torch.full(
             (k_cache.shape[0],), cache_seqlens, dtype=torch.int32, device=k_cache.device
         )
+        cache_seqlens = maybe_contiguous(cache_seqlens)
+    cache_batch_idx = maybe_contiguous(cache_batch_idx)
     out, softmax_lse = flash_attn_cuda.fwd_kvcache(
         q,
         k_cache,
@@ -1023,6 +1030,7 @@ def flash_attn_with_kvcache(
         cache_seqlens,
         rotary_cos,
         rotary_sin,
+        cache_batch_idx,
         None,
         softmax_scale,
         causal,

xformers-0.0.23.dev657/third_party/flash-attention/version.txt

@@ -0,0 +1 @@
+v2.3.2

{xformers-0.0.23.dev651/third_party/flash-attention/flash_attn → xformers-0.0.23.dev657/xformers/_flash_attn}/__init__.py

@@ -1,4 +1,4 @@
-__version__ = "2.3.0"
+__version__ = "2.3.2"
 
 from flash_attn.flash_attn_interface import (
     flash_attn_func,

{xformers-0.0.23.dev651 → xformers-0.0.23.dev657}/xformers/_flash_attn/flash_attn_interface.py

@@ -928,6 +928,7 @@ def flash_attn_with_kvcache(
     rotary_cos=None,
     rotary_sin=None,
     cache_seqlens: Optional[Union[(int, torch.Tensor)]] = None,
+    cache_batch_idx: Optional[torch.Tensor] = None,
     softmax_scale=None,
     causal=False,
     window_size=(-1, -1),  # -1 means infinite context window
@@ -978,8 +979,8 @@ def flash_attn_with_kvcache(
 
     Arguments:
         q: (batch_size, seqlen, nheads, headdim)
-        k_cache: (batch_size, seqlen_cache, nheads_k, headdim)
-        v_cache: (batch_size, seqlen_cache, nheads_k, headdim)
+        k_cache: (batch_size_cache, seqlen_cache, nheads_k, headdim)
+        v_cache: (batch_size_cache, seqlen_cache, nheads_k, headdim)
         k [optional]: (batch_size, seqlen_new, nheads_k, headdim). If not None, we concatenate
             k with k_cache, starting at the indices specified by cache_seqlens.
         v [optional]: (batch_size, seqlen_new, nheads_k, headdim). Similar to k.
@@ -988,6 +989,10 @@ def flash_attn_with_kvcache(
         rotary_sin [optional]: (seqlen_ro, rotary_dim / 2). Similar to rotary_cos.
         cache_seqlens: int, or (batch_size,), dtype torch.int32. The sequence lengths of the
             KV cache.
+        cache_batch_idx: (batch_size,), dtype torch.int32. The indices used to index into the KV cache.
+            If None, we assume that the batch indices are [0, 1, 2, ..., batch_size - 1].
+            If the indices are not distinct, and k and v are provided, the values updated in the cache
+                 might come from any of the duplicate indices.
         softmax_scale: float. The scaling of QK^T before applying softmax.
             Default to 1 / sqrt(headdim).
         causal: bool. Whether to apply causal attention mask (e.g., for auto-regressive modeling).
@@ -1014,6 +1019,8 @@ def flash_attn_with_kvcache(
         cache_seqlens = torch.full(
             (k_cache.shape[0],), cache_seqlens, dtype=torch.int32, device=k_cache.device
         )
+        cache_seqlens = maybe_contiguous(cache_seqlens)
+    cache_batch_idx = maybe_contiguous(cache_batch_idx)
     out, softmax_lse = flash_attn_cuda.fwd_kvcache(
         q,
         k_cache,
@@ -1023,6 +1030,7 @@ def flash_attn_with_kvcache(
         cache_seqlens,
         rotary_cos,
         rotary_sin,
+        cache_batch_idx,
         None,
         softmax_scale,
         causal,

xformers-0.0.23.dev657/xformers/version.py

@@ -0,0 +1,2 @@
+# noqa: C801
+__version__ = "0.0.23.dev657"

{xformers-0.0.23.dev651 → xformers-0.0.23.dev657/xformers.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: xformers
-Version: 0.0.23.dev651
+Version: 0.0.23.dev657
 Summary: XFormers: A collection of composable Transformer building blocks.
 Home-page: https://facebookresearch.github.io/xformers/
 Author: Facebook AI Research

xformers-0.0.23.dev651/third_party/flash-attention/version.txt

@@ -1 +0,0 @@
-v2.3.0

xformers-0.0.23.dev651/xformers/version.py

@@ -1,2 +0,0 @@
-# noqa: C801
-__version__ = "0.0.23.dev651"