ggml : add Flash Attention (#5021)

* ggml : add ggml_flash_attn_ext API

* ggml : fix GQA support in ggml_flash_attn_ext

* ggml : online attention (CPU)

* metal : initial implementation

* metal : f16 precision

* metal : reduce branches

* metal : specialize for head size

* wip : 8 rows per simd group

* wip : 4 rows per simd group

* wip : template for rows per warp

* metal : parallelize across KV size

* metal : parallel reduce across heads

* metal : efficient flash_attn_f16 implementation

* metal : avoid redundant loads of the attention

* metal : scale and mask in matrix form

* metal : fix comment

* llama : avoid ggml_cast, use F32 query

* metal : add parallel reduce version (disabled)

* metal : move output into local memory + optimize

- the result from each simdgroup now stays in the registers
- significantly reduced SRAM usage
- more efficient skipping of -INF blocks
- avoid simdgroup barrier in hot loop
- add comments

* metal : add tests, fix scaling, support C > 32

* metal : improve precision

* ggml : fix f16 mad

* metal : minor

* metal : support Q > 8

* tests : add ATTN tests

* metal : disable buffer allocation logs

* tests : more

* metal : faster inner loop for C == 32

* metal : fix array initialization

* tests : ifdef

* ggml : switch to padded F16 mask for ggml_soft_max, ggml_flash_attn_ext

* ggml : fix ggml_soft_max mask requirement

* cuda : fix soft_max to use correct mask size

* cuda : add flash_attn kernel (wip)

* metal : optimize softmax for C > 32

* metal : optimize softmax

* tests : minor fix

* cuda : avoid zeroing fragments

* tests : update dims

* cuda : fix __hisinf() result check

* cuda : avoid warp_reduce for smax

* cuda : use int instead of int64_t

Noticeably improves performance (thanks to Johannes)

* cuda : make loops use the same loop values

Thanks Johannes again for the tip

* cuda : unroll some of the loops

* cuda : avoid __hisinf branches

* cuda : use half2 in softmax

* cuda : switch to 1 warp for bs > 16

* cuda : speed-up reduce part of the kernel

* cuda : unroll Q*K^T loop

* cuda : fix -INF block check

* cuda : simplify softmax

* cuda : fix matrix names

* cuda : minor

* llama : adapt to F16 KQ_pos

* llama : adapt new models to F16 KQ_mask

* ggml : fix F16 store (ARM NEON)

* llama : fix type of KQ_mask and KQ_pos

* ggml : fix CPU soft_max

* tests : add hs=256

* cuda : fix build

* metal : improve perf via smaller int registers

* cuda : adapt soft_max to F16 mask and pos

* CUDA: faster FlashAttention, kernel for bs == 1

* 16 cols for Phi-2

* no vec for hs, no hs==256 ncols==32 for Volta

* adjust kernel selection logic

* 4 warps, 256 stride for all D

* no ncols == 64

* Multiple parallel blocks for batch size 1

* fix compile warnings

* fix excessive KQ_b loads

* fix cmake build

* fix KV cache padding, NaN from INFINITY (#6438)

* llama : flash_attn cparam + fix defrag

* server: support flash_attn param

* server: bench: enable flash_attn param

* CUDA: refactor host code, dyn. par. blocks

* fix flash_attn_vec_f16 race condition

* flush softmax exp below threshold to 0

* store temp KQ in registers

* Calculate KQ as FP32 if KQV has GGML_PREC_F32

* Add __hgt2_mask implementation for CUDA 11

* fix KQ FP32 precision fpr parallel_blocks > 1

* llama-bench : add -fa,--flash-attn arg

* metal : add BS=1 kernel for flash attention (#6508)

* metal : add BS=1 kernel for flash attention (wip)

* metal : support more than 1 warps

* metal : opts

* metal : opt

* metal : switch to parallel reduce

* metal : reduce registers

* metal : simplify

* metal : initial FA vec kernel

* metal : use F32 attention accumulators

* batched-bench : add fattn arg

* llama : simplify llama_build_kv_store

ggml-ci

* llama : adapt build_olmo to changes

* ggml : fix arm fp16 store on windows

* metal : clean-up

* metal : clean-up kernel code

* metal : minor

* tests : remove benchmarks

ggml-ci

* ggml : fix avx512 const correctness

ggml-ci

* ggml : fix soft_max with bias on CPU

ggml-ci

* common : print --flash-attn in help

* ggml : fix num dimensions in ggml_flash_attn_ext

* llama : force disable flash attention for incompatible models

* ggml : ggml_soft_max support F16/F32 mask/pos

ggml-ci

* cuda : uint -> uint32_t

* cuda : "constexpr dim3" -> "const dim3"

ggml-ci

* cuda : try to fix __hgt2_mask

ggml-ci

* ggml : add TODO's for F16/F32 mask/pos support in other backends

* llama : replace bool need_kq_pos with use_alibi

* llama : prep ALiBi support for BERT models

ggml-ci

* llama : fix n_batch requirements

ggml-ci

* cont

* server : add help for --flash-attn arg

* llama : disable FA for AMD

* tests : remove TMP_ATTN_BENCH

ggml-ci

* llama : support save/load state with FA enabled

ggml-ci

* ci : add CUDA save-load-state tests

ggml-ci

* llama : llama_kv_cache_clear zeroes data + fix save-load seq

ggml-ci

* llama : fix copy-paste errors, add TODO

* llama : disallow incompatible states

* llama : update llama_state_get_size after v_trans field

* metal : remove tmp log

* llama : add static reminder for llama_state_get_size

* metal : fix max nsg

ggml-ci

* ci : fix arg order

ggml-ci

---------

Co-authored-by: Johannes Gäßler <[email protected]>
Co-authored-by: Pierrick HYMBERT <[email protected]>

Author: 3 people

ci/run.sh

@@ -336,7 +336,8 @@ function gg_run_open_llama_3b_v2 {
 
     (time ./bin/imatrix --model ${model_f16} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-imatrix.log
 
-    (time ./bin/save-load-state --model ${model_q4_0} ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
+    (time ./bin/save-load-state     --model ${model_q4_0} ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
+    (time ./bin/save-load-state -fa --model ${model_q4_0} ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
 
     function check_ppl {
         qnt="$1"
@@ -517,7 +518,10 @@ function gg_run_open_llama_7b_v2 {
 
     (time ./bin/imatrix --model ${model_f16} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-imatrix.log
 
-    (time ./bin/save-load-state --model ${model_q4_0} ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
+    (time ./bin/save-load-state     -ngl 10 --model ${model_q4_0} ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
+    (time ./bin/save-load-state -fa -ngl 10 --model ${model_q4_0} ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
+    (time ./bin/save-load-state     -ngl 99 --model ${model_q4_0} ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
+    (time ./bin/save-load-state -fa -ngl 99 --model ${model_q4_0} ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
 
     function check_ppl {
         qnt="$1"

common/common.cpp

@@ -947,6 +947,10 @@ bool gpt_params_find_arg(int argc, char ** argv, const std::string & arg, gpt_pa
         params.cont_batching = true;
         return true;
     }
+    if (arg == "-fa" || arg == "--flash-attn") {
+        params.flash_attn = true;
+        return true;
+    }
     if (arg == "--color") {
         params.use_color = true;
         return true;
@@ -1513,6 +1517,7 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) {
     printf("  -ns N, --sequences N  number of sequences to decode (default: %d)\n", params.n_sequences);
     printf("  -ps N, --p-split N    speculative decoding split probability (default: %.1f)\n", (double)params.p_split);
     printf("  -cb, --cont-batching  enable continuous batching (a.k.a dynamic batching) (default: disabled)\n");
+    printf("  -fa, --flash-attn     enable Flash Attention (default: %s)\n", params.flash_attn ? "enabled" : "disabled");
     printf("  --mmproj MMPROJ_FILE  path to a multimodal projector file for LLaVA. see examples/llava/README.md\n");
     printf("  --image IMAGE_FILE    path to an image file. use with multimodal models. Specify multiple times for batching\n");
     if (llama_supports_mlock()) {
@@ -1885,6 +1890,7 @@ struct llama_context_params llama_context_params_from_gpt_params(const gpt_param
     cparams.cb_eval           = params.cb_eval;
     cparams.cb_eval_user_data = params.cb_eval_user_data;
     cparams.offload_kqv       = !params.no_kv_offload;
+    cparams.flash_attn        = params.flash_attn;
 
     cparams.type_k = kv_cache_type_from_str(params.cache_type_k);
     cparams.type_v = kv_cache_type_from_str(params.cache_type_v);
@@ -2707,6 +2713,7 @@ void dump_non_result_info_yaml(FILE * stream, const gpt_params & params, const l
     fprintf(stream, "seed: %u # default: -1 (random seed)\n", params.seed);
     fprintf(stream, "simple_io: %s # default: false\n", params.simple_io ? "true" : "false");
     fprintf(stream, "cont_batching: %s # default: false\n", params.cont_batching ? "true" : "false");
+    fprintf(stream, "flash_attn: %s # default: false\n", params.flash_attn ? "true" : "false");
     fprintf(stream, "temp: %f # default: 0.8\n", sparams.temp);
 
     const std::vector<float> tensor_split_vector(params.tensor_split, params.tensor_split + llama_max_devices());

common/common.h

@@ -150,6 +150,7 @@ struct gpt_params {
     bool multiline_input   = false; // reverse the usage of `\`
     bool simple_io         = false; // improves compatibility with subprocesses and limited consoles
     bool cont_batching     = true;  // insert new sequences for decoding on-the-fly
+    bool flash_attn        = false; // flash attention
 
     bool input_prefix_bos  = false; // prefix BOS to user inputs, preceding input_prefix
     bool ignore_eos        = false; // ignore generated EOS tokens

examples/batched-bench/batched-bench.cpp

@@ -32,7 +32,7 @@ int main(int argc, char ** argv) {
     gpt_params params;
 
     if (argc == 1 || argv[1][0] == '-') {
-        printf("usage: %s MODEL_PATH [N_KV_MAX] [N_BATCH] [N_UBATCH] [IS_PP_SHARED] [NGL] <PP> <TG> <PL>\n" , argv[0]);
+        printf("usage: %s MODEL_PATH [N_KV_MAX] [N_BATCH] [N_UBATCH] [FATTN] [IS_PP_SHARED] [NGL] <PP> <TG> <PL>\n" , argv[0]);
         printf("  <PP>, <TG> and PL are comma-separated lists of numbers without spaces\n\n");
         printf("  example: %s ggml-model-f16.gguf 2048 2048 512 0 999 128,256,512 128,256 1,2,4,8,16,32\n\n", argv[0]);
         return 1 ;
@@ -41,6 +41,7 @@ int main(int argc, char ** argv) {
     int n_kv_max     = 2048;
     int n_batch      = 2048;
     int n_ubatch     = 512;
+    bool flash_attn  = false;
     int is_pp_shared = 0;
     int n_gpu_layers = 0;
 
@@ -66,23 +67,27 @@ int main(int argc, char ** argv) {
     }
 
     if (argc >= 6) {
-        is_pp_shared = std::atoi(argv[5]);
+        flash_attn = std::atoi(argv[5]);
     }
 
     if (argc >= 7) {
-        n_gpu_layers = std::atoi(argv[6]);
+        is_pp_shared = std::atoi(argv[6]);
     }
 
     if (argc >= 8) {
-        n_pp = parse_list(argv[7]);
+        n_gpu_layers = std::atoi(argv[7]);
     }
 
     if (argc >= 9) {
-        n_tg = parse_list(argv[8]);
+        n_pp = parse_list(argv[8]);
     }
 
     if (argc >= 10) {
-        n_pl = parse_list(argv[9]);
+        n_tg = parse_list(argv[9]);
+    }
+
+    if (argc >= 11) {
+        n_pl = parse_list(argv[10]);
     }
 
     // init LLM
@@ -108,10 +113,11 @@ int main(int argc, char ** argv) {
 
     llama_context_params ctx_params = llama_context_default_params();
 
-    ctx_params.seed      = 1234;
-    ctx_params.n_ctx     = n_kv_max;
-    ctx_params.n_batch   = n_batch;
-    ctx_params.n_ubatch  = n_ubatch;
+    ctx_params.seed       = 1234;
+    ctx_params.n_ctx      = n_kv_max;
+    ctx_params.n_batch    = n_batch;
+    ctx_params.n_ubatch   = n_ubatch;
+    ctx_params.flash_attn = flash_attn;
 
     ctx_params.n_threads       = params.n_threads;
     ctx_params.n_threads_batch = params.n_threads_batch == -1 ? params.n_threads : params.n_threads_batch;
@@ -169,7 +175,7 @@ int main(int argc, char ** argv) {
     }
 
     LOG_TEE("\n");
-    LOG_TEE("%s: n_kv_max = %d, n_batch = %d, n_ubatch = %d, is_pp_shared = %d, n_gpu_layers = %d, n_threads = %u, n_threads_batch = %u\n", __func__, n_kv_max, n_batch, n_ubatch, is_pp_shared, n_gpu_layers, ctx_params.n_threads, ctx_params.n_threads_batch);
+    LOG_TEE("%s: n_kv_max = %d, n_batch = %d, n_ubatch = %d, flash_attn = %d, is_pp_shared = %d, n_gpu_layers = %d, n_threads = %u, n_threads_batch = %u\n", __func__, n_kv_max, n_batch, n_ubatch, flash_attn, is_pp_shared, n_gpu_layers, ctx_params.n_threads, ctx_params.n_threads_batch);
     LOG_TEE("\n");
 
     LOG_TEE("|%6s | %6s | %4s | %6s | %8s | %8s | %8s | %8s | %8s | %8s |\n", "PP",     "TG",     "B",    "N_KV",     "T_PP s",   "S_PP t/s", "T_TG s",   "S_TG t/s", "T s",      "S t/s");

examples/llama-bench/llama-bench.cpp

@@ -174,6 +174,7 @@ struct cmd_params {
     std::vector<llama_split_mode> split_mode;
     std::vector<int> main_gpu;
     std::vector<bool> no_kv_offload;
+    std::vector<bool> flash_attn;
     std::vector<std::vector<float>> tensor_split;
     std::vector<bool> use_mmap;
     std::vector<bool> embeddings;
@@ -195,6 +196,7 @@ static const cmd_params cmd_params_defaults = {
     /* split_mode    */ {LLAMA_SPLIT_MODE_LAYER},
     /* main_gpu      */ {0},
     /* no_kv_offload */ {false},
+    /* flash_attn    */ {false},
     /* tensor_split  */ {std::vector<float>(llama_max_devices(), 0.0f)},
     /* use_mmap      */ {true},
     /* embeddings    */ {false},
@@ -220,6 +222,7 @@ static void print_usage(int /* argc */, char ** argv) {
     printf("  -sm, --split-mode <none|layer|row>  (default: %s)\n", join(transform_to_str(cmd_params_defaults.split_mode, split_mode_str), ",").c_str());
     printf("  -mg, --main-gpu <i>                 (default: %s)\n", join(cmd_params_defaults.main_gpu, ",").c_str());
     printf("  -nkvo, --no-kv-offload <0|1>        (default: %s)\n", join(cmd_params_defaults.no_kv_offload, ",").c_str());
+    printf("  -fa, --flash-attn <0|1>             (default: %s)\n", join(cmd_params_defaults.flash_attn, ",").c_str());
     printf("  -mmp, --mmap <0|1>                  (default: %s)\n", join(cmd_params_defaults.use_mmap, ",").c_str());
     printf("  -embd, --embeddings <0|1>           (default: %s)\n", join(cmd_params_defaults.embeddings, ",").c_str());
     printf("  -ts, --tensor-split <ts0/ts1/..>    (default: 0)\n");
@@ -393,6 +396,13 @@ static cmd_params parse_cmd_params(int argc, char ** argv) {
             }
             auto p = split<bool>(argv[i], split_delim);
             params.no_kv_offload.insert(params.no_kv_offload.end(), p.begin(), p.end());
+        } else if (arg == "-fa" || arg == "--flash-attn") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            auto p = split<bool>(argv[i], split_delim);
+            params.flash_attn.insert(params.flash_attn.end(), p.begin(), p.end());
         } else if (arg == "-mmp" || arg == "--mmap") {
             if (++i >= argc) {
                 invalid_param = true;
@@ -477,6 +487,7 @@ static cmd_params parse_cmd_params(int argc, char ** argv) {
     if (params.split_mode.empty())   { params.split_mode = cmd_params_defaults.split_mode; }
     if (params.main_gpu.empty())     { params.main_gpu = cmd_params_defaults.main_gpu; }
     if (params.no_kv_offload.empty()){ params.no_kv_offload = cmd_params_defaults.no_kv_offload; }
+    if (params.flash_attn.empty())   { params.flash_attn = cmd_params_defaults.flash_attn; }
     if (params.tensor_split.empty()) { params.tensor_split = cmd_params_defaults.tensor_split; }
     if (params.use_mmap.empty())     { params.use_mmap = cmd_params_defaults.use_mmap; }
     if (params.embeddings.empty())   { params.embeddings = cmd_params_defaults.embeddings; }
@@ -498,6 +509,7 @@ struct cmd_params_instance {
     llama_split_mode split_mode;
     int main_gpu;
     bool no_kv_offload;
+    bool flash_attn;
     std::vector<float> tensor_split;
     bool use_mmap;
     bool embeddings;
@@ -532,6 +544,7 @@ struct cmd_params_instance {
         cparams.type_k = type_k;
         cparams.type_v = type_v;
         cparams.offload_kqv = !no_kv_offload;
+        cparams.flash_attn = flash_attn;
         cparams.embeddings = embeddings;
 
         return cparams;
@@ -554,6 +567,7 @@ static std::vector<cmd_params_instance> get_cmd_params_instances(const cmd_param
     for (const auto & tk : params.type_k)
     for (const auto & tv : params.type_v)
     for (const auto & nkvo : params.no_kv_offload)
+    for (const auto & fa : params.flash_attn)
     for (const auto & nt : params.n_threads) {
         for (const auto & n_prompt : params.n_prompt) {
             if (n_prompt == 0) {
@@ -572,6 +586,7 @@ static std::vector<cmd_params_instance> get_cmd_params_instances(const cmd_param
                 /* .split_mode   = */ sm,
                 /* .main_gpu     = */ mg,
                 /* .no_kv_offload= */ nkvo,
+                /* .flash_attn   = */ fa,
                 /* .tensor_split = */ ts,
                 /* .use_mmap     = */ mmp,
                 /* .embeddings   = */ embd,
@@ -596,6 +611,7 @@ static std::vector<cmd_params_instance> get_cmd_params_instances(const cmd_param
                 /* .split_mode   = */ sm,
                 /* .main_gpu     = */ mg,
                 /* .no_kv_offload= */ nkvo,
+                /* .flash_attn   = */ fa,
                 /* .tensor_split = */ ts,
                 /* .use_mmap     = */ mmp,
                 /* .embeddings   = */ embd,
@@ -633,6 +649,7 @@ struct test {
     llama_split_mode split_mode;
     int main_gpu;
     bool no_kv_offload;
+    bool flash_attn;
     std::vector<float> tensor_split;
     bool use_mmap;
     bool embeddings;
@@ -657,6 +674,7 @@ struct test {
         split_mode = inst.split_mode;
         main_gpu = inst.main_gpu;
         no_kv_offload = inst.no_kv_offload;
+        flash_attn = inst.flash_attn;
         tensor_split = inst.tensor_split;
         use_mmap = inst.use_mmap;
         embeddings = inst.embeddings;
@@ -731,7 +749,7 @@ struct test {
             "n_batch", "n_ubatch",
             "n_threads", "type_k", "type_v",
             "n_gpu_layers", "split_mode",
-            "main_gpu", "no_kv_offload",
+            "main_gpu", "no_kv_offload", "flash_attn",
             "tensor_split", "use_mmap", "embeddings",
             "n_prompt", "n_gen", "test_time",
             "avg_ns", "stddev_ns",
@@ -753,7 +771,7 @@ struct test {
         }
         if (field == "cuda" || field == "opencl"  || field == "vulkan" || field == "kompute" || field == "metal" ||
             field == "gpu_blas" || field == "blas" || field == "sycl" ||field == "f16_kv" || field == "no_kv_offload" ||
-            field == "use_mmap" || field == "embeddings") {
+            field == "flash_attn" || field == "use_mmap" || field == "embeddings") {
             return BOOL;
         }
         if (field == "avg_ts" || field == "stddev_ts") {
@@ -787,7 +805,7 @@ struct test {
             std::to_string(n_batch), std::to_string(n_ubatch),
             std::to_string(n_threads), ggml_type_name(type_k), ggml_type_name(type_v),
             std::to_string(n_gpu_layers), split_mode_str(split_mode),
-            std::to_string(main_gpu), std::to_string(no_kv_offload),
+            std::to_string(main_gpu), std::to_string(no_kv_offload), std::to_string(flash_attn),
             tensor_split_str, std::to_string(use_mmap), std::to_string(embeddings),
             std::to_string(n_prompt), std::to_string(n_gen), test_time,
             std::to_string(avg_ns()), std::to_string(stdev_ns()),
@@ -955,6 +973,9 @@ struct markdown_printer : public printer {
         if (field == "no_kv_offload") {
             return "nkvo";
         }
+        if (field == "flash_attn") {
+            return "fa";
+        }
         if (field == "use_mmap") {
             return "mmap";
         }
@@ -1001,6 +1022,9 @@ struct markdown_printer : public printer {
         if (params.no_kv_offload.size() > 1 || params.no_kv_offload != cmd_params_defaults.no_kv_offload) {
             fields.emplace_back("no_kv_offload");
         }
+        if (params.flash_attn.size() > 1 || params.flash_attn != cmd_params_defaults.flash_attn) {
+            fields.emplace_back("flash_attn");
+        }
         if (params.tensor_split.size() > 1 || params.tensor_split != cmd_params_defaults.tensor_split) {
             fields.emplace_back("tensor_split");
         }

examples/server/bench/bench.py

@@ -268,6 +268,7 @@ def start_server_background(args):
     server_args.extend(['--defrag-thold', "0.1"])
     server_args.append('--cont-batching')
     server_args.append('--metrics')
+    server_args.append('--flash-attn')
     server_args.extend(['--log-format', "text"])
     args = [str(arg) for arg in [server_path, *server_args]]
     print(f"bench: starting server with: {' '.join(args)}")

examples/server/server.cpp

@@ -2377,6 +2377,7 @@ static void server_print_usage(const char * argv0, const gpt_params & params, co
     printf("  --embeddings              enable embedding vector output (default: %s)\n", params.embedding ? "enabled" : "disabled");
     printf("  -np N, --parallel N       number of slots for process requests (default: %d)\n", params.n_parallel);
     printf("  -cb, --cont-batching      enable continuous batching (a.k.a dynamic batching) (default: enabled)\n");
+    printf("  -fa, --flash-attn         enable Flash Attention (default: %s)\n", params.flash_attn ? "enabled" : "disabled");
     printf("  -spf FNAME, --system-prompt-file FNAME\n");
     printf("                            set a file to load a system prompt (initial prompt of all slots), this is useful for chat applications.\n");
     printf("  -ctk TYPE, --cache-type-k TYPE\n");
@@ -2742,6 +2743,8 @@ static void server_params_parse(int argc, char ** argv, server_params & sparams,
             params.embedding = true;
         } else if (arg == "-cb" || arg == "--cont-batching") {
             params.cont_batching = true;
+        } else if (arg == "-fa" || arg == "--flash-attn") {
+            params.flash_attn = true;
         } else if (arg == "-np" || arg == "--parallel") {
             if (++i >= argc) {
                 invalid_param = true;

ggml-cuda.cu

@@ -14,6 +14,7 @@
 #include "ggml-cuda/cpy.cuh"
 #include "ggml-cuda/diagmask.cuh"
 #include "ggml-cuda/dmmv.cuh"
+#include "ggml-cuda/fattn.cuh"
 #include "ggml-cuda/getrows.cuh"
 #include "ggml-cuda/im2col.cuh"
 #include "ggml-cuda/mmq.cuh"
@@ -140,6 +141,7 @@ static ggml_cuda_device_info ggml_cuda_init() {
         info.devices[id].cc = 100*prop.major + 10*prop.minor;
 #endif // defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
         info.devices[id].smpb = prop.sharedMemPerBlock;
+        info.devices[id].nsm  = prop.multiProcessorCount;
     }
 
     for (int id = 0; id < info.device_count; ++id) {
@@ -2290,6 +2292,9 @@ static bool ggml_cuda_compute_forward(ggml_backend_cuda_context & ctx, struct gg
         case GGML_OP_ARGSORT:
             ggml_cuda_op_argsort(ctx, dst);
             break;
+        case GGML_OP_FLASH_ATTN_EXT:
+            ggml_cuda_flash_attn_ext(ctx, dst);
+            break;
         default:
             return false;
     }
@@ -2564,6 +2569,7 @@ GGML_CALL static bool ggml_backend_cuda_supports_op(ggml_backend_t backend, cons
         case GGML_OP_ARANGE:
         case GGML_OP_TIMESTEP_EMBEDDING:
         case GGML_OP_LEAKY_RELU:
+        case GGML_OP_FLASH_ATTN_EXT:
             return true;
         default:
             return false;