From ef47ec18da469423c276b683dd9b5741cee7023e Mon Sep 17 00:00:00 2001
From: Georgi Gerganov <ggerganov@gmail.com>
Date: Fri, 1 Dec 2023 10:51:24 +0200
Subject: ggml : add ggml_soft_max_ext (#4256)

* metal : implement soft_max_ext

* cuda : implement soft_max_ext

* ggml : implement soft_max_ext (CPU)

* batched-bench : print threads

ggml-ci

* metal : simplify soft_max encoding

ggml-ci

* cuda : use 512 threads for soft_max instead of 32

* ggml : update soft max cpu

* cuda : do warp-based block reduce

* cuda : increase max block size to 1024

* cuda : fix warp reduction initialization of shared mem

* metal : warp-based reduction for soft max kernel

* metal : warp-based reduce for rms_norm

* metal : simplify soft max kernel

ggml-ci

* alloc : fix build with debug
---
 llama.cpp | 33 ++++++++++++++++++++-------------
 1 file changed, 20 insertions(+), 13 deletions(-)

(limited to 'llama.cpp')

diff --git a/llama.cpp b/llama.cpp
index 1e00ea4a..e74fd723 100644
--- a/llama.cpp
+++ b/llama.cpp
@@ -3704,22 +3704,28 @@ static struct ggml_tensor * llm_build_kqv(
     struct ggml_tensor * kq = ggml_mul_mat(ctx, k, q);
     cb(kq, "kq", il);
 
-    kq = ggml_scale(ctx, kq, kq_scale);
-    cb(kq, "kq_scaled", il);
-
     if (max_alibi_bias > 0.0f) {
-        // TODO: n_head or n_head_kv
-        // TODO: K-shift is likely not working
-        // TODO: change to ggml_add
-        kq = ggml_alibi(ctx, kq, /*n_past*/ 0, n_head, max_alibi_bias);
-        cb(kq, "kq_scaled_alibi", il);
-    }
+        // temporary branch until we figure out how to handle ggml_alibi through ggml_add
+        kq = ggml_scale(ctx, kq, kq_scale);
+        cb(kq, "kq_scaled", il);
+
+        if (max_alibi_bias > 0.0f) {
+            // TODO: n_head or n_head_kv
+            // TODO: K-shift is likely not working
+            // TODO: change to ggml_add
+            kq = ggml_alibi(ctx, kq, /*n_past*/ 0, n_head, max_alibi_bias);
+            cb(kq, "kq_scaled_alibi", il);
+        }
 
-    kq = ggml_add(ctx, kq, kq_mask);
-    cb(kq, "kq_masked", il);
+        kq = ggml_add(ctx, kq, kq_mask);
+        cb(kq, "kq_masked", il);
 
-    kq = ggml_soft_max(ctx, kq);
-    cb(kq, "kq_soft_max", il);
+        kq = ggml_soft_max(ctx, kq);
+        cb(kq, "kq_soft_max", il);
+    } else {
+        kq = ggml_soft_max_ext(ctx, kq, kq_mask, 1.0f/sqrtf(float(n_embd_head)));
+        cb(kq, "kq_soft_max_ext", il);
+    }
 
     // split cached v into n_head heads
     struct ggml_tensor * v =
@@ -5041,6 +5047,7 @@ static const std::unordered_map<const char *, llm_offload_func_e> k_offload_map
     { "kq_scaled_alibi",            OFFLOAD_FUNC_KQ  },
     { "kq_masked",                  OFFLOAD_FUNC_KQ  },
     { "kq_soft_max",                OFFLOAD_FUNC_V   },
+    { "kq_soft_max_ext",            OFFLOAD_FUNC_V   },
     { "v",                          OFFLOAD_FUNC_V   },
     { "kqv",                        OFFLOAD_FUNC_V   },
     { "kqv_merged",                 OFFLOAD_FUNC_V   },
-- 
cgit v1.2.3