llama: refactor fused ops (#24646)

diff --git a/src/llama-context.cpp b/src/llama-context.cpp
index 0465430df43a5dd77d33c272793ecc98fbc6a2a9..5edfc85abfdac202fda43f86db29a5628d6a3583 100644 (file)
--- a/src/llama-context.cpp
+++ b/src/llama-context.cpp
@@ -17,6 +17,7 @@
 #include <cstring>
 #include <limits>
 #include <stdexcept>
+#include <string>
 
 //
 // llama_context
@@ -30,6 +31,30 @@ static llm_graph_type ctx_type_to_graph_type(llama_context_type ctx_type) {
     throw std::runtime_error("Unsupported ctx type");
 }
 
+struct llm_fused_op_probe {
+    llm_fused_op op;
+    const char * name;
+    uint32_t n_tokens_per_seq;
+};
+
+static const llm_fused_op_probe llm_fused_op_flash_attn_probe = {
+    /*.op               =*/ LLM_FUSED_OP_FLASH_ATTN,
+    /*.name             =*/ "Flash Attention",
+    /*.n_tokens_per_seq =*/ 1,
+};
+
+static const llm_fused_op_probe llm_fused_op_gdn_ar_probe = {
+    /*.op               =*/ LLM_FUSED_OP_GDN_AR,
+    /*.name             =*/ "fused Gated Delta Net (autoregressive)",
+    /*.n_tokens_per_seq =*/ 1,
+};
+
+static const llm_fused_op_probe llm_fused_op_gdn_ch_probe = {
+    /*.op               =*/ LLM_FUSED_OP_GDN_CH,
+    /*.name             =*/ "fused Gated Delta Net (chunked)",
+    /*.n_tokens_per_seq =*/ 16,
+};
+
 llama_context::llama_context(
         const llama_model & model,
               llama_context_params params) :
@@ -436,6 +461,69 @@ llama_context::~llama_context() {
     ggml_opt_free(opt_ctx);
 }
 
+void llama_context::resolve_fused_ops(const llama_memory_context_i * mctx, uint32_t n_seqs) {
+    const char * func = __func__;
+    auto resolve = [&](const llm_fused_op_probe & probe, bool & enabled) {
+        if (!enabled) {
+            return;
+        }
+
+        const uint32_t n_tokens_probe = probe.n_tokens_per_seq*n_seqs;
+
+        auto * gf = graph_reserve(n_tokens_probe, n_seqs, n_tokens_probe, mctx, true);
+        if (!gf) {
+            throw std::runtime_error(std::string("failed to reserve graph for ") + probe.name + " check");
+        }
+
+        bool device_mismatch = false;
+        for (const auto & node : get_gf_res_reserve()->get_fused_nodes()) {
+            if (node.op != probe.op) {
+                continue;
+            }
+
+            GGML_ASSERT(node.il >= 0);
+
+            ggml_backend_t backend_fused = ggml_backend_sched_get_tensor_backend(sched.get(), node.tensor);
+            ggml_backend_dev_t device_fused = backend_fused ? ggml_backend_get_device(backend_fused) : nullptr;
+
+            // TODO: make this descriptor-specific; model.dev_layer() preserves the current behavior,
+            // but is still wrong for cases like --no-kv-offload.
+            ggml_backend_dev_t device_layer = model.dev_layer(node.il);
+
+            if (device_fused != device_layer) {
+                LLAMA_LOG_WARN("%s: layer %d is assigned to device %s but %s "
+                        "is assigned to device %s (usually due to missing support)\n",
+                        func, node.il,
+                        device_layer ? ggml_backend_dev_name(device_layer) : "none",
+                        probe.name,
+                        device_fused ? ggml_backend_dev_name(device_fused) : "none");
+                device_mismatch = true;
+                break;
+            }
+        }
+
+        if (device_mismatch) {
+            enabled = false;
+            LLAMA_LOG_WARN("%s: %s not supported, set to disabled\n", func, probe.name);
+        } else {
+            enabled = true;
+            LLAMA_LOG_INFO("%s: %s enabled\n", func, probe.name);
+        }
+    };
+
+    if (cparams.auto_fa) {
+        resolve(llm_fused_op_flash_attn_probe, cparams.flash_attn);
+        cparams.auto_fa = false;
+    }
+
+    if (cparams.auto_fgdn) {
+        LLAMA_LOG_INFO("%s: resolving fused Gated Delta Net support:\n", func);
+        resolve(llm_fused_op_gdn_ar_probe, cparams.fused_gdn_ar);
+        resolve(llm_fused_op_gdn_ch_probe, cparams.fused_gdn_ch);
+        cparams.auto_fgdn = false;
+    }
+}
+
 void llama_context::sched_reserve() {
     if (!sched_need_reserve) {
         return;
@@ -475,128 +563,7 @@ void llama_context::sched_reserve() {
 
     LLAMA_LOG_DEBUG("%s: worst-case: n_tokens = %d, n_seqs = %d, n_outputs = %d\n", __func__, n_tokens, n_seqs, n_outputs);
 
-    // resolve automatic Flash Attention use
-    if (cparams.auto_fa) {
-        auto * gf = graph_reserve(1, n_seqs, n_outputs, mctx.get(), true);
-        if (!gf) {
-            throw std::runtime_error("failed to reserve graph for Flash Attention check");
-        }
-
-        const size_t prefix_len = strlen(LLAMA_TENSOR_NAME_FATTN) + 1;
-        bool fa_device_mismatch = false;
-        for (int i = 0; i < ggml_graph_n_nodes(gf); i++) {
-            ggml_tensor * n = ggml_graph_node(gf, i);
-            if (n->op != GGML_OP_FLASH_ATTN_EXT) {
-                continue;
-            }
-            ggml_backend_dev_t device_fa = ggml_backend_get_device(ggml_backend_sched_get_tensor_backend(sched.get(), n));
-
-            // TODO: instead of the tensor names, use a map to keep track of which (FA) tensors belong to which layer
-            GGML_ASSERT(strncmp(n->name, LLAMA_TENSOR_NAME_FATTN "-", prefix_len) == 0);
-            const int il = std::stoi(n->name + prefix_len);
-            ggml_backend_dev_t device_kv = model.dev_layer(il);
-            if (device_fa != device_kv) {
-                LLAMA_LOG_WARN("%s: layer %d is assigned to device %s but the Flash Attention tensor "
-                        "is assigned to device %s (usually due to missing support)\n",
-                        __func__, il, ggml_backend_dev_name(device_kv), ggml_backend_dev_name(device_fa));
-                // FIXME: fa_device_mismatch logic is wrong for --no-kv-offload, but this is broken anyways
-                fa_device_mismatch = true;
-                break;
-            }
-        }
-
-        if (fa_device_mismatch) {
-            cparams.flash_attn = false;
-            LLAMA_LOG_WARN("%s: Flash Attention was auto, set to disabled\n", __func__);
-        } else {
-            cparams.flash_attn = true;
-            LLAMA_LOG_INFO("%s: Flash Attention was auto, set to enabled\n", __func__);
-        }
-
-        cparams.auto_fa = false;
-    }
-
-    if (cparams.auto_fgdn) {
-        LLAMA_LOG_INFO("%s: resolving fused Gated Delta Net support:\n", __func__);
-
-        if (cparams.fused_gdn_ar) {
-            auto * gf = graph_reserve(1, n_seqs, n_outputs, mctx.get(), true);
-            if (!gf) {
-                throw std::runtime_error("failed to reserve graph for fused Gated Delta Net check (autoregressive)");
-            }
-
-            const size_t prefix_len = strlen(LLAMA_TENSOR_NAME_FGDN_AR) + 1;
-            bool gdn_device_mismatch = false;
-            for (int i = 0; i < ggml_graph_n_nodes(gf); i++) {
-                ggml_tensor * n = ggml_graph_node(gf, i);
-                if (n->op != GGML_OP_GATED_DELTA_NET) {
-                    continue;
-                }
-                ggml_backend_dev_t device_gdn = ggml_backend_get_device(ggml_backend_sched_get_tensor_backend(sched.get(), n));
-
-                GGML_ASSERT(strncmp(n->name, LLAMA_TENSOR_NAME_FGDN_AR "-", prefix_len) == 0);
-                const int il = std::stoi(n->name + prefix_len);
-                ggml_backend_dev_t device_kv = model.dev_layer(il);
-                if (device_gdn != device_kv) {
-                    LLAMA_LOG_WARN("%s: layer %d is assigned to device %s but the fused Gated Delta Net tensor "
-                            "is assigned to device %s (usually due to missing support)\n",
-                            __func__, il, ggml_backend_dev_name(device_kv), ggml_backend_dev_name(device_gdn));
-                    gdn_device_mismatch = true;
-                    break;
-                }
-            }
-
-            if (gdn_device_mismatch) {
-                cparams.fused_gdn_ar = false;
-                LLAMA_LOG_WARN("%s: fused Gated Delta Net (autoregressive) not supported, set to disabled\n", __func__);
-            } else {
-                LLAMA_LOG_INFO("%s: fused Gated Delta Net (autoregressive) enabled\n", __func__);
-            }
-        }
-
-        if (cparams.fused_gdn_ch) {
-            // more than one token in the batch per sequence in order to take the chunked path
-            // note: n_outputs must match n_tokens for embedding models with mean/rank pooling,
-            // because build_pooling creates inp_mean with shape [n_tokens, n_seqs] and multiplies
-            // it with t_embd which is reduced to [n_outputs, ...] via out_ids. if n_outputs != n_tokens,
-            // the ggml_mul_mat assertion fails.
-            const uint32_t n_tokens_ch = 16*n_seqs;
-            auto * gf = graph_reserve(n_tokens_ch, n_seqs, n_tokens_ch, mctx.get(), true);
-            if (!gf) {
-                throw std::runtime_error("failed to reserve graph for fused Gated Delta Net check (chunked)");
-            }
-
-            const size_t prefix_len = strlen(LLAMA_TENSOR_NAME_FGDN_CH) + 1;
-            bool gdn_device_mismatch = false;
-            for (int i = 0; i < ggml_graph_n_nodes(gf); i++) {
-                ggml_tensor * n = ggml_graph_node(gf, i);
-                if (n->op != GGML_OP_GATED_DELTA_NET) {
-                    continue;
-                }
-                ggml_backend_dev_t device_gdn = ggml_backend_get_device(ggml_backend_sched_get_tensor_backend(sched.get(), n));
-
-                GGML_ASSERT(strncmp(n->name, LLAMA_TENSOR_NAME_FGDN_CH "-", prefix_len) == 0);
-                const int il = std::stoi(n->name + prefix_len);
-                ggml_backend_dev_t device_kv = model.dev_layer(il);
-                if (device_gdn != device_kv) {
-                    LLAMA_LOG_WARN("%s: layer %d is assigned to device %s but the fused Gated Delta Net tensor "
-                            "is assigned to device %s (usually due to missing support)\n",
-                            __func__, il, ggml_backend_dev_name(device_kv), ggml_backend_dev_name(device_gdn));
-                    gdn_device_mismatch = true;
-                    break;
-                }
-            }
-
-            if (gdn_device_mismatch) {
-                cparams.fused_gdn_ch = false;
-                LLAMA_LOG_WARN("%s: fused Gated Delta Net (chunked) not supported, set to disabled\n", __func__);
-            } else {
-                LLAMA_LOG_INFO("%s: fused Gated Delta Net (chunked) enabled\n", __func__);
-            }
-        }
-
-        cparams.auto_fgdn = false;
-    }
+    resolve_fused_ops(mctx.get(), n_seqs);
 
     // reserve worst-case graph
     int n_splits_pp = -1;

diff --git a/src/llama-context.h b/src/llama-context.h
index f8b7805871ef60d70693b9495c9d9047b7768e80..bf91daa8b562aa66d15b08ca559b6baa09ab7855 100644 (file)
--- a/src/llama-context.h
+++ b/src/llama-context.h
@@ -262,6 +262,10 @@ private:
 
     llm_graph_cb graph_get_cb() const;
 
+    // disable auto fused ops (Flash Attention, Gated Delta Net) whose op lands on a device
+    // that differs from the layer it belongs to (usually due to missing backend support)
+    void resolve_fused_ops(const llama_memory_context_i * mctx, uint32_t n_seqs);
+
     // TODO: read/write lora adapters and cvec
     size_t state_write_data(llama_io_write_i & io);
     size_t state_read_data (llama_io_read_i  & io);

diff --git a/src/llama-graph.cpp b/src/llama-graph.cpp
index d78ddf827b3b5d4bdba6c27028fba3bb0299b4b2..8c9228b38f9efd0064a822242e58baa2744e0f18 100644 (file)
--- a/src/llama-graph.cpp
+++ b/src/llama-graph.cpp
@@ -1192,6 +1192,7 @@ void llm_graph_result::reset() {
     params = {};
 
     inputs.clear();
+    fused_nodes.clear();
 
     buf_compute_meta.resize(ggml_tensor_overhead()*max_nodes + ggml_graph_overhead_custom(max_nodes, false));
 
@@ -1293,6 +1294,10 @@ llm_graph_input_i * llm_graph_result::add_input(llm_graph_input_ptr input) {
     return inputs.back().get();
 }
 
+void llm_graph_result::add_fused_node(llm_graph_fused_node result) {
+    fused_nodes.push_back(result);
+}
+
 void llm_graph_result::set_params(const llm_graph_params & params) {
     this->params = params;
 }
@@ -1352,6 +1357,8 @@ void llm_graph_context::cb(ggml_tensor * cur, const char * name, int il) const {
     }
 }
 
+
+
 ggml_tensor * llm_graph_context::build_cvec(
          ggml_tensor * cur,
                  int   il) const {
@@ -2402,7 +2409,7 @@ ggml_tensor * llm_graph_context::build_attn_mha(
 
         cur = ggml_flash_attn_ext(ctx0, q, k, v, kq_mask, kq_scale, hparams.f_max_alibi_bias,
                                   hparams.attn_soft_cap ? hparams.f_attn_logit_softcapping : 0.0f);
-        cb(cur, LLAMA_TENSOR_NAME_FATTN, il);
+        res->add_fused_node({LLM_FUSED_OP_FLASH_ATTN, cur, il});
 
         ggml_flash_attn_ext_add_sinks(cur, sinks);
         ggml_flash_attn_ext_set_prec (cur, GGML_PREC_F32);

diff --git a/src/llama-graph.h b/src/llama-graph.h
index 4b5b75c632abfe301c2de6baa38bdde5d09a912c..97141ef93be7a43c0a4700d3c13397a3276e35da 100644 (file)
--- a/src/llama-graph.h
+++ b/src/llama-graph.h
@@ -38,6 +38,12 @@ enum llm_graph_type {
     LLM_GRAPH_TYPE_DECODER_MTP,
 };
 
+enum llm_fused_op {
+    LLM_FUSED_OP_FLASH_ATTN,
+    LLM_FUSED_OP_GDN_AR,
+    LLM_FUSED_OP_GDN_CH,
+};
+
 enum llm_ffn_op_type : int {
     LLM_FFN_NONE = 0,           // sentinel: unset; archs must assign before use
     LLM_FFN_SILU,
@@ -775,6 +781,12 @@ struct llm_graph_params {
     }
 };
 
+struct llm_graph_fused_node {
+    llm_fused_op op;
+    ggml_tensor * tensor;
+    int il;
+};
+
 class llm_graph_result {
 public:
     llm_graph_result(int64_t max_nodes);
@@ -808,6 +820,10 @@ public:
 
     llm_graph_input_i * add_input(llm_graph_input_ptr input);
 
+    void add_fused_node(llm_graph_fused_node result);
+
+    const std::vector<llm_graph_fused_node> & get_fused_nodes() const { return fused_nodes; }
+
     void set_params(const llm_graph_params & params);
 
     // important graph nodes
@@ -826,6 +842,7 @@ public:
     std::map<llama_seq_id, ggml_tensor *> t_sampled_probs;
 
     std::vector<llm_graph_input_ptr> inputs;
+    std::vector<llm_graph_fused_node> fused_nodes;
 
     ggml_context_ptr ctx_compute;
 

diff --git a/src/llama-impl.h b/src/llama-impl.h
index 2d06752c94c971ce8fef7e2d7734fe83edd7659b..4988b06d2ca0acc204944db4a8f29baf207d5ff6 100644 (file)
--- a/src/llama-impl.h
+++ b/src/llama-impl.h
@@ -103,7 +103,3 @@ std::string llama_format_tensor_shape(const std::vector<int64_t> & ne);
 std::string llama_format_tensor_shape(const struct ggml_tensor * t);
 
 std::string gguf_kv_to_str(const struct gguf_context * ctx_gguf, int i);
-
-#define LLAMA_TENSOR_NAME_FATTN   "__fattn__"
-#define LLAMA_TENSOR_NAME_FGDN_AR "__fgdn_ar__"
-#define LLAMA_TENSOR_NAME_FGDN_CH "__fgdn_ch__"

diff --git a/src/models/delta-net-base.cpp b/src/models/delta-net-base.cpp
index cf5e380950191dc654dd37285abd258389b47c6a..ad66126477369412c0b77d13044be9fa909f906c 100644 (file)
--- a/src/models/delta-net-base.cpp
+++ b/src/models/delta-net-base.cpp
@@ -401,9 +401,9 @@ std::pair<ggml_tensor *, ggml_tensor *> llm_build_delta_net_base::build_delta_ne
     // K=1: output carries the final state only. state s is 4D [S_v, S_v, H_v, n_seqs].
     ggml_tensor * result = ggml_gated_delta_net(ctx0, q, k, v, g, b, s, /*K=*/1);
     if (n_tokens == 1) {
-        cb(result, LLAMA_TENSOR_NAME_FGDN_AR, il);
+        res->add_fused_node({LLM_FUSED_OP_GDN_AR, result, il});
     } else {
-        cb(result, LLAMA_TENSOR_NAME_FGDN_CH, il);
+        res->add_fused_node({LLM_FUSED_OP_GDN_CH, result, il});
     }
 
     ggml_tensor * output = ggml_view_4d(ctx0, result,
@@ -566,9 +566,9 @@ ggml_tensor * llm_build_delta_net_base::build_recurrent_attn(
     // state s is 4D [S_v, S_v, H_v, n_seqs]; K snapshot slots are written into the output.
     ggml_tensor * gdn_out = ggml_gated_delta_net(ctx0, q, k, v, g, b, s, K);
     if (n_seq_tokens > 1) {
-        cb(gdn_out, LLAMA_TENSOR_NAME_FGDN_CH, il);
+        res->add_fused_node({LLM_FUSED_OP_GDN_CH, gdn_out, il});
     } else {
-        cb(gdn_out, LLAMA_TENSOR_NAME_FGDN_AR, il);
+        res->add_fused_node({LLM_FUSED_OP_GDN_AR, gdn_out, il});
     }
 
     const int64_t attn_score_elems    = S_v * H_v * n_seq_tokens * n_seqs;