llama : make all KQ masks f16 if FA is used, remove zero attention bias, remove raw_k...

* llama : make all KQ masks (except the lightning indexer one) f16 if FA is used and remove zero attention bias in DeepSeek V4

* llama : remove dead code that repeats unified raw_k cache for each stream in DeepSeek V4 - no longer needed as raw_k is always non-unified.

---------

Co-authored-by: Stanisław Szymczyk <redacted>

diff --git a/src/llama-graph.cpp b/src/llama-graph.cpp
index 8c9228b38f9efd0064a822242e58baa2744e0f18..a8fd11ebc385ebd4d8c90a97794c7a6dd7b57231 100644 (file)
--- a/src/llama-graph.cpp
+++ b/src/llama-graph.cpp
@@ -646,7 +646,7 @@ static void dsv4_set_kq_mask(
         return;
     }
 
-    GGML_ASSERT(dst->type == GGML_TYPE_F32);
+    GGML_ASSERT(dst->type == GGML_TYPE_F32 || dst->type == GGML_TYPE_F16);
     GGML_ASSERT(n_stream > 0);
     GGML_ASSERT(n_tokens%n_stream == 0);
     GGML_ASSERT(dst->ne[0] == plan.n_kv);
@@ -656,13 +656,27 @@ static void dsv4_set_kq_mask(
     GGML_ASSERT((int64_t) plan.n_visible.size() == (int64_t) n_tokens);
     GGML_ASSERT(ggml_backend_buffer_is_host(dst->buffer));
 
-    float * data = (float *) dst->data;
+    if (dst->type == GGML_TYPE_F32) {
+        float * data = (float *) dst->data;
 
-    for (int64_t i = 0; i < (int64_t) n_tokens; ++i) {
-        const int32_t n_visible = plan.n_visible[i];
+        for (int64_t i = 0; i < (int64_t) n_tokens; ++i) {
+            const int32_t n_visible = plan.n_visible[i];
 
-        for (int64_t j = 0; j < dst->ne[0]; ++j) {
-            data[i*dst->ne[0] + j] = j < n_visible ? 0.0f : -INFINITY;
+            for (int64_t j = 0; j < dst->ne[0]; ++j) {
+                data[i*dst->ne[0] + j] = j < n_visible ? 0.0f : -INFINITY;
+            }
+        }
+    } else if (dst->type == GGML_TYPE_F16) {
+        ggml_fp16_t * data = (ggml_fp16_t *) dst->data;
+        const ggml_fp16_t fp16_ninf = llama_cast<ggml_fp16_t>(-INFINITY);
+        const ggml_fp16_t fp16_zero = llama_cast<ggml_fp16_t>(0.0f);
+
+        for (int64_t i = 0; i < (int64_t) n_tokens; ++i) {
+            const int32_t n_visible = plan.n_visible[i];
+
+            for (int64_t j = 0; j < dst->ne[0]; ++j) {
+                data[i*dst->ne[0] + j] = j < n_visible ? fp16_zero : fp16_ninf;
+            }
         }
     }
 }
@@ -679,8 +693,7 @@ static ggml_tensor * dsv4_build_raw_kq_mask(
     GGML_ASSERT(n_stream > 0);
     GGML_ASSERT(n_tokens%n_stream == 0);
 
-    const bool use_fattn = cparams.flash_attn && (!cparams.kv_unified || n_stream == 1);
-    const auto type = use_fattn ? GGML_TYPE_F16 : GGML_TYPE_F32;
+    const auto type = cparams.flash_attn ? GGML_TYPE_F16 : GGML_TYPE_F32;
 
     ggml_tensor * res = ggml_new_tensor_4d(ctx, type, n_kv, n_tokens/n_stream, 1, n_stream);
     ggml_set_input(res);
@@ -814,6 +827,7 @@ static void dsv4_build_comp_inputs(
         llm_graph_input_dsv4::comp_input & inp,
         const llama_kv_cache_dsv4_context::comp_plan & plan,
         const char * name,
+        const llama_cparams & cparams,
         int64_t n_stream) {
     inp.state_pos = dsv4_build_input_1d(ctx, GGML_TYPE_I32, plan.state_pos.size(), std::string("dsv4_") + name + "_state_pos");
     inp.state_persist_src_idxs = dsv4_build_input_1d(ctx, GGML_TYPE_I32, plan.state_persist_src_idxs.size(), std::string("dsv4_") + name + "_state_persist_src_idxs");
@@ -828,7 +842,7 @@ static void dsv4_build_comp_inputs(
         GGML_ASSERT(n_stream > 0);
         GGML_ASSERT(n_tokens%n_stream == 0);
 
-        inp.kq_mask = ggml_new_tensor_4d(ctx, GGML_TYPE_F32, plan.n_kv, n_tokens/n_stream, 1, n_stream);
+        inp.kq_mask = ggml_new_tensor_4d(ctx, cparams.flash_attn && strcmp(name, "lid") != 0 ? GGML_TYPE_F16 : GGML_TYPE_F32, plan.n_kv, n_tokens/n_stream, 1, n_stream);
         ggml_set_input(inp.kq_mask);
         ggml_set_name(inp.kq_mask, (std::string("dsv4_") + name + "_kq_mask").c_str());
     }
@@ -3076,9 +3090,9 @@ llm_graph_input_dsv4 * llm_graph_context::build_inp_dsv4() const {
     inp_raw->self_k_rot = raw_ctx->build_input_k_rot(ctx0);
     auto inp = std::make_unique<llm_graph_input_dsv4>(cparams, std::move(inp_raw), mctx_cur);
 
-    dsv4_build_comp_inputs(ctx0, inp->inp_csa, mctx_cur->get_csa_plan(ubatch), "csa", n_stream);
-    dsv4_build_comp_inputs(ctx0, inp->inp_hca, mctx_cur->get_hca_plan(ubatch), "hca", n_stream);
-    dsv4_build_comp_inputs(ctx0, inp->inp_lid, mctx_cur->get_lid_plan(ubatch), "lid", n_stream);
+    dsv4_build_comp_inputs(ctx0, inp->inp_csa, mctx_cur->get_csa_plan(ubatch), "csa", cparams, n_stream);
+    dsv4_build_comp_inputs(ctx0, inp->inp_hca, mctx_cur->get_hca_plan(ubatch), "hca", cparams, n_stream);
+    dsv4_build_comp_inputs(ctx0, inp->inp_lid, mctx_cur->get_lid_plan(ubatch), "lid", cparams, n_stream);
     inp->inp_csa.k_rot = mctx_cur->get_csa()->build_input_k_rot(ctx0);
     inp->inp_hca.k_rot = mctx_cur->get_hca()->build_input_k_rot(ctx0);
     inp->inp_lid.k_rot = mctx_cur->get_lid()->build_input_k_rot(ctx0);

diff --git a/src/models/deepseek4.cpp b/src/models/deepseek4.cpp
index 3fb5bff1bb51b996167a4662e8f0a245ed38a39e..07aa477e1e47ed05b52f21632e3c563574dfc5d6 100644 (file)
--- a/src/models/deepseek4.cpp
+++ b/src/models/deepseek4.cpp
@@ -184,32 +184,6 @@ static ggml_tensor * dsv4_with_zero_dep(ggml_context * ctx, ggml_tensor * t, ggm
     return ggml_add(ctx, t, zero);
 }
 
-// Raw SWA K is stored once, but compressed K/masks can carry a stream axis.
-// Repeat raw K at graph build time before concatenating raw and compressed K.
-static ggml_tensor * dsv4_repeat_streams(ggml_context * ctx, ggml_tensor * t, int64_t n_stream) {
-    if (t->ne[3] == n_stream) {
-        return t;
-    }
-
-    GGML_ASSERT(t->ne[3] == 1);
-    return ggml_repeat_4d(ctx, t, t->ne[0], t->ne[1], t->ne[2], n_stream);
-}
-
-static ggml_tensor * dsv4_build_kq_zero_bias(
-        ggml_context * ctx,
-        const llama_cparams & cparams,
-        ggml_tensor * kq_mask,
-        int64_t n_head) {
-    if (!cparams.kv_unified || !cparams.flash_attn || kq_mask->ne[3] == 1) {
-        return nullptr;
-    }
-
-    // Keep multi-stream unified DSV4 on the explicit attention path.
-    ggml_tensor * res = ggml_new_tensor_4d(ctx, GGML_TYPE_F32,
-            kq_mask->ne[0], kq_mask->ne[1], n_head, kq_mask->ne[3]);
-    return ggml_fill(ctx, res, 0.0f);
-}
-
 static constexpr int64_t DSV4_CSA_RATIO  = 4;
 static constexpr int64_t DSV4_HCA_RATIO  = 128;
 
@@ -624,7 +598,7 @@ ggml_tensor * llama_model_deepseek4::graph::build_top_k_mask(
     ggml_tensor * top_k_3d = ggml_view_4d(ctx0, top_k, top_k->ne[0], top_k->ne[1], top_k->ne[3], 1,
             top_k->nb[1], top_k->nb[2], top_k->ne[3]*top_k->nb[3], 0);
 
-    ggml_tensor * zeros = ggml_new_tensor_4d(ctx0, GGML_TYPE_F32, 1, top_k_3d->ne[0], top_k_3d->ne[1], top_k_3d->ne[2]);
+    ggml_tensor * zeros = ggml_new_tensor_4d(ctx0, cparams.flash_attn ? GGML_TYPE_F16 : GGML_TYPE_F32, 1, top_k_3d->ne[0], top_k_3d->ne[1], top_k_3d->ne[2]);
     zeros = ggml_fill(ctx0, zeros, 0.0f);
 
     ggml_tensor * kq_mask_top_k = ggml_set_rows(ctx0, kq_mask_all, zeros, top_k_3d);
@@ -681,26 +655,16 @@ ggml_tensor * llama_model_deepseek4::graph::build_csa_lid_attention(
             csa_k->nb[1], csa_k->nb[2], csa_k->nb[3], 0);
     cb(csa_k, "csa_comp_k", il);
 
-    raw_k = dsv4_repeat_streams(ctx0, raw_k, csa_k->ne[3]);
-
     ggml_tensor * k_all = ggml_concat(ctx0, raw_k, csa_k, 2);
     cb(k_all, "csa_k_all", il);
 
     ggml_tensor * raw_mask = inp_attn->get_kq_mask();
     ggml_tensor * csa_mask = build_top_k_mask(inp_csa.kq_mask, top_k, "csa_top_k_mask", il);
-    const bool use_fattn = cparams.flash_attn && (!cparams.kv_unified || csa_mask->ne[3] == 1);
-    if (use_fattn && csa_mask->type != GGML_TYPE_F16) {
-        csa_mask = ggml_cast(ctx0, csa_mask, GGML_TYPE_F16);
-    }
-    if (raw_mask->type != csa_mask->type) {
-        raw_mask = ggml_cast(ctx0, raw_mask, csa_mask->type);
-    }
 
     ggml_tensor * kq_mask = ggml_concat(ctx0, raw_mask, csa_mask, 0);
     cb(kq_mask, "csa_lid_kq_mask", il);
 
-    ggml_tensor * kq_b = dsv4_build_kq_zero_bias(ctx0, cparams, kq_mask, q->ne[1]);
-    ggml_tensor * out = build_attn_mha(q, k_all, k_all, kq_b, kq_mask, sinks, nullptr, kq_scale, il);
+    ggml_tensor * out = build_attn_mha(q, k_all, k_all, nullptr, kq_mask, sinks, nullptr, kq_scale, il);
     if (k_rot) {
         out = llama_mul_mat_hadamard(ctx0, out, k_rot);
     }
@@ -746,26 +710,16 @@ ggml_tensor * llama_model_deepseek4::graph::build_hca_attention(
             hca_k->nb[1], hca_k->nb[2], hca_k->nb[3], 0);
     cb(hca_k, "hca_comp_k", il);
 
-    raw_k = dsv4_repeat_streams(ctx0, raw_k, hca_k->ne[3]);
-
     ggml_tensor * k_all = ggml_concat(ctx0, raw_k, hca_k, 2);
     cb(k_all, "hca_k_all", il);
 
     ggml_tensor * raw_mask = inp_attn->get_kq_mask();
     ggml_tensor * hca_mask = inp_hca.kq_mask;
-    const bool use_fattn = cparams.flash_attn && (!cparams.kv_unified || hca_mask->ne[3] == 1);
-    if (use_fattn && hca_mask->type != GGML_TYPE_F16) {
-        hca_mask = ggml_cast(ctx0, hca_mask, GGML_TYPE_F16);
-    }
-    if (raw_mask->type != hca_mask->type) {
-        raw_mask = ggml_cast(ctx0, raw_mask, hca_mask->type);
-    }
 
     ggml_tensor * kq_mask = ggml_concat(ctx0, raw_mask, hca_mask, 0);
     cb(kq_mask, "hca_kq_mask", il);
 
-    ggml_tensor * kq_b = dsv4_build_kq_zero_bias(ctx0, cparams, kq_mask, q->ne[1]);
-    ggml_tensor * out = build_attn_mha(q, k_all, k_all, kq_b, kq_mask, sinks, nullptr, kq_scale, il);
+    ggml_tensor * out = build_attn_mha(q, k_all, k_all, nullptr, kq_mask, sinks, nullptr, kq_scale, il);
     if (k_rot) {
         out = llama_mul_mat_hadamard(ctx0, out, k_rot);
     }
@@ -800,10 +754,8 @@ ggml_tensor * llama_model_deepseek4::graph::build_raw_attention(
     ggml_tensor * kq_mask = inp_attn->get_kq_mask();
 
     ggml_tensor * k = mctx_cur->get_k(ctx0, il);
-    k = dsv4_repeat_streams(ctx0, k, kq_mask->ne[3]);
 
-    ggml_tensor * kq_b = dsv4_build_kq_zero_bias(ctx0, cparams, kq_mask, q->ne[1]);
-    ggml_tensor * out = build_attn_mha(q, k, k, kq_b, kq_mask, sinks, nullptr, kq_scale, il);
+    ggml_tensor * out = build_attn_mha(q, k, k, nullptr, kq_mask, sinks, nullptr, kq_scale, il);
     if (k_rot) {
         out = llama_mul_mat_hadamard(ctx0, out, k_rot);
     }