model : support for DeepseekV32ForCausalLM with generic DeepSeek Sparse Attention...

* llama : support DeepSeek V3.2 model family (with DSA lightning indexer)

* convert : handle DeepseekV32ForCausalLM architecture

* ggml : support for f16 GGML_OP_FILL

* memory : separate hparams argument in llama_kv_cache constructor

* memory : add llama_kv_cache_dsa memory (KV cache + lightning indexer cache)

* llama : support for LLM_ARCH_DEEPSEEK32

* model : llama_model_deepseek32 implementation

* model : merge two scale operations into one in DSA lightning indexer implementation

* chore : remove unused code

* model : support NVFP4 in DeepSeek V3.2

Co-authored-by: Sigbjørn Skjæret <redacted>
* memory : refactoring TODO

Co-authored-by: ggerganov <redacted>
---------

Co-authored-by: Stanisław Szymczyk <redacted>
Co-authored-by: Sigbjørn Skjæret <redacted>
Co-authored-by: ggerganov <redacted>

diff --git a/ggml/src/ggml-cpu/ops.cpp b/ggml/src/ggml-cpu/ops.cpp
index 7485ba4fc861332d91b230a415ff3be72493fbe9..dc73696ad9fef155509ec217ad1cd6c9c787e1d1 100644 (file)
--- a/ggml/src/ggml-cpu/ops.cpp
+++ b/ggml/src/ggml-cpu/ops.cpp
@@ -2235,8 +2235,42 @@ static void ggml_compute_forward_fill_f32(const ggml_compute_params * params, gg
     }
 }
 
+static void ggml_compute_forward_fill_f16(const ggml_compute_params * params, ggml_tensor * dst) {
+    const ggml_fp16_t c = GGML_CPU_FP32_TO_FP16(ggml_get_op_params_f32(dst, 0));
+
+    GGML_TENSOR_LOCALS(int64_t, ne, dst, ne);
+    GGML_TENSOR_LOCALS(size_t,  nb, dst, nb);
+
+    const auto [ir0, ir1] = get_thread_range(params, dst);
+
+    for (int64_t ir = ir0; ir < ir1; ++ir) {
+        const int64_t i03 = ir/(ne2*ne1);
+        const int64_t i02 = (ir - i03*ne2*ne1)/ne1;
+        const int64_t i01 = (ir - i03*ne2*ne1 - i02*ne1);
+
+        ggml_fp16_t * dst_ptr  = (ggml_fp16_t *) ((char *) dst->data + i03*nb3 + i02*nb2 + i01*nb1);
+
+        ggml_vec_set_f16(ne0, dst_ptr, c);
+    }
+}
+
 void ggml_compute_forward_fill(const ggml_compute_params * params, ggml_tensor * dst) {
-    ggml_compute_forward_fill_f32(params, dst);
+    const ggml_tensor * src0 = dst->src[0];
+
+    switch (src0->type) {
+        case GGML_TYPE_F32:
+            {
+                ggml_compute_forward_fill_f32(params, dst);
+            } break;
+        case GGML_TYPE_F16:
+            {
+                ggml_compute_forward_fill_f16(params, dst);
+            } break;
+        default:
+            {
+                GGML_ABORT("unsupported type for ggml_compute_forward_fill: %s", ggml_type_name(src0->type));
+            }
+    }
 }
 
 // ggml_compute_tri

diff --git a/ggml/src/ggml.c b/ggml/src/ggml.c
index 476c30797956773eb1c6b5dca16d3cea8984cfad..8815c67d8bcc0cd808c9fac4a61d3183a4a41856 100644 (file)
--- a/ggml/src/ggml.c
+++ b/ggml/src/ggml.c
@@ -5223,7 +5223,7 @@ static struct ggml_tensor * ggml_fill_impl(
     struct ggml_tensor  * a,
     float                 c,
     bool                  inplace) {
-    GGML_ASSERT(a->type == GGML_TYPE_F32);
+    GGML_ASSERT(a->type == GGML_TYPE_F32 || a->type == GGML_TYPE_F16);
     GGML_ASSERT(ggml_is_contiguous(a));
 
     struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a);