Commit 5a6caa05f for llama.cpp

commit 5a6caa05fc806cfd532a499cad8f340d92362010
Author: Georgi Gerganov <ggerganov@gmail.com>
Date:   Tue Sep 8 09:06:24 2026 +0300

    ggml : update ggml_prec specification (#26675)

    * ggml : update ggml_prec specification

    [no ci]

    * cont : add GGML_PREC_BF16

    * cont : rework API

    * cont : use new API

    * cont : swap arg order

    * cont : support for MUL_MAT_ID

    * cont : fix accidental remove of "break;"

    * cont : return bools, add doc TAG_GGML_PREC, clean-up

    * cont : add search tag

    * cont : ws

diff --git a/ggml/include/ggml.h b/ggml/include/ggml.h
index b88b7e54a..85a1ae7ae 100644
--- a/ggml/include/ggml.h
+++ b/ggml/include/ggml.h
@@ -433,10 +433,21 @@ extern "C" {
         GGML_TYPE_COUNT   = 43,
     };

-    // precision
+    // [TAG_GGML_PREC]
+    // this enum is used to declare the allowed numerical precision/data-types types that can be used during the compute of an op
+    // the declared types can be:
+    //  - result accumulation type
+    //  - source tensor data representation type
+    //  - etc.
+    // the precision parameters are stored as ggml_tensor.op_params to the respective ops
     enum ggml_prec {
-        GGML_PREC_DEFAULT =  0, // stored as ggml_tensor.op_params, 0 by default
-        GGML_PREC_F32     = 10,
+        GGML_PREC_UNDEFINED = 0,
+        GGML_PREC_DEFAULT   = 0,  // note: deprecated, use GGML_PREC_UNDEFINED
+        GGML_PREC_F32       = 10,
+        GGML_PREC_BF16      = 15,
+        GGML_PREC_F16       = 20,
+        GGML_PREC_Q8        = 30,
+        GGML_PREC_Q4        = 40,
     };

     // op hint
@@ -1429,6 +1440,42 @@ extern "C" {
             struct ggml_tensor  * b,
             float                 eps);

+    // [TAG_GGML_PREC]
+    // set the minimum required accumulator type for the implementation to use during the compute
+    // for example:
+    //  - GGML_PREC_F32  - requires accumulation of the results in F32
+    //  - GGML_PREC_BF16 - can accumulate the results in BF16, F32
+    //  - GGML_PREC_F16  - can accumulate the results in F16, F32
+    //  - GGML_PREC_Q8   - not allowed
+    //  - GGML_PREC_Q4   - not allowed
+    //
+    // return false on faliure
+    GGML_API bool ggml_prec_set_acc(
+            struct ggml_tensor * a,
+            enum ggml_prec       prec);
+
+    // [TAG_GGML_PREC]
+    // set the smallest rank that the implementation can use to internally convert the src[idx] data to
+    // ranks in decreasing order:
+    //  - GGML_PREC_F32  - GGML_TYPE_F32
+    //  - GGML_PREC_BF16 - GGML_TYPE_BF16
+    //  - GGML_PREC_F16  - GGML_TYPE_F16,
+    //  - GGML_PREC_Q8   - GGML_TYPE_Q8_0, GGML_TYPE_Q8_1, GGML_TYPE_Q8_K, etc.
+    //  - GGML_PREC_Q4   - GGML_TYPE_Q4_0, GGML_TYPE_Q4_1, GGML_TYPE_Q4_K, GGML_TYPE_NVFP4, GGML_TYPE_MXFP4, etc.
+    //
+    // for example:
+    //   - ggml_prec_set_src(a, GGML_PREC_Q8, 1):
+    //     - allows the implementation to quantize F32, BF16, F16 data of src[1] down to GGML_TYPE_Q8_0
+    //     - cannot quantize it down to GGML_TYPE_Q4_0 or GGML_TYPE_NVFP4
+    //   - ggml_prec_set_src(a, GGML_PREC_Q4, 1):
+    //     - allows the implementation to quantize F32, BF16, F16 data of src[1] down to 4-bit datatypes such as GGML_TYPE_Q4_K, GGML_TYPE_NVFP4 etc.
+    //
+    // return false on faliure
+    GGML_API bool ggml_prec_set_src(
+            struct ggml_tensor * a,
+            enum ggml_prec       prec,
+            int                  idx);
+
     // A: k columns, n rows => [ne03, ne02, n, k]
     // B: k columns, m rows  (i.e. we transpose it internally) => [ne03 * x, ne02 * y, m, k]
     // result is n columns, m rows => [ne03 * x, ne02 * y, m, n]
@@ -1439,9 +1486,10 @@ extern "C" {

     // change the precision of a matrix multiplication
     // set to GGML_PREC_F32 for higher precision (useful for phi-2)
-    GGML_API void ggml_mul_mat_set_prec(
+    GGML_DEPRECATED(GGML_API void ggml_mul_mat_set_prec(
             struct ggml_tensor * a,
-            enum ggml_prec       prec);
+            enum ggml_prec       prec),
+        "use ggml_prec_set_acc() instead");

     // change the hint of a matrix multiplication
     GGML_API void ggml_mul_mat_set_hint(
@@ -2446,9 +2494,10 @@ extern "C" {
             float                 max_bias,
             float                 logit_softcap);

-    GGML_API void ggml_flash_attn_ext_set_prec(
+    GGML_DEPRECATED(GGML_API void ggml_flash_attn_ext_set_prec(
             struct ggml_tensor * a,
-            enum ggml_prec       prec);
+            enum ggml_prec       prec),
+        "use ggml_prec_set_acc() instead");

     GGML_API enum ggml_prec ggml_flash_attn_ext_get_prec(
             const struct ggml_tensor * a);
diff --git a/ggml/src/ggml-impl.h b/ggml/src/ggml-impl.h
index 62b76abbc..ae26e0c23 100644
--- a/ggml/src/ggml-impl.h
+++ b/ggml/src/ggml-impl.h
@@ -160,6 +160,18 @@ static float ggml_get_op_params_f32(const struct ggml_tensor * tensor, uint32_t
     return ((const float *)(tensor->op_params))[i];
 }

+// [TAG_GGML_PREC]
+// - GGML_OP_MUL_MAT
+//   0 - acc
+//   1 - hint
+//   2 - src0 precision
+//   3 - src1 precision
+//
+// - GGML_OP_MUL_MAT_ID
+//   0 - acc
+//   1 - hint
+//   2 - src0 precision
+//   3 - src1 precision
 static void ggml_set_op_params_i32(struct ggml_tensor * tensor, uint32_t i, int32_t value) {
     assert(i < GGML_MAX_OP_PARAMS / sizeof(int32_t));
     ((int32_t *)(tensor->op_params))[i] = value;
diff --git a/ggml/src/ggml.c b/ggml/src/ggml.c
index 6257cdbe5..5ef03e190 100644
--- a/ggml/src/ggml.c
+++ b/ggml/src/ggml.c
@@ -3277,6 +3277,57 @@ struct ggml_tensor * ggml_l2_norm_inplace(
     return ggml_l2_norm_impl(ctx, a, eps, true);
 }

+// ggml_prec
+
+bool ggml_prec_set_acc(
+        struct ggml_tensor * a,
+        enum ggml_prec       prec) {
+    switch (a->op) {
+        case GGML_OP_MUL_MAT:
+        case GGML_OP_MUL_MAT_ID:
+            {
+                const int32_t prec_i32 = (int32_t) prec;
+                ggml_set_op_params_i32(a, 0, prec_i32);
+            }
+            break;
+        case GGML_OP_FLASH_ATTN_EXT:
+            {
+                const int32_t prec_i32 = (int32_t) prec;
+                ggml_set_op_params_i32(a, 3, prec_i32);
+            }
+            break;
+        default:
+            return false;
+    };
+
+    return true;
+}
+
+bool ggml_prec_set_src(
+        struct ggml_tensor * a,
+        enum ggml_prec       prec,
+        int                  idx) {
+    GGML_ASSERT(idx >= 0 && idx < GGML_MAX_SRC);
+
+    switch (a->op) {
+        case GGML_OP_MUL_MAT:
+        case GGML_OP_MUL_MAT_ID:
+            {
+                if (idx != 1) {
+                    return false;
+                }
+
+                const int32_t prec_i32 = (int32_t) prec;
+                ggml_set_op_params_i32(a, 2 + idx, prec_i32);
+            }
+            break;
+        default:
+            return false;
+    };
+
+    return true;
+}
+
 // ggml_mul_mat

 static inline bool ggml_can_mul_mat(const struct ggml_tensor * t0, const struct ggml_tensor * t1) {
diff --git a/src/llama-graph.cpp b/src/llama-graph.cpp
index 4cbd5fe21..5855393ef 100644
--- a/src/llama-graph.cpp
+++ b/src/llama-graph.cpp
@@ -1926,7 +1926,7 @@ ggml_tensor * llm_graph_context::build_ffn(
         cur = build_lora_mm(down, cur);
         if (arch == LLM_ARCH_GLM4 || arch == LLM_ARCH_GLM4_MOE || arch == LLM_ARCH_JAIS2) {
             // GLM4, GLM4_MOE, and JAIS2 seem to have numerical issues with half-precision accumulators
-            ggml_mul_mat_set_prec(cur, GGML_PREC_F32);
+            ggml_prec_set_acc(cur, GGML_PREC_F32);
         }
     }

@@ -2024,7 +2024,7 @@ ggml_tensor * llm_graph_context::build_moe_ffn(
     if (probs_in == nullptr) {
         logits = build_lora_mm(gate_inp, cur); // [n_expert, n_tokens]
         if (gating_op == LLAMA_EXPERT_GATING_FUNC_TYPE_SQRT_SOFTPLUS) {
-            ggml_mul_mat_set_prec(logits, GGML_PREC_F32);
+            ggml_prec_set_acc(logits, GGML_PREC_F32);
         }
         cb(logits, "ffn_moe_logits", il);
     } else {
@@ -2636,7 +2636,7 @@ ggml_tensor * llm_graph_context::build_attn_mha(
         ggml_flash_attn_ext_add_sinks(cur, sinks);
         GGML_ASSERT(n_kv_max >= 0 && n_kv_max <= INT32_MAX);
         ggml_flash_attn_ext_set_n_kv_max(cur, static_cast<int32_t>(n_kv_max));
-        ggml_flash_attn_ext_set_prec (cur, GGML_PREC_F32);
+        ggml_prec_set_acc(cur, GGML_PREC_F32);

         if (v_mla) {
 #if 0
@@ -2662,7 +2662,7 @@ ggml_tensor * llm_graph_context::build_attn_mha(

         // note: this op tends to require high floating point range
         //       while for some models F16 is enough, for others it is not, so we default to F32 here
-        ggml_mul_mat_set_prec(kq, GGML_PREC_F32);
+        ggml_prec_set_acc(kq, GGML_PREC_F32);

         if (arch == LLM_ARCH_GROK) {
             // need to do the following:
@@ -2895,7 +2895,7 @@ ggml_tensor * llm_graph_context::build_attn(
         if (arch == LLM_ARCH_GLM4 || arch == LLM_ARCH_GLM4_MOE || arch == LLM_ARCH_JAIS2) {
             // GLM4, GLM4_MOE, and JAIS2 seem to have numerical issues with half-precision accumulators
             cur = build_lora_mm(wo, cur);
-            ggml_mul_mat_set_prec(cur, GGML_PREC_F32);
+            ggml_prec_set_acc(cur, GGML_PREC_F32);
             if (wo_s) {
                 cur = ggml_mul(ctx0, cur, wo_s);
             }
@@ -2982,7 +2982,7 @@ ggml_tensor * llm_graph_context::build_attn(
         if (arch == LLM_ARCH_GLM4 || arch == LLM_ARCH_GLM4_MOE) {
             // GLM4 and GLM4_MOE seem to have numerical issues with half-precision accumulators
             cur = build_lora_mm(wo, cur);
-            ggml_mul_mat_set_prec(cur, GGML_PREC_F32);
+            ggml_prec_set_acc(cur, GGML_PREC_F32);
             if (wo_s) {
                 cur = ggml_mul(ctx0, cur, wo_s);
             }
diff --git a/src/models/minimax-m3.cpp b/src/models/minimax-m3.cpp
index 80260a629..f3b64b210 100644
--- a/src/models/minimax-m3.cpp
+++ b/src/models/minimax-m3.cpp
@@ -191,7 +191,7 @@ ggml_tensor * llama_model_minimax_m3::graph::build_attn_msa_fa(

     ggml_tensor * o = ggml_flash_attn_ext(ctx0, q, k, v, mask, kq_scale,
                                           hparams.f_max_alibi_bias, 0.0f);
-    ggml_flash_attn_ext_set_prec(o, GGML_PREC_F32);
+    ggml_prec_set_acc(o, GGML_PREC_F32);
     cb(o, "msa_fattn", il);

     // [D, Gp, R, C] -> [D, Gp, C, R] -> [n_embd, T]
@@ -389,7 +389,7 @@ llama_model_minimax_m3::graph::graph(const llama_model & model, const llm_graph_
                     ggml_tensor * iq4 = ggml_reshape_4d(ctx0, iq, n_idx_dim, Hd, 1, ns);
                     ggml_tensor * sc  = ggml_mul_mat(ctx0,
                             ggml_reshape_4d(ctx0, ikp, n_idx_dim, n_ps, 1, ns), iq4);
-                    ggml_mul_mat_set_prec(sc, GGML_PREC_F32);
+                    ggml_prec_set_acc(sc, GGML_PREC_F32);
                     // unmapped positions come out -inf, so they can never rank into the top-k
                     sc = ggml_add_inplace(ctx0, sc,
                             ggml_reshape_4d(ctx0, msa->pos_mask, n_ps, 1, 1, ns));
@@ -471,7 +471,7 @@ llama_model_minimax_m3::graph::graph(const llama_model & model, const llm_graph_
                         ggml_tensor * sc = ggml_mul_mat(ctx0, ikp,
                                 ggml_reshape_2d(ctx0, iq_s, n_idx_dim, Hd*n_tps));
                         // indexer scores run in F32
-                        ggml_mul_mat_set_prec(sc, GGML_PREC_F32);
+                        ggml_prec_set_acc(sc, GGML_PREC_F32);
                         sc = ggml_reshape_3d(ctx0, sc, n_ps, Hd, n_tps);
                         // unmapped positions (holes, padding, empty cells) come out -inf
                         sc = ggml_add_inplace(ctx0, sc, pm_s);
diff --git a/tests/test-backend-ops.cpp b/tests/test-backend-ops.cpp
index c335e793b..5121578ff 100644
--- a/tests/test-backend-ops.cpp
+++ b/tests/test-backend-ops.cpp
@@ -7659,7 +7659,7 @@ struct test_flash_attn_ext : public test_case {
         ggml_tensor * out = ggml_flash_attn_ext(ctx, q, k, v, m, 1.0f/sqrtf(hsk), max_bias, logit_softcap);
         ggml_flash_attn_ext_add_sinks(out, s);
         ggml_flash_attn_ext_set_n_kv_max(out, n_kv_max);
-        ggml_flash_attn_ext_set_prec (out, prec);
+        ggml_prec_set_acc(out, prec);
         ggml_set_name(out, "out");

         return out;
diff --git a/tools/mtmd/clip.cpp b/tools/mtmd/clip.cpp
index 74f4e2b5a..cd6421def 100644
--- a/tools/mtmd/clip.cpp
+++ b/tools/mtmd/clip.cpp
@@ -780,7 +780,7 @@ ggml_tensor * clip_graph::build_attn(
         }

         cur = ggml_flash_attn_ext(ctx0, q, k, v, kq_mask, kq_scale, 0.0f, 0.0f);
-        ggml_flash_attn_ext_set_prec(cur, GGML_PREC_F32);
+        ggml_prec_set_acc(cur, GGML_PREC_F32);
         if (sinks != nullptr) {
             ggml_flash_attn_ext_add_sinks(cur, sinks);
         }
@@ -793,7 +793,7 @@ ggml_tensor * clip_graph::build_attn(

         ggml_tensor * kq = ggml_mul_mat(ctx0, k, q);
         // F32 may not needed for vision encoders?
-        // ggml_mul_mat_set_prec(kq, GGML_PREC_F32);
+        // ggml_prec_set_acc(kq, GGML_PREC_F32);

         kq = ggml_soft_max_ext(ctx0, kq, kq_mask, kq_scale, 0.0f);
         if (sinks != nullptr) {
diff --git a/tools/mtmd/models/mimovl.cpp b/tools/mtmd/models/mimovl.cpp
index 6ff1124a0..e1fbe2671 100644
--- a/tools/mtmd/models/mimovl.cpp
+++ b/tools/mtmd/models/mimovl.cpp
@@ -2,7 +2,7 @@

 ggml_tensor * clip_graph_mimovl::build_mm(ggml_tensor * w, ggml_tensor * x) const {
     ggml_tensor * cur = ggml_mul_mat(ctx0, w, x);
-    ggml_mul_mat_set_prec(cur, GGML_PREC_F32);
+    ggml_prec_set_acc(cur, GGML_PREC_F32);
     return cur;
 }

diff --git a/tools/mtmd/models/qwen3tts-spkenc.cpp b/tools/mtmd/models/qwen3tts-spkenc.cpp
index d4659fd63..405fbb9cb 100644
--- a/tools/mtmd/models/qwen3tts-spkenc.cpp
+++ b/tools/mtmd/models/qwen3tts-spkenc.cpp
@@ -27,7 +27,7 @@ ggml_tensor * clip_graph_qwen3tts_spkenc::conv1d_same(ggml_tensor * x, ggml_tens

     ggml_tensor * w2d = ggml_reshape_2d(ctx0, w, (int64_t) K * IC, OC);
     ggml_tensor * y   = ggml_mul_mat(ctx0, w2d, col); // [OC, T_out]
-    ggml_mul_mat_set_prec(y, GGML_PREC_F32);
+    ggml_prec_set_acc(y, GGML_PREC_F32);

     ggml_tensor * b2d = ggml_reshape_2d(ctx0, b, OC, 1);
     y = ggml_add(ctx0, y, b2d);
diff --git a/tools/tuning/fa-vec.cpp b/tools/tuning/fa-vec.cpp
index f90437969..3d6cbeb2c 100644
--- a/tools/tuning/fa-vec.cpp
+++ b/tools/tuning/fa-vec.cpp
@@ -56,7 +56,7 @@ static ggml_tensor * fa_build_graph(ggml_context * ctx, const fa_shape & s) {
     ggml_set_name(m, "m");

     ggml_tensor * out = ggml_flash_attn_ext(ctx, q, k, v, m, 1.0f / sqrtf((float) s.dk), 0.0f, 0.0f);
-    ggml_flash_attn_ext_set_prec(out, GGML_PREC_F32);
+    ggml_prec_set_acc(out, GGML_PREC_F32);
     ggml_set_name(out, "out");

     return out;