Add Q2_0 quantization: type definition and CPU backend (#24448 )

spec : fix naming, spacing (#25410 )
CUDA: Fuse MMVQ post-scale for NVFP4 (#24481 )
2026-07-09 14:05:54 +02:00 · 2026-07-07 12:05:47 -07:00 · 2026-07-07 18:52:30 +03:00 · 2026-07-07 17:12:19 +02:00 · 2026-07-07 17:20:42 +03:00 · 2026-07-07 16:13:03 +02:00
27 changed files with 1120 additions and 213 deletions
@@ -2221,6 +2221,112 @@ int32_t common_speculative_n_max(const common_params_speculative * spec) {
    return n_max;
 }

+common_params common_base_params_to_speculative(const common_params & params) {
+    const bool has_draft = params.speculative.has_dft();
+
+    const auto & params_spec = params.speculative.draft;
+    common_params result = params;
+
+    if (has_draft) {
+        result.devices               = params_spec.devices;
+        result.model                 = params_spec.mparams;
+        result.n_gpu_layers          = params_spec.n_gpu_layers;
+        result.tensor_buft_overrides = params_spec.tensor_buft_overrides;
+
+        if (params_spec.cpuparams.n_threads > 0) {
+            result.cpuparams.n_threads       = params_spec.cpuparams.n_threads;
+            result.cpuparams_batch.n_threads = params_spec.cpuparams_batch.n_threads;
+        }
+    }
+
+    result.cache_type_k  = params_spec.cache_type_k;
+    result.cache_type_v  = params_spec.cache_type_v;
+    result.n_outputs_max = params.n_parallel;
+
+    return result;
+}
+
+struct common_speculative_init_result::impl {
+    impl() = default;
+    ~impl() = default;
+
+    // note: the order in which model, context, etc. are declared matters because their destructors will be called bottom-to-top
+    llama_model_ptr   model;
+    llama_context_ptr context;
+};
+
+common_speculative_init_result::common_speculative_init_result(
+    common_params & params,
+      llama_model * model_tgt,
+    llama_context * ctx_tgt) :
+    pimpl(new impl{}) {
+    const bool has_draft = params.speculative.has_dft();
+    const bool spec_mtp = std::find(params.speculative.types.begin(),
+                                    params.speculative.types.end(),
+                                    COMMON_SPECULATIVE_TYPE_DRAFT_MTP) != params.speculative.types.end();
+    GGML_ASSERT(has_draft || spec_mtp);
+
+    auto mparams = common_model_params_to_llama(params);
+    auto cparams = common_context_params_to_llama(params);
+
+    if (spec_mtp) {
+        cparams.ctx_type = LLAMA_CONTEXT_TYPE_MTP;
+    }
+
+    // note: for small models maybe we can set this to the maximum possible draft from all speculative types
+    //       the extra memory for small models is likely negligible?
+    cparams.n_rs_seq  = 0;
+    cparams.ctx_other = ctx_tgt;
+
+    std::string model_path;
+    if (has_draft) {
+        model_path = params.speculative.draft.mparams.path;
+        LOG_TRC("%s: loading draft model '%s'\n", __func__, model_path.c_str());
+
+        llama_model * model_dft = llama_model_load_from_file(params.model.path.c_str(), mparams);
+        if (model_dft == NULL) {
+            LOG_ERR("%s: failed to load draft model, '%s'\n", __func__, model_path.c_str());
+            return;
+        }
+
+        pimpl->model.reset(model_dft);
+
+        llama_context * ctx_dft = llama_init_from_model(model_dft, cparams);
+        if (ctx_dft == nullptr) {
+            LOG_ERR("%s: failed to create MTP context\n", __func__);
+            return;
+        }
+
+        pimpl->context.reset(ctx_dft);
+    } else if (spec_mtp) {
+        model_path = params.model.path;
+
+        LOG_TRC("%s: creating MTP draft context against the target model '%s'\n", __func__, model_path.c_str());
+
+        llama_context * ctx_dft = llama_init_from_model(model_tgt, cparams);
+        if (ctx_dft == nullptr) {
+            LOG_ERR("%s: failed to create MTP context\n", __func__);
+            return;
+        }
+
+        pimpl->context.reset(ctx_dft);
+    }
+}
+
+common_speculative_init_result::~common_speculative_init_result() = default;
+
+llama_model * common_speculative_init_result::model() {
+    return pimpl->model.get();
+}
+
+llama_context * common_speculative_init_result::context() {
+    return pimpl->context.get();
+}
+
+common_speculative_init_result_ptr common_speculative_init_from_params(common_params & params, llama_model * model_tgt, llama_context * ctx_tgt) {
+    return std::make_unique<common_speculative_init_result>(params, model_tgt, ctx_tgt);
+}
+
 // initialization of the speculative decoding system
 //
 common_speculative * common_speculative_init(common_params_speculative & params, uint32_t n_seq) {
@@ -23,6 +23,8 @@ std::string common_speculative_type_to_str(enum common_speculative_type type);
 // return the max number of draft tokens based on the speculative parameters
 int32_t common_speculative_n_max(const common_params_speculative * spec);

+common_params common_base_params_to_speculative(const common_params & params);
+
 common_speculative * common_speculative_init(common_params_speculative & params, uint32_t n_seq);

 void common_speculative_free(common_speculative * spec);
@@ -80,3 +82,19 @@ struct common_speculative_deleter {
 };

 typedef std::unique_ptr<common_speculative, common_speculative_deleter> common_speculative_ptr;
+
+struct common_speculative_init_result {
+    common_speculative_init_result(common_params & params, llama_model * model_tgt, llama_context * ctx_tgt);
+    ~common_speculative_init_result();
+
+    llama_model   * model();
+    llama_context * context();
+
+private:
+    struct impl;
+    std::unique_ptr<impl> pimpl;
+};
+
+using common_speculative_init_result_ptr = std::unique_ptr<common_speculative_init_result>;
+
+common_speculative_init_result_ptr common_speculative_init_from_params(common_params & params, llama_model * model_tgt, llama_context * ctx_tgt);
@@ -429,7 +429,8 @@ extern "C" {
        GGML_TYPE_MXFP4   = 39, // MXFP4 (1 block)
        GGML_TYPE_NVFP4   = 40, // NVFP4 (4 blocks, E4M3 scale)
        GGML_TYPE_Q1_0    = 41,
-        GGML_TYPE_COUNT   = 42,
+        GGML_TYPE_Q2_0    = 42,
+        GGML_TYPE_COUNT   = 43,
    };

    // precision
@@ -473,6 +474,7 @@ extern "C" {
        GGML_FTYPE_MOSTLY_MXFP4   = 25, // except 1d tensors
        GGML_FTYPE_MOSTLY_NVFP4   = 26, // except 1d tensors
        GGML_FTYPE_MOSTLY_Q1_0    = 27, // except 1d tensors
+        GGML_FTYPE_MOSTLY_Q2_0    = 28, // except 1d tensors
    };

    // available tensor operations:
@@ -96,6 +96,9 @@ typedef sycl::half2 ggml_half2;
 #define QI1_0 (QK1_0 / 32)
 #define QR1_0 1

+#define QI2_0 (QK2_0 / 32)
+#define QR2_0 1
+

 #define QI4_0 (QK4_0 / (4 * QR4_0))
 #define QR4_0 2
@@ -181,6 +184,13 @@ typedef struct {
 } block_q1_0;
 static_assert(sizeof(block_q1_0) == sizeof(ggml_half) + QK1_0 / 8, "wrong q1_0 block size/padding");

+#define QK2_0 64
+typedef struct {
+    ggml_half d;              // delta (scale)
+    uint8_t qs[QK2_0 / 4];   // 2 bits per element
+} block_q2_0;
+static_assert(sizeof(block_q2_0) == sizeof(ggml_half) + QK2_0 / 4, "wrong q2_0 block size/padding");
+
 #define QK4_0 32
 typedef struct {
    ggml_half d;           // delta
@@ -17,6 +17,7 @@
 #define ggml_vec_dot_mxfp4_q8_0_generic ggml_vec_dot_mxfp4_q8_0
 #define ggml_vec_dot_nvfp4_q8_0_generic ggml_vec_dot_nvfp4_q8_0
 #define ggml_vec_dot_q1_0_q8_0_generic ggml_vec_dot_q1_0_q8_0
+#define ggml_vec_dot_q2_0_q8_0_generic ggml_vec_dot_q2_0_q8_0
 #define ggml_vec_dot_tq1_0_q8_K_generic ggml_vec_dot_tq1_0_q8_K
 #define ggml_vec_dot_tq2_0_q8_K_generic ggml_vec_dot_tq2_0_q8_K
 #define ggml_vec_dot_q2_K_q8_K_generic ggml_vec_dot_q2_K_q8_K
@@ -82,6 +83,7 @@
 #define ggml_gemm_q2_K_8x8_q8_K_generic ggml_gemm_q2_K_8x8_q8_K
 #elif defined(__x86_64__) || defined(__i386__) || defined(_M_IX86) || defined(_M_X64)
 // quants.c
+#define ggml_vec_dot_q2_0_q8_0_generic ggml_vec_dot_q2_0_q8_0
 // repack.cpp
 #define ggml_quantize_mat_q8_0_4x4_generic ggml_quantize_mat_q8_0_4x4
 #define ggml_quantize_mat_q8_K_4x4_generic ggml_quantize_mat_q8_K_4x4
@@ -113,6 +115,7 @@
 #define quantize_row_q8_K_generic quantize_row_q8_K
 #define ggml_vec_dot_nvfp4_q8_0_generic ggml_vec_dot_nvfp4_q8_0
 #define ggml_vec_dot_q1_0_q8_0_generic ggml_vec_dot_q1_0_q8_0
+#define ggml_vec_dot_q2_0_q8_0_generic ggml_vec_dot_q2_0_q8_0
 #define ggml_vec_dot_tq1_0_q8_K_generic ggml_vec_dot_tq1_0_q8_K
 #define ggml_vec_dot_tq2_0_q8_K_generic ggml_vec_dot_tq2_0_q8_K
 #define ggml_vec_dot_iq1_m_q8_K_generic ggml_vec_dot_iq1_m_q8_K
@@ -162,6 +165,7 @@
 #define ggml_vec_dot_mxfp4_q8_0_generic ggml_vec_dot_mxfp4_q8_0
 #define ggml_vec_dot_nvfp4_q8_0_generic ggml_vec_dot_nvfp4_q8_0
 #define ggml_vec_dot_q1_0_q8_0_generic ggml_vec_dot_q1_0_q8_0
+#define ggml_vec_dot_q2_0_q8_0_generic ggml_vec_dot_q2_0_q8_0
 // repack.cpp
 #define ggml_quantize_mat_q8_0_4x4_generic ggml_quantize_mat_q8_0_4x4
 #define ggml_quantize_mat_q8_0_4x8_generic ggml_quantize_mat_q8_0_4x8
@@ -202,6 +206,7 @@
 #elif defined(__riscv)
 // quants.c
 #define ggml_vec_dot_nvfp4_q8_0_generic ggml_vec_dot_nvfp4_q8_0
+#define ggml_vec_dot_q2_0_q8_0_generic ggml_vec_dot_q2_0_q8_0
 // repack.cpp
 #define ggml_quantize_mat_q8_0_4x1_generic ggml_quantize_mat_q8_0_4x1
 #define ggml_quantize_mat_q8_0_4x4_generic ggml_quantize_mat_q8_0_4x4
@@ -243,6 +248,7 @@
 #define quantize_row_q8_K_generic quantize_row_q8_K
 #define ggml_vec_dot_nvfp4_q8_0_generic ggml_vec_dot_nvfp4_q8_0
 #define ggml_vec_dot_q1_0_q8_0_generic ggml_vec_dot_q1_0_q8_0
+#define ggml_vec_dot_q2_0_q8_0_generic ggml_vec_dot_q2_0_q8_0
 #define ggml_vec_dot_tq1_0_q8_K_generic ggml_vec_dot_tq1_0_q8_K
 #define ggml_vec_dot_tq2_0_q8_K_generic ggml_vec_dot_tq2_0_q8_K
 #define ggml_vec_dot_q2_K_q8_K_generic ggml_vec_dot_q2_K_q8_K
@@ -306,6 +312,7 @@
 #define ggml_vec_dot_mxfp4_q8_0_generic ggml_vec_dot_mxfp4_q8_0
 #define ggml_vec_dot_nvfp4_q8_0_generic ggml_vec_dot_nvfp4_q8_0
 #define ggml_vec_dot_q1_0_q8_0_generic ggml_vec_dot_q1_0_q8_0
+#define ggml_vec_dot_q2_0_q8_0_generic ggml_vec_dot_q2_0_q8_0
 // repack.cpp
 #define ggml_quantize_mat_q8_0_4x4_generic ggml_quantize_mat_q8_0_4x4
 #define ggml_quantize_mat_q8_0_4x8_generic ggml_quantize_mat_q8_0_4x8
@@ -219,6 +219,80 @@ void ggml_vec_dot_q1_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
 #endif
 }

+void ggml_vec_dot_q2_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
+    const int qk = QK2_0;
+    const int nb = n / qk;
+
+    assert(n % qk == 0);
+    assert(nrc == 1);
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const block_q2_0 * GGML_RESTRICT x = vx;
+    const block_q8_0 * GGML_RESTRICT y = vy;
+
+    float sumf = 0.0f;
+
+#if defined(__ARM_NEON)
+    // Replicate pattern: each byte repeated 4 times
+    static const uint8_t tbl_idx_lo[16] = {0,0,0,0, 1,1,1,1, 2,2,2,2, 3,3,3,3};
+    static const uint8_t tbl_idx_hi[16] = {4,4,4,4, 5,5,5,5, 6,6,6,6, 7,7,7,7};
+    // Right-shift amounts: 0,2,4,6 repeated for each group of 4
+    static const int8_t shift_vals[16] = {0,-2,-4,-6, 0,-2,-4,-6, 0,-2,-4,-6, 0,-2,-4,-6};
+
+    const uint8x16_t idx_lo  = vld1q_u8(tbl_idx_lo);
+    const uint8x16_t idx_hi  = vld1q_u8(tbl_idx_hi);
+    const int8x16_t  shifts  = vld1q_s8(shift_vals);
+    const uint8x16_t mask2   = vdupq_n_u8(0x03);
+    const int8x16_t  one     = vdupq_n_s8(1);
+
+    float32x4_t sumv = vdupq_n_f32(0.0f);
+
+    for (int i = 0; i < nb; i++) {
+        const float d0 = GGML_CPU_FP16_TO_FP32(x[i].d);
+
+        // group 64: one Q2_0 block (64 weights) maps to two Q8_0 blocks (2 * 32 = 64)
+        for (int k = 0; k < 2; k++) {
+            const block_q8_0 * GGML_RESTRICT yb = &y[i * 2 + k];
+            const float d1 = GGML_CPU_FP16_TO_FP32(yb->d);
+
+            // Load 8 bytes of packed 2-bit values
+            const uint8x8_t raw = vld1_u8(&x[i].qs[k * 8]);
+            const uint8x16_t raw16 = vcombine_u8(raw, raw);
+
+            // First 16 elements: replicate bytes 0-3, shift, mask, subtract 1
+            uint8x16_t bytes0 = vqtbl1q_u8(raw16, idx_lo);
+            int8x16_t qv0 = vsubq_s8(
+                vreinterpretq_s8_u8(vandq_u8(vshlq_u8(bytes0, shifts), mask2)),
+                one);
+
+            // Second 16 elements: replicate bytes 4-7, shift, mask, subtract 1
+            uint8x16_t bytes1 = vqtbl1q_u8(raw16, idx_hi);
+            int8x16_t qv1 = vsubq_s8(
+                vreinterpretq_s8_u8(vandq_u8(vshlq_u8(bytes1, shifts), mask2)),
+                one);
+
+            // Load Q8_0 values and dot product
+            const int8x16_t y0 = vld1q_s8(yb->qs);
+            const int8x16_t y1 = vld1q_s8(yb->qs + 16);
+
+            int32x4_t p0 = ggml_vdotq_s32(vdupq_n_s32(0), qv0, y0);
+            int32x4_t p1 = ggml_vdotq_s32(p0, qv1, y1);
+
+            sumv = vmlaq_n_f32(sumv, vcvtq_f32_s32(p1), d0 * d1);
+        }
+    }
+
+    sumf = vaddvq_f32(sumv);
+#else
+    ggml_vec_dot_q2_0_q8_0_generic(n, s, bs, vx, bx, vy, by, nrc);
+    return;
+#endif
+
+    *s = sumf;
+}

 void ggml_vec_dot_q4_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
    const int qk = QK8_0;
@@ -230,6 +230,12 @@ static const struct ggml_type_traits_cpu type_traits_cpu[GGML_TYPE_COUNT] = {
        .vec_dot_type             = GGML_TYPE_Q8_0,
        .nrows                    = 1,
    },
+    [GGML_TYPE_Q2_0] = {
+        .from_float               = quantize_row_q2_0,
+        .vec_dot                  = ggml_vec_dot_q2_0_q8_0,
+        .vec_dot_type             = GGML_TYPE_Q8_0,
+        .nrows                    = 1,
+    },
    [GGML_TYPE_Q4_0] = {
        .from_float               = quantize_row_q4_0,
        .vec_dot                  = ggml_vec_dot_q4_0_q8_0,
@@ -665,6 +665,7 @@ void ggml_compute_forward_add(
                ggml_compute_forward_add_non_quantized(params, dst);
            } break;
        case GGML_TYPE_Q1_0:
+        case GGML_TYPE_Q2_0:
        case GGML_TYPE_Q4_0:
        case GGML_TYPE_Q4_1:
        case GGML_TYPE_Q5_0:
@@ -1115,6 +1116,7 @@ void ggml_compute_forward_add1(
                }
            } break;
        case GGML_TYPE_Q1_0:
+        case GGML_TYPE_Q2_0:
        case GGML_TYPE_Q4_0:
        case GGML_TYPE_Q4_1:
        case GGML_TYPE_Q5_0:
@@ -1245,6 +1247,7 @@ void ggml_compute_forward_acc(
        case GGML_TYPE_F16:
        case GGML_TYPE_BF16:
        case GGML_TYPE_Q1_0:
+        case GGML_TYPE_Q2_0:
        case GGML_TYPE_Q4_0:
        case GGML_TYPE_Q4_1:
        case GGML_TYPE_Q5_0:
@@ -4454,6 +4457,7 @@ void ggml_compute_forward_out_prod(

    switch (src0->type) {
        case GGML_TYPE_Q1_0:
+        case GGML_TYPE_Q2_0:
        case GGML_TYPE_Q4_0:
        case GGML_TYPE_Q4_1:
        case GGML_TYPE_Q5_0:
@@ -4730,6 +4734,7 @@ void ggml_compute_forward_set(
        case GGML_TYPE_F16:
        case GGML_TYPE_BF16:
        case GGML_TYPE_Q1_0:
+        case GGML_TYPE_Q2_0:
        case GGML_TYPE_Q4_0:
        case GGML_TYPE_Q4_1:
        case GGML_TYPE_Q5_0:
@@ -4954,6 +4959,7 @@ void ggml_compute_forward_get_rows(

    switch (src0->type) {
        case GGML_TYPE_Q1_0:
+        case GGML_TYPE_Q2_0:
        case GGML_TYPE_Q4_0:
        case GGML_TYPE_Q4_1:
        case GGML_TYPE_Q5_0:
@@ -5680,6 +5686,7 @@ void ggml_compute_forward_clamp(
            } break;
        case GGML_TYPE_BF16:
        case GGML_TYPE_Q1_0:
+        case GGML_TYPE_Q2_0:
        case GGML_TYPE_Q4_0:
        case GGML_TYPE_Q4_1:
        case GGML_TYPE_Q5_0:
@@ -26,6 +26,10 @@ void quantize_row_q1_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, in
    quantize_row_q1_0_ref(x, y, k);
 }

+void quantize_row_q2_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k) {
+    quantize_row_q2_0_ref(x, y, k);
+}
+
 void quantize_row_q4_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k) {
    quantize_row_q4_0_ref(x, y, k);
 }
@@ -170,6 +174,53 @@ void ggml_vec_dot_q1_0_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, c
    *s = sumf;
 }

+void ggml_vec_dot_q2_0_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
+    const int qk = QK2_0;
+    const int nb = n / qk;
+
+    assert(n % qk == 0);
+    assert(nrc == 1);
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const block_q2_0 * GGML_RESTRICT x = vx;
+    const block_q8_0 * GGML_RESTRICT y = vy;
+
+    float sumf = 0.0f;
+
+    for (int i = 0; i < nb; i++) {
+        const float d0 = GGML_CPU_FP16_TO_FP32(x[i].d);
+
+        float sumi = 0.0f;
+
+        // group 64: one Q2_0 block (64 weights) maps to two Q8_0 blocks (2 * 32 = 64)
+        for (int k = 0; k < 2; k++) {
+            const block_q8_0 * GGML_RESTRICT yb = &y[i * 2 + k];
+            const float d1 = GGML_CPU_FP16_TO_FP32(yb->d);
+            int sumi_block = 0;
+
+            const uint8_t * GGML_RESTRICT qs = &x[i].qs[k * 8];
+            const int8_t  * GGML_RESTRICT qy = yb->qs;
+
+            for (int b = 0; b < 8; ++b) {
+                const uint8_t byte = qs[b];
+                // Extract 4 two-bit values, map {0,1,2,3} -> {-1,0,1,2}
+                sumi_block += ((int)((byte >> 0) & 3) - 1) * qy[b*4 + 0];
+                sumi_block += ((int)((byte >> 2) & 3) - 1) * qy[b*4 + 1];
+                sumi_block += ((int)((byte >> 4) & 3) - 1) * qy[b*4 + 2];
+                sumi_block += ((int)((byte >> 6) & 3) - 1) * qy[b*4 + 3];
+            }
+
+            sumi += d1 * sumi_block;
+        }
+
+        sumf += d0 * sumi;
+    }
+
+    *s = sumf;
+}

 void ggml_vec_dot_q4_0_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
    const int qk = QK8_0;
@@ -13,6 +13,7 @@ extern "C" {

 // Quantization
 void quantize_row_q1_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+void quantize_row_q2_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
 void quantize_row_q4_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
 void quantize_row_q4_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
 void quantize_row_q5_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
@@ -38,6 +39,7 @@ void quantize_row_iq4_xs (const float * GGML_RESTRICT x, void * GGML_RESTRICT y,

 // Dot product
 void ggml_vec_dot_q1_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_q2_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
 void ggml_vec_dot_q4_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
 void ggml_vec_dot_q4_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
 void ggml_vec_dot_q5_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
@@ -71,6 +73,7 @@ void quantize_row_q8_0_generic(const float * GGML_RESTRICT x, void * GGML_RESTRI
 void quantize_row_q8_1_generic(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k);
 void quantize_row_q8_K_generic(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
 void ggml_vec_dot_q1_0_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_q2_0_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
 void ggml_vec_dot_q4_0_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
 void ggml_vec_dot_q4_1_q8_1_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
 void ggml_vec_dot_q5_0_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
@@ -1505,12 +1505,16 @@ struct ggml_cuda_mm_fusion_args_host {
    const ggml_tensor * x_bias = nullptr;
    const ggml_tensor * gate = nullptr;
    const ggml_tensor * gate_bias = nullptr;
+    const ggml_tensor * x_scale = nullptr;
+    const ggml_tensor * gate_scale = nullptr;
    ggml_glu_op glu_op;
 };
 struct ggml_cuda_mm_fusion_args_device {
    const void * x_bias = nullptr;
    const void * gate = nullptr;
    const void * gate_bias = nullptr;
+    const void * x_scale = nullptr;
+    const void * gate_scale = nullptr;
    ggml_glu_op glu_op;
 };

@@ -1582,12 +1582,18 @@ static bool ggml_cuda_should_fuse_mul_mat(const ggml_tensor * ffn_up,
                                          const ggml_tensor * ffn_gate,
                                          const ggml_tensor * glu,
                                          const ggml_tensor * ffn_up_bias = nullptr,
-                                          const ggml_tensor * ffn_gate_bias = nullptr) {
+                                          const ggml_tensor * ffn_gate_bias = nullptr,
+                                          const ggml_tensor * ffn_up_scale = nullptr,
+                                          const ggml_tensor * ffn_gate_scale = nullptr) {
    const bool has_bias = ffn_up_bias != nullptr || ffn_gate_bias != nullptr;
+    const bool has_scale = ffn_up_scale != nullptr || ffn_gate_scale != nullptr;

    if (has_bias && (!ffn_up_bias || !ffn_gate_bias)) {
        return false;
    }
+    if (has_scale && (!ffn_up_scale || !ffn_gate_scale)) {
+        return false;
+    }

    const bool is_mul_mat     = ffn_up->op == GGML_OP_MUL_MAT     && ffn_gate->op == GGML_OP_MUL_MAT     && glu->op == GGML_OP_GLU;
    const bool is_mul_mat_id  = ffn_up->op == GGML_OP_MUL_MAT_ID  && ffn_gate->op == GGML_OP_MUL_MAT_ID  && glu->op == GGML_OP_GLU;
@@ -1599,34 +1605,45 @@ static bool ggml_cuda_should_fuse_mul_mat(const ggml_tensor * ffn_up,
    }

    const ggml_op expected_bias_op = is_mul_mat ? GGML_OP_ADD : GGML_OP_ADD_ID;
+    const ggml_tensor * ffn_up_bias_src   = has_scale ? ffn_up_scale   : ffn_up;
+    const ggml_tensor * ffn_gate_bias_src = has_scale ? ffn_gate_scale : ffn_gate;
+    const ggml_tensor * ffn_up_out        = has_bias ? ffn_up_bias     : ffn_up_bias_src;
+    const ggml_tensor * ffn_gate_out      = has_bias ? ffn_gate_bias   : ffn_gate_bias_src;
+
+    if (glu->src[0] != ffn_gate_out || glu->src[1] != ffn_up_out) {
+        return false;
+    }
+
+    if (has_scale) {
+        if (ffn_up_scale->op != GGML_OP_MUL || ffn_gate_scale->op != GGML_OP_MUL) {
+            return false;
+        }
+        const bool up_has_mm   = ffn_up_scale->src[0] == ffn_up || ffn_up_scale->src[1] == ffn_up;
+        const bool gate_has_mm = ffn_gate_scale->src[0] == ffn_gate || ffn_gate_scale->src[1] == ffn_gate;
+        if (!up_has_mm || !gate_has_mm) {
+            return false;
+        }
+    }

    if (has_bias) {
        if (ffn_up_bias->op != expected_bias_op || ffn_gate_bias->op != expected_bias_op) {
            return false;
        }

-        if (glu->src[0] != ffn_gate_bias || glu->src[1] != ffn_up_bias) {
-            return false;
-        }
-
        if (expected_bias_op == GGML_OP_ADD) {
-            const bool up_has_mul   = ffn_up_bias->src[0] == ffn_up || ffn_up_bias->src[1] == ffn_up;
-            const bool gate_has_mul = ffn_gate_bias->src[0] == ffn_gate || ffn_gate_bias->src[1] == ffn_gate;
+            const bool up_has_mul   = ffn_up_bias->src[0] == ffn_up_bias_src || ffn_up_bias->src[1] == ffn_up_bias_src;
+            const bool gate_has_mul = ffn_gate_bias->src[0] == ffn_gate_bias_src || ffn_gate_bias->src[1] == ffn_gate_bias_src;
            if (!up_has_mul || !gate_has_mul) {
                return false;
            }
        } else { // GGML_OP_ADD_ID
-            if (ffn_up_bias->src[0] != ffn_up || ffn_gate_bias->src[0] != ffn_gate) {
+            if (ffn_up_bias->src[0] != ffn_up_bias_src || ffn_gate_bias->src[0] != ffn_gate_bias_src) {
                return false;
            }
            if (ffn_up_bias->src[2] != ffn_up->src[2] || ffn_gate_bias->src[2] != ffn_gate->src[2]) {
                return false;
            }
        }
-    } else {
-        if (glu->src[0] != ffn_gate && glu->src[1] != ffn_up) {
-            return false;
-        }
    }

    if (ffn_up->src[0]->type != ffn_gate->src[0]->type || !ggml_are_same_shape(ffn_up->src[0], ffn_gate->src[0]) ||
@@ -1638,7 +1655,7 @@ static bool ggml_cuda_should_fuse_mul_mat(const ggml_tensor * ffn_up,
        return false;
    }

-    if (ffn_up->src[2] && (ffn_up->src[2] != ffn_gate->src[2])) {
+    if (is_mul_mat_id && ffn_up->src[2] != ffn_gate->src[2]) {
        return false;
    }

@@ -3204,10 +3221,240 @@ static int ggml_cuda_try_fuse(ggml_backend_cuda_context * cuda_ctx, ggml_cgraph
    bool fused_mul_mat_vec = false;
    int  fused_node_count  = 0;

-    // gate + glu + up
+    auto get_mul_mat_scale = [](const ggml_tensor * scale_node, const ggml_tensor * mm_node) -> const ggml_tensor * {
+        const bool scale_lhs_mm = scale_node->src[0] == mm_node;
+        const bool scale_rhs_mm = scale_node->src[1] == mm_node;
+        if (!scale_lhs_mm && !scale_rhs_mm) {
+            return nullptr;
+        }
+
+        const ggml_tensor * scale = scale_lhs_mm ? scale_node->src[1] : scale_node->src[0];
+        if (mm_node->src[0]->type != GGML_TYPE_NVFP4 || scale_node->type != GGML_TYPE_F32 ||
+                scale->type != GGML_TYPE_F32 || !ggml_is_contiguous(scale) || ggml_nelements(scale) != 1 ||
+                !ggml_are_same_shape(scale_node, mm_node)) {
+            return nullptr;
+        }
+
+        return scale;
+    };
+
+    auto get_mul_mat_id_scale = [](const ggml_tensor * reshape, const ggml_tensor * repeat, const ggml_tensor * getrows,
+            const ggml_tensor * scale_node, const ggml_tensor * mm_node) -> const ggml_tensor * {
+        if (repeat->src[0] != reshape || getrows->src[0] != repeat || getrows->src[1] != mm_node->src[2]) {
+            return nullptr;
+        }
+        if (!((scale_node->src[0] == mm_node && scale_node->src[1] == getrows) ||
+                (scale_node->src[0] == getrows && scale_node->src[1] == mm_node))) {
+            return nullptr;
+        }
+
+        const ggml_tensor * scale = reshape->src[0];
+        if (mm_node->src[0]->type != GGML_TYPE_NVFP4 || scale_node->type != GGML_TYPE_F32 ||
+                scale->type != GGML_TYPE_F32 || !ggml_is_contiguous(scale) || ggml_nelements(scale) != mm_node->src[0]->ne[2] ||
+                !ggml_are_same_shape(scale_node, mm_node)) {
+            return nullptr;
+        }
+
+        return scale;
+    };
+
+    auto get_bias_tensor = [](const ggml_tensor * bias_node, const ggml_tensor * mul_node, ggml_op op_bias) -> const ggml_tensor * {
+        if (op_bias == GGML_OP_ADD) {
+            if (bias_node->src[0] == mul_node) {
+                return bias_node->src[1];
+            }
+            if (bias_node->src[1] == mul_node) {
+                return bias_node->src[0];
+            }
+            return nullptr;
+        }
+        GGML_ASSERT(op_bias == GGML_OP_ADD_ID);
+        GGML_ASSERT(bias_node->src[0] == mul_node);
+        return bias_node->src[1];
+    };
+
+    // gate + glu + up, with optional scale/bias on both lanes.
    for (ggml_op op : { GGML_OP_MUL_MAT, GGML_OP_MUL_MAT_ID }) {
        const ggml_op bias_op = op == GGML_OP_MUL_MAT ? GGML_OP_ADD : GGML_OP_ADD_ID;

+        if (op == GGML_OP_MUL_MAT) {
+            for (const bool with_bias : { false, true }) {
+                const int gate_idx       = i;
+                const int gate_scale_idx = i + 1;
+                const int gate_bias_idx  = with_bias ? i + 2 : -1;
+                const int up_idx         = with_bias ? i + 3 : i + 2;
+                const int up_scale_idx   = up_idx + 1;
+                const int up_bias_idx    = with_bias ? up_idx + 2 : -1;
+                const int glu_idx        = with_bias ? up_idx + 3 : up_idx + 2;
+
+                const int out_nodes[] = { glu_idx };
+                ggml_op ops[7];
+                if (with_bias) {
+                    ops[0] = op;
+                    ops[1] = GGML_OP_MUL;
+                    ops[2] = bias_op;
+                    ops[3] = op;
+                    ops[4] = GGML_OP_MUL;
+                    ops[5] = bias_op;
+                    ops[6] = GGML_OP_GLU;
+                } else {
+                    ops[0] = op;
+                    ops[1] = GGML_OP_MUL;
+                    ops[2] = op;
+                    ops[3] = GGML_OP_MUL;
+                    ops[4] = GGML_OP_GLU;
+                }
+                const int n_ops = with_bias ? 7 : 5;
+
+                if (!ggml_can_fuse_subgraph(cgraph, i, n_ops, ops, out_nodes, 1) ||
+                        !ggml_cuda_check_fusion_memory_ranges(cgraph, i, n_ops, out_nodes, 1)) {
+                    continue;
+                }
+
+                ggml_tensor * gate_n       = cgraph->nodes[gate_idx];
+                ggml_tensor * gate_scale_n = cgraph->nodes[gate_scale_idx];
+                ggml_tensor * gate_out_n   = with_bias ? cgraph->nodes[gate_bias_idx] : gate_scale_n;
+                ggml_tensor * up_n         = cgraph->nodes[up_idx];
+                ggml_tensor * up_scale_n   = cgraph->nodes[up_scale_idx];
+                ggml_tensor * up_out_n     = with_bias ? cgraph->nodes[up_bias_idx] : up_scale_n;
+                const ggml_tensor * glu = cgraph->nodes[glu_idx];
+
+                if (!ggml_cuda_should_fuse_mul_mat(up_n, gate_n, glu,
+                        with_bias ? up_out_n : nullptr, with_bias ? gate_out_n : nullptr, up_scale_n, gate_scale_n)) {
+                    continue;
+                }
+
+                const ggml_tensor * gate_scale = get_mul_mat_scale(gate_scale_n, gate_n);
+                const ggml_tensor * up_scale   = get_mul_mat_scale(up_scale_n, up_n);
+                if (!gate_scale || !up_scale) {
+                    continue;
+                }
+
+                const ggml_tensor * up_bias   = with_bias ? get_bias_tensor(up_out_n, up_scale_n, bias_op) : nullptr;
+                const ggml_tensor * gate_bias = with_bias ? get_bias_tensor(gate_out_n, gate_scale_n, bias_op) : nullptr;
+                if (with_bias && (!ggml_are_same_shape(gate_out_n->src[0], gate_out_n->src[1]) ||
+                        !ggml_are_same_shape(up_out_n->src[0], up_out_n->src[1]))) {
+                    continue;
+                }
+
+                const ggml_tensor * src0 = up_n->src[0];
+                const ggml_tensor * src1 = up_n->src[1];
+                const ggml_tensor * ids  = up_n->src[2];
+
+                ggml_cuda_mm_fusion_args_host fusion_data{};
+                fusion_data.gate       = gate_n->src[0];
+                fusion_data.x_bias     = up_bias;
+                fusion_data.gate_bias  = gate_bias;
+                fusion_data.x_scale    = up_scale;
+                fusion_data.gate_scale = gate_scale;
+                fusion_data.glu_op     = ggml_get_glu_op(glu);
+
+                if (ggml_cuda_should_fuse_mul_mat_vec_q(up_n)) {
+                    ggml_cuda_mul_mat_vec_q(*cuda_ctx, src0, src1, ids, cgraph->nodes[glu_idx], &fusion_data);
+                    fused_mul_mat_vec = true;
+                    fused_node_count  = n_ops;
+                    break;
+                }
+            }
+
+            if (fused_mul_mat_vec) {
+                break;
+            }
+        } else {
+            for (const bool with_bias : { false, true }) {
+                const int gate_idx       = i;
+                const int gate_scale_idx = i + 4;
+                const int gate_bias_idx  = with_bias ? i + 5 : -1;
+                const int up_idx         = with_bias ? i + 6 : i + 5;
+                const int up_scale_idx   = up_idx + 4;
+                const int up_bias_idx    = with_bias ? up_idx + 5 : -1;
+                const int glu_idx        = with_bias ? up_idx + 6 : up_idx + 5;
+
+                const int out_nodes[] = { glu_idx };
+                ggml_op ops[13];
+                if (with_bias) {
+                    ops[0]  = op;
+                    ops[1]  = GGML_OP_RESHAPE;
+                    ops[2]  = GGML_OP_REPEAT;
+                    ops[3]  = GGML_OP_GET_ROWS;
+                    ops[4]  = GGML_OP_MUL;
+                    ops[5]  = bias_op;
+                    ops[6]  = op;
+                    ops[7]  = GGML_OP_RESHAPE;
+                    ops[8]  = GGML_OP_REPEAT;
+                    ops[9]  = GGML_OP_GET_ROWS;
+                    ops[10] = GGML_OP_MUL;
+                    ops[11] = bias_op;
+                    ops[12] = GGML_OP_GLU;
+                } else {
+                    ops[0]  = op;
+                    ops[1]  = GGML_OP_RESHAPE;
+                    ops[2]  = GGML_OP_REPEAT;
+                    ops[3]  = GGML_OP_GET_ROWS;
+                    ops[4]  = GGML_OP_MUL;
+                    ops[5]  = op;
+                    ops[6]  = GGML_OP_RESHAPE;
+                    ops[7]  = GGML_OP_REPEAT;
+                    ops[8]  = GGML_OP_GET_ROWS;
+                    ops[9]  = GGML_OP_MUL;
+                    ops[10] = GGML_OP_GLU;
+                }
+                const int n_ops = with_bias ? 13 : 11;
+
+                if (!ggml_can_fuse_subgraph(cgraph, i, n_ops, ops, out_nodes, 1) ||
+                        !ggml_cuda_check_fusion_memory_ranges(cgraph, i, n_ops, out_nodes, 1)) {
+                    continue;
+                }
+
+                ggml_tensor * gate_n       = cgraph->nodes[gate_idx];
+                ggml_tensor * gate_scale_n = cgraph->nodes[gate_scale_idx];
+                ggml_tensor * gate_out_n   = with_bias ? cgraph->nodes[gate_bias_idx] : gate_scale_n;
+                ggml_tensor * up_n         = cgraph->nodes[up_idx];
+                ggml_tensor * up_scale_n   = cgraph->nodes[up_scale_idx];
+                ggml_tensor * up_out_n     = with_bias ? cgraph->nodes[up_bias_idx] : up_scale_n;
+                const ggml_tensor * glu = cgraph->nodes[glu_idx];
+
+                if (!ggml_cuda_should_fuse_mul_mat(up_n, gate_n, glu,
+                        with_bias ? up_out_n : nullptr, with_bias ? gate_out_n : nullptr, up_scale_n, gate_scale_n)) {
+                    continue;
+                }
+
+                const ggml_tensor * gate_scale = get_mul_mat_id_scale(cgraph->nodes[gate_idx + 1], cgraph->nodes[gate_idx + 2],
+                        cgraph->nodes[gate_idx + 3], gate_scale_n, gate_n);
+                const ggml_tensor * up_scale = get_mul_mat_id_scale(cgraph->nodes[up_idx + 1], cgraph->nodes[up_idx + 2],
+                        cgraph->nodes[up_idx + 3], up_scale_n, up_n);
+                if (!gate_scale || !up_scale) {
+                    continue;
+                }
+
+                const ggml_tensor * up_bias   = with_bias ? get_bias_tensor(up_out_n, up_scale_n, bias_op) : nullptr;
+                const ggml_tensor * gate_bias = with_bias ? get_bias_tensor(gate_out_n, gate_scale_n, bias_op) : nullptr;
+
+                const ggml_tensor * src0 = up_n->src[0];
+                const ggml_tensor * src1 = up_n->src[1];
+                const ggml_tensor * ids  = up_n->src[2];
+
+                ggml_cuda_mm_fusion_args_host fusion_data{};
+                fusion_data.gate       = gate_n->src[0];
+                fusion_data.x_bias     = up_bias;
+                fusion_data.gate_bias  = gate_bias;
+                fusion_data.x_scale    = up_scale;
+                fusion_data.gate_scale = gate_scale;
+                fusion_data.glu_op     = ggml_get_glu_op(glu);
+
+                if (ggml_cuda_should_fuse_mul_mat_vec_q(up_n)) {
+                    ggml_cuda_mul_mat_vec_q(*cuda_ctx, src0, src1, ids, cgraph->nodes[glu_idx], &fusion_data);
+                    fused_mul_mat_vec = true;
+                    fused_node_count  = n_ops;
+                    break;
+                }
+            }
+
+            if (fused_mul_mat_vec) {
+                break;
+            }
+        }
+
        if (ggml_cuda_can_fuse(cgraph, i, { op, bias_op, op, bias_op, GGML_OP_GLU }, {})) {
            ggml_tensor * glu         = cgraph->nodes[i + 4];
            ggml_tensor * gate_bias_n = glu->src[0];
@@ -3227,23 +3474,8 @@ static int ggml_cuda_try_fuse(ggml_backend_cuda_context * cuda_ctx, ggml_cgraph
                continue;
            }

-            auto get_bias_tensor = [](const ggml_tensor * bias_node, const ggml_tensor * mul_node, ggml_op op_bias) {
-                if (op_bias == GGML_OP_ADD) {
-                    if (bias_node->src[0] == mul_node) {
-                        return bias_node->src[1];
-                    }
-                    if (bias_node->src[1] == mul_node) {
-                        return bias_node->src[0];
-                    }
-                    return (ggml_tensor *) nullptr;
-                }
-                GGML_ASSERT(op_bias == GGML_OP_ADD_ID);
-                GGML_ASSERT(bias_node->src[0] == mul_node);
-                return bias_node->src[1];
-            };
-
-            ggml_tensor * up_bias_tensor   = get_bias_tensor(up_bias_n, up_n, bias_op);
-            ggml_tensor * gate_bias_tensor = get_bias_tensor(gate_bias_n, gate_n, bias_op);
+            const ggml_tensor * up_bias_tensor   = get_bias_tensor(up_bias_n, up_n, bias_op);
+            const ggml_tensor * gate_bias_tensor = get_bias_tensor(gate_bias_n, gate_n, bias_op);

            if (!up_bias_tensor || !gate_bias_tensor) {
                continue;
@@ -3331,7 +3563,95 @@ static int ggml_cuda_try_fuse(ggml_backend_cuda_context * cuda_ctx, ggml_cgraph
    fused_mul_mat_vec = false;
    fused_node_count  = 0;

-    // gate + add + glu + up + add
+    // mul_mat + scale + optional bias
+    for (ggml_op op : { GGML_OP_MUL_MAT, GGML_OP_MUL_MAT_ID }) {
+        const ggml_op bias_op = op == GGML_OP_MUL_MAT ? GGML_OP_ADD : GGML_OP_ADD_ID;
+
+        for (const bool with_bias : { false, true }) {
+            const int n_ops = op == GGML_OP_MUL_MAT ? (with_bias ? 3 : 2) : (with_bias ? 6 : 5);
+            const int out_nodes[] = { i + n_ops - 1 };
+            ggml_op ops[6];
+            if (op == GGML_OP_MUL_MAT) {
+                if (with_bias) {
+                    ops[0] = op;
+                    ops[1] = GGML_OP_MUL;
+                    ops[2] = bias_op;
+                } else {
+                    ops[0] = op;
+                    ops[1] = GGML_OP_MUL;
+                }
+            } else {
+                if (with_bias) {
+                    ops[0] = op;
+                    ops[1] = GGML_OP_RESHAPE;
+                    ops[2] = GGML_OP_REPEAT;
+                    ops[3] = GGML_OP_GET_ROWS;
+                    ops[4] = GGML_OP_MUL;
+                    ops[5] = bias_op;
+                } else {
+                    ops[0] = op;
+                    ops[1] = GGML_OP_RESHAPE;
+                    ops[2] = GGML_OP_REPEAT;
+                    ops[3] = GGML_OP_GET_ROWS;
+                    ops[4] = GGML_OP_MUL;
+                }
+            }
+
+            if (!ggml_can_fuse_subgraph(cgraph, i, n_ops, ops, out_nodes, 1) ||
+                    !ggml_cuda_check_fusion_memory_ranges(cgraph, i, n_ops, out_nodes, 1)) {
+                continue;
+            }
+
+            ggml_tensor * mm_node    = cgraph->nodes[i];
+            ggml_tensor * scale_node = op == GGML_OP_MUL_MAT ? cgraph->nodes[i + 1] : cgraph->nodes[i + 4];
+            ggml_tensor * out_node   = with_bias ? cgraph->nodes[i + n_ops - 1] : scale_node;
+
+            const ggml_tensor * scale = nullptr;
+            if (op == GGML_OP_MUL_MAT) {
+                scale = get_mul_mat_scale(scale_node, mm_node);
+            } else {
+                scale = get_mul_mat_id_scale(cgraph->nodes[i + 1], cgraph->nodes[i + 2], cgraph->nodes[i + 3], scale_node, mm_node);
+            }
+            if (!scale) {
+                continue;
+            }
+
+            const ggml_tensor * bias = with_bias ? get_bias_tensor(out_node, scale_node, bias_op) : nullptr;
+            if (with_bias && !bias) {
+                continue;
+            }
+            if (with_bias && bias_op == GGML_OP_ADD && !ggml_are_same_shape(out_node->src[0], out_node->src[1])) {
+                continue;
+            }
+            if (with_bias && bias_op == GGML_OP_ADD_ID && out_node->src[2] != mm_node->src[2]) {
+                continue;
+            }
+
+            const ggml_tensor * src0 = mm_node->src[0];
+            const ggml_tensor * src1 = mm_node->src[1];
+            const ggml_tensor * ids  = mm_node->src[2];
+
+            ggml_cuda_mm_fusion_args_host fusion_data{};
+            fusion_data.x_bias  = bias;
+            fusion_data.x_scale = scale;
+
+            if (ggml_cuda_should_fuse_mul_mat_vec_q(mm_node)) {
+                ggml_cuda_mul_mat_vec_q(*cuda_ctx, src0, src1, ids, out_node, &fusion_data);
+                fused_mul_mat_vec = true;
+                fused_node_count  = n_ops;
+                break;
+            }
+        }
+        if (fused_mul_mat_vec) {
+            break;
+        }
+    }
+
+    if (fused_mul_mat_vec) {
+        return fused_node_count - 1;
+    }
+
+    // mul_mat + add
    for (ggml_op op : { GGML_OP_MUL_MAT, GGML_OP_MUL_MAT_ID }) {
        const ggml_op bias_op = op == GGML_OP_MUL_MAT ? GGML_OP_ADD : GGML_OP_ADD_ID;

@@ -3562,12 +3882,6 @@ static void ggml_cuda_graph_evaluate_and_capture(ggml_backend_cuda_context * cud
                    }
                }

-#ifdef GGML_CUDA_DEBUG
-                const int nodes_fused = i - prev_i - 1;
-                if (nodes_fused > 0) {
-                    GGML_LOG_INFO("nodes_fused: %d\n", nodes_fused);
-                }
-#endif
                prev_i = i;

                if (ggml_cuda_is_view_or_noop(node)) {
@@ -3581,6 +3895,12 @@ static void ggml_cuda_graph_evaluate_and_capture(ggml_backend_cuda_context * cud
                int nodes_to_skip = ggml_cuda_try_fuse(cuda_ctx, cgraph, i);

                if (nodes_to_skip != 0) {
+#ifdef GGML_CUDA_DEBUG
+                    const int last_fused = i + nodes_to_skip;
+                    GGML_LOG_INFO("nodes_fused: %d, first: %s (%s), last: %s (%s)\n",
+                            nodes_to_skip + 1, ggml_op_name(node->op), node->name,
+                            ggml_op_name(cgraph->nodes[last_fused]->op), cgraph->nodes[last_fused]->name);
+#endif
                    i += nodes_to_skip;
                    continue;
                }
@@ -521,9 +521,13 @@ static __global__ void mul_mat_vec_q(
    bool use_gate = false;
    bool use_bias = false;
    bool use_gate_bias = false;
+    bool use_scale = false;
+    bool use_gate_scale = false;
    [[maybe_unused]] const void * vgate = nullptr;
    const float * x_bias = nullptr;
    const float * gate_bias = nullptr;
+    const float * x_scale = nullptr;
+    const float * gate_scale = nullptr;
    ggml_glu_op active_glu;

    if constexpr (has_fusion) {
@@ -534,34 +538,47 @@ static __global__ void mul_mat_vec_q(
        x_bias        = (const float *) fusion.x_bias;
        gate_bias     = (const float *) fusion.gate_bias;
        active_glu    = fusion.glu_op;
+        if constexpr (type == GGML_TYPE_NVFP4) {
+            use_scale      = fusion.x_scale    != nullptr;
+            use_gate_scale = fusion.gate_scale != nullptr && use_gate;
+            x_scale        = (const float *) fusion.x_scale;
+            gate_scale     = (const float *) fusion.gate_scale;
+        }
    }


    [[maybe_unused]] float x_biases[ncols_dst]    = { 0.0f };
    [[maybe_unused]] float gate_biases[ncols_dst] = { 0.0f };
+    [[maybe_unused]] float x_scales;
+    [[maybe_unused]] float gate_scales;
    if constexpr (has_fusion) {
+        // 1. Hide latency by prefetching bias, gates and scales here
+        // 2. load only on threads that won't die after partial sum calculation
        const uint32_t channel_bias = ids ? channel_x : channel_dst;
-        if (use_bias) {
-            x_bias = x_bias + sample_dst*stride_sample_dst + channel_bias*stride_channel_dst + row0;
-            // 1. Hide latency by prefetching bias and gate here
-            // 2. load only on threads that won't die after partial sum calculation
-            if (threadIdx.x < rows_per_cuda_block && threadIdx.y == 0 &&
-                (rows_per_cuda_block == 1 || uint32_t(row0 + threadIdx.x) < stride_col_dst)) {
+        if (threadIdx.x < rows_per_cuda_block && threadIdx.y == 0 &&
+            (rows_per_cuda_block == 1 || uint32_t(row0 + threadIdx.x) < stride_col_dst)) {
+            if (use_bias) {
+                x_bias = x_bias + sample_dst * stride_sample_dst + channel_bias * stride_channel_dst + row0;
 #pragma unroll
                for (int j = 0; j < ncols_dst; ++j) {
                    x_biases[j] = x_bias[j * stride_col_dst + threadIdx.x];
                }
            }
-        }
-        if (use_gate_bias) {
-            gate_bias = gate_bias + sample_dst*stride_sample_dst + channel_bias*stride_channel_dst + row0;
-            if (threadIdx.x < rows_per_cuda_block && threadIdx.y == 0 &&
-                (rows_per_cuda_block == 1 || uint32_t(row0 + threadIdx.x) < stride_col_dst)) {
+            if (use_gate_bias) {
+                gate_bias = gate_bias + sample_dst * stride_sample_dst + channel_bias * stride_channel_dst + row0;
 #pragma unroll
                for (int j = 0; j < ncols_dst; ++j) {
                    gate_biases[j] = gate_bias[j * stride_col_dst + threadIdx.x];
                }
            }
+            if constexpr (type == GGML_TYPE_NVFP4) {
+                if (use_scale) {
+                    x_scales = x_scale[ids ? channel_x : 0];
+                }
+                if (use_gate_scale) {
+                    gate_scales = gate_scale[ids ? channel_x : 0];
+                }
+            }
        }
    }

@@ -643,11 +660,21 @@ static __global__ void mul_mat_vec_q(
        if (threadIdx.x < rows_per_cuda_block && (rows_per_cuda_block == 1 || uint32_t(row0 + threadIdx.x) < stride_col_dst)) {
            float result = tmp[j][threadIdx.x];
            if constexpr (has_fusion) {
+                if constexpr (type == GGML_TYPE_NVFP4) {
+                    if (use_scale) {
+                        result *= x_scales;
+                    }
+                }
                if (use_bias) {
                    result += x_biases[j];
                }
                if (use_gate) {
                    float gate_value = tmp_gate[j][threadIdx.x];
+                    if constexpr (type == GGML_TYPE_NVFP4) {
+                        if (use_gate_scale) {
+                            gate_value *= gate_scales;
+                        }
+                    }
                    if (use_gate_bias) {
                        gate_value += gate_biases[j];
                    }
@@ -673,7 +700,10 @@ static __global__ void mul_mat_vec_q(
    }

    if constexpr (!has_fusion) {
-        GGML_UNUSED_VARS(use_gate, use_bias, use_gate_bias, active_glu, gate_bias, x_bias, tmp_gate);
+        GGML_UNUSED_VARS(use_gate, use_bias, use_gate_bias, use_scale, use_gate_scale, active_glu, gate_bias, x_bias, x_scale, gate_scale, tmp_gate);
+    }
+    if constexpr (type != GGML_TYPE_NVFP4) {
+        GGML_UNUSED_VARS(use_scale, use_gate_scale, x_scale, gate_scale, x_scales, gate_scales);
    }
 }

@@ -769,7 +799,8 @@ static void mul_mat_vec_q_switch_fusion(
        const dim3 & block_nums, const dim3 & block_dims, const int nbytes_shared,
        const uint32_t ids_stride, cudaStream_t stream) {

-    const bool has_fusion = fusion.gate != nullptr || fusion.x_bias != nullptr || fusion.gate_bias != nullptr;
+    const bool has_fusion = fusion.gate != nullptr || fusion.x_bias != nullptr || fusion.gate_bias != nullptr ||
+                            fusion.x_scale != nullptr || fusion.gate_scale != nullptr;
    if constexpr (c_ncols_dst == 1) {
        if (has_fusion) {
            const ggml_cuda_kernel_launch_params launch_params = ggml_cuda_kernel_launch_params(block_nums, block_dims, nbytes_shared, stream);
@@ -834,7 +865,6 @@ static void mul_mat_vec_q_switch_ncols_dst(
    const int warp_size = ggml_cuda_info().devices[device].warp_size;
    const mmvq_parameter_table_id table_id  = get_device_table_id(cc);

-    const bool has_fusion = fusion.gate != nullptr || fusion.x_bias != nullptr || fusion.gate_bias != nullptr;
    const bool has_ids = ids != nullptr;

    const auto should_use_small_k = [&](int c_ncols_dst) {
@@ -973,8 +1003,6 @@ static void mul_mat_vec_q_switch_ncols_dst(
            GGML_ABORT("fatal error");
            break;
    }
-
-    GGML_UNUSED(has_fusion);
 }
 static void mul_mat_vec_q_switch_type(
        const void * vx, const ggml_type type_x, const void * vy, const int32_t * ids, const ggml_cuda_mm_fusion_args_device fusion, float * dst,
@@ -1154,6 +1182,9 @@ void ggml_cuda_mul_mat_vec_q(
    if (fusion) {
        GGML_ASSERT( !ids || dst->ne[2] == 1);
        GGML_ASSERT(  ids || dst->ne[1] == 1);
+        // Scale fusion is only allowed for NVFP4 currently as the cost of checking this at run-time in the prologue is
+        // non-negligible for some models such as gpt-oss-20b
+        GGML_ASSERT((fusion->x_scale == nullptr && fusion->gate_scale == nullptr) || src0->type == GGML_TYPE_NVFP4);

        if (fusion->x_bias) {
            GGML_ASSERT(fusion->x_bias->type == GGML_TYPE_F32);
@@ -1171,6 +1202,18 @@ void ggml_cuda_mul_mat_vec_q(
            GGML_ASSERT(!ids || fusion->gate_bias->ne[1] == src0->ne[2]);
            fusion_local.gate_bias = fusion->gate_bias->data;
        }
+        if (fusion->x_scale) {
+            GGML_ASSERT(fusion->x_scale->type == GGML_TYPE_F32);
+            GGML_ASSERT(ggml_is_contiguous(fusion->x_scale));
+            GGML_ASSERT(ggml_nelements(fusion->x_scale) == (ids ? src0->ne[2] : 1));
+            fusion_local.x_scale = fusion->x_scale->data;
+        }
+        if (fusion->gate_scale) {
+            GGML_ASSERT(fusion->gate_scale->type == GGML_TYPE_F32);
+            GGML_ASSERT(ggml_is_contiguous(fusion->gate_scale));
+            GGML_ASSERT(ggml_nelements(fusion->gate_scale) == (ids ? src0->ne[2] : 1));
+            fusion_local.gate_scale = fusion->gate_scale->data;
+        }
        fusion_local.glu_op = fusion->glu_op;
    }

@@ -71,6 +71,44 @@ void quantize_row_q1_0_ref(const float * GGML_RESTRICT x, block_q1_0 * GGML_REST
    }
 }

+void quantize_row_q2_0_ref(const float * GGML_RESTRICT x, block_q2_0 * GGML_RESTRICT y, int64_t k) {
+    static const int qk = QK2_0;
+
+    assert(k % qk == 0);
+
+    const int nb = k / qk;
+
+    for (int i = 0; i < nb; i++) {
+        // Compute scale as max absolute value in the block
+        float amax = 0.0f;
+        for (int j = 0; j < qk; j++) {
+            const float a = fabsf(x[i*qk + j]);
+            if (a > amax) amax = a;
+        }
+        const float d = amax;
+        const float id = d > 0.0f ? 1.0f / d : 0.0f;
+
+        y[i].d = GGML_FP32_TO_FP16(d);
+
+        // Clear quant bytes
+        for (int j = 0; j < qk / 4; ++j) {
+            y[i].qs[j] = 0;
+        }
+
+        // Encode 2-bit values: round(w/d) clamped to [-1, 2], then add 1
+        // 00 (-1) = -scale, 01 (0) = 0, 10 (+1) = +scale, 11 (+2) = 2*scale
+        for (int j = 0; j < qk; ++j) {
+            const float w = x[i*qk + j];
+            int q = (int)roundf(w * id) + 1;
+            if (q < 0) q = 0;
+            if (q > 3) q = 3;
+            const int byte_index = j / 4;
+            const int bit_offset = (j % 4) * 2;
+            y[i].qs[byte_index] |= ((uint8_t)q << bit_offset);
+        }
+    }
+}
+
 // reference implementation for deterministic creation of model files
 void quantize_row_q4_0_ref(const float * GGML_RESTRICT x, block_q4_0 * GGML_RESTRICT y, int64_t k) {
    static const int qk = QK4_0;
@@ -398,6 +436,26 @@ void dequantize_row_q1_0(const block_q1_0 * GGML_RESTRICT x, float * GGML_RESTRI
    }
 }

+void dequantize_row_q2_0(const block_q2_0 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k) {
+    static const int qk = QK2_0;
+
+    assert(k % qk == 0);
+
+    const int nb = k / qk;
+
+    for (int i = 0; i < nb; i++) {
+        const float d = GGML_FP16_TO_FP32(x[i].d);
+
+        for (int j = 0; j < qk; ++j) {
+            const int byte_index = j / 4;
+            const int bit_offset = (j % 4) * 2;
+            const uint8_t q = (x[i].qs[byte_index] >> bit_offset) & 0x03;
+            // 00=-1, 01=0, 10=+1, 11=+2
+            y[i*qk + j] = ((int)q - 1) * d;
+        }
+    }
+}
+
 void dequantize_row_q4_0(const block_q4_0 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k) {
    static const int qk = QK4_0;

@@ -2052,6 +2110,20 @@ size_t quantize_q1_0(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst,
    return nrow * row_size;
 }

+size_t quantize_q2_0(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) {
+    if (!quant_weights) {
+        quantize_row_q2_0_ref(src, dst, (int64_t)nrow*n_per_row);
+        return nrow * ggml_row_size(GGML_TYPE_Q2_0, n_per_row);
+    }
+    size_t row_size = ggml_row_size(GGML_TYPE_Q2_0, n_per_row);
+    char * qrow = (char *)dst;
+    for (int64_t row = 0; row < nrow; ++row) {
+        quantize_row_q2_0_ref(src, (block_q2_0*)qrow, n_per_row);
+        src += n_per_row;
+        qrow += row_size;
+    }
+    return nrow * row_size;
+}

 size_t quantize_q4_0(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) {
    if (!quant_weights) {
@@ -5461,6 +5533,10 @@ bool ggml_validate_row_data(enum ggml_type type, const void * data, size_t nbyte
            {
                VALIDATE_ROW_DATA_D_F16_IMPL(block_q1_0, data, nb);
            } break;
+        case GGML_TYPE_Q2_0:
+            {
+                VALIDATE_ROW_DATA_D_F16_IMPL(block_q2_0, data, nb);
+            } break;
        case GGML_TYPE_Q4_0:
            {
                VALIDATE_ROW_DATA_D_F16_IMPL(block_q4_0, data, nb);
@@ -15,6 +15,7 @@ extern "C" {

 // Quantization
 GGML_API void quantize_row_q1_0_ref(const float * GGML_RESTRICT x, block_q1_0 * GGML_RESTRICT y, int64_t k);
+GGML_API void quantize_row_q2_0_ref(const float * GGML_RESTRICT x, block_q2_0 * GGML_RESTRICT y, int64_t k);
 GGML_API void quantize_row_q4_0_ref(const float * GGML_RESTRICT x, block_q4_0 * GGML_RESTRICT y, int64_t k);
 GGML_API void quantize_row_q4_1_ref(const float * GGML_RESTRICT x, block_q4_1 * GGML_RESTRICT y, int64_t k);
 GGML_API void quantize_row_q5_0_ref(const float * GGML_RESTRICT x, block_q5_0 * GGML_RESTRICT y, int64_t k);
@@ -43,6 +44,7 @@ GGML_API void quantize_row_iq2_s_ref  (const float * GGML_RESTRICT x, block_iq2_

 // Dequantization
 GGML_API void dequantize_row_q1_0(const block_q1_0 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
+GGML_API void dequantize_row_q2_0(const block_q2_0 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
 GGML_API void dequantize_row_q4_0(const block_q4_0 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
 GGML_API void dequantize_row_q4_1(const block_q4_1 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
 GGML_API void dequantize_row_q5_0(const block_q5_0 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
@@ -93,6 +95,7 @@ GGML_API size_t quantize_q4_K(const float * GGML_RESTRICT src, void * GGML_RESTR
 GGML_API size_t quantize_q5_K(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
 GGML_API size_t quantize_q6_K(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
 GGML_API size_t quantize_q1_0(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
+GGML_API size_t quantize_q2_0(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
 GGML_API size_t quantize_q4_0(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
 GGML_API size_t quantize_q4_1(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
 GGML_API size_t quantize_q5_0(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
@@ -681,6 +681,14 @@ static const struct ggml_type_traits type_traits[GGML_TYPE_COUNT] = {
        .to_float                 = (ggml_to_float_t) dequantize_row_q1_0,
        .from_float_ref           = (ggml_from_float_t) quantize_row_q1_0_ref,
    },
+    [GGML_TYPE_Q2_0] = {
+        .type_name                = "q2_0",
+        .blck_size                = QK2_0,
+        .type_size                = sizeof(block_q2_0),
+        .is_quantized             = true,
+        .to_float                 = (ggml_to_float_t) dequantize_row_q2_0,
+        .from_float_ref           = (ggml_from_float_t) quantize_row_q2_0_ref,
+    },
    [GGML_TYPE_Q4_0] = {
        .type_name                = "q4_0",
        .blck_size                = QK4_0,
@@ -1417,6 +1425,7 @@ enum ggml_type ggml_ftype_to_ggml_type(enum ggml_ftype ftype) {
        case GGML_FTYPE_MOSTLY_Q4_0:          wtype = GGML_TYPE_Q4_0;  break;
        case GGML_FTYPE_MOSTLY_Q4_1:          wtype = GGML_TYPE_Q4_1;  break;
        case GGML_FTYPE_MOSTLY_Q1_0:          wtype = GGML_TYPE_Q1_0;  break;
+        case GGML_FTYPE_MOSTLY_Q2_0:          wtype = GGML_TYPE_Q2_0;  break;
        case GGML_FTYPE_MOSTLY_Q5_0:          wtype = GGML_TYPE_Q5_0;  break;
        case GGML_FTYPE_MOSTLY_Q5_1:          wtype = GGML_TYPE_Q5_1;  break;
        case GGML_FTYPE_MOSTLY_Q8_0:          wtype = GGML_TYPE_Q8_0;  break;
@@ -7419,6 +7428,10 @@ static int ggml_node_list_find_tensor(const struct ggml_cgraph * cgraph,
    return -1;
 }

+static bool ggml_is_constant(const struct ggml_tensor * tensor) {
+    return tensor->buffer != NULL && ggml_backend_buffer_get_usage(tensor->buffer) == GGML_BACKEND_BUFFER_USAGE_WEIGHTS && (tensor->flags & GGML_TENSOR_FLAG_PARAM) == 0;
+}
+
 bool ggml_can_fuse_subgraph_ext(const struct ggml_cgraph * cgraph,
                                const int *                node_idxs,
                                int                        count,
@@ -7464,10 +7477,11 @@ bool ggml_can_fuse_subgraph_ext(const struct ggml_cgraph * cgraph,
            return false;
        }

-        // if node is a view, check if the view_src and all it's parent view_srcs are within the subgraph
+        // if node is a view, check if the view_src and all its parent view_srcs are within the subgraph.
+        // external view sources are allowed only for weight tensors, which are constant for this graph execution.
        struct ggml_tensor * view_src = node->view_src;
        while (view_src) {
-            if (ggml_node_list_find_tensor(cgraph, node_idxs, count, view_src) == -1) {
+            if (ggml_node_list_find_tensor(cgraph, node_idxs, count, view_src) == -1 && !ggml_is_constant(view_src)) {
                return false;
            }
            view_src = view_src->view_src;
@@ -7739,6 +7753,7 @@ size_t ggml_quantize_chunk(

    switch (type) {
        case GGML_TYPE_Q1_0:    result = quantize_q1_0   (src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;
+        case GGML_TYPE_Q2_0:    result = quantize_q2_0   (src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;
        case GGML_TYPE_Q4_0:    result = quantize_q4_0   (src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;
        case GGML_TYPE_Q4_1:    result = quantize_q4_1   (src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;
        case GGML_TYPE_Q5_0:    result = quantize_q5_0   (src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;
@@ -4533,6 +4533,7 @@ class GGMLQuantizationType(IntEnum):
    MXFP4   = 39
    NVFP4   = 40
    Q1_0    = 41
+    Q2_0    = 42


 class ExpertGatingFuncType(IntEnum):
@@ -4588,6 +4589,7 @@ class LlamaFileType(IntEnum):
    MOSTLY_MXFP4_MOE     = 38  # except 1d tensors
    MOSTLY_NVFP4         = 39  # except 1d tensors
    MOSTLY_Q1_0          = 40  # except 1d tensors
+    MOSTLY_Q2_0          = 41  # except 1d tensors

    GUESSED              = 1024  # not specified in the model file

@@ -4713,6 +4715,7 @@ GGML_QUANT_SIZES: dict[GGMLQuantizationType, tuple[int, int]] = {
    GGMLQuantizationType.MXFP4:   (32, 1 + 16),
    GGMLQuantizationType.NVFP4:   (64, 4 + 32),
    GGMLQuantizationType.Q1_0:    (128, 2 + 16),
+    GGMLQuantizationType.Q2_0:    (64, 2 + 16),
 }


@@ -155,6 +155,7 @@ extern "C" {
        LLAMA_FTYPE_MOSTLY_MXFP4_MOE     = 38, // except 1d tensors
        LLAMA_FTYPE_MOSTLY_NVFP4         = 39, // except 1d tensors
        LLAMA_FTYPE_MOSTLY_Q1_0          = 40, // except 1d tensors
+        LLAMA_FTYPE_MOSTLY_Q2_0          = 41, // except 1d tensors

        LLAMA_FTYPE_GUESSED = 1024, // not specified in the model file
    };
@@ -37,6 +37,7 @@ const char * llama_ftype_name(llama_ftype ftype) {
        case LLAMA_FTYPE_MOSTLY_F16:       name = LLAMA_FTYPE_PREFIX "F16"; break;
        case LLAMA_FTYPE_MOSTLY_BF16:      name = LLAMA_FTYPE_PREFIX "BF16"; break;
        case LLAMA_FTYPE_MOSTLY_Q1_0:      name = LLAMA_FTYPE_PREFIX "Q1_0"; break;
+        case LLAMA_FTYPE_MOSTLY_Q2_0:      name = LLAMA_FTYPE_PREFIX "Q2_0"; break;
        case LLAMA_FTYPE_MOSTLY_Q4_0:      name = LLAMA_FTYPE_PREFIX "Q4_0"; break;
        case LLAMA_FTYPE_MOSTLY_Q4_1:      name = LLAMA_FTYPE_PREFIX "Q4_1"; break;
        case LLAMA_FTYPE_MOSTLY_Q5_0:      name = LLAMA_FTYPE_PREFIX "Q5_0"; break;
@@ -767,6 +768,7 @@ llama_model_loader::llama_model_loader(
            case GGML_TYPE_IQ3_S:   ftype = LLAMA_FTYPE_MOSTLY_IQ3_S;   break;
            case GGML_TYPE_NVFP4:   ftype = LLAMA_FTYPE_MOSTLY_NVFP4;   break;
            case GGML_TYPE_Q1_0:    ftype = LLAMA_FTYPE_MOSTLY_Q1_0;    break;
+            case GGML_TYPE_Q2_0:    ftype = LLAMA_FTYPE_MOSTLY_Q2_0;    break;
            default:
                {
                    LLAMA_LOG_WARN("%s: unknown type %s\n", __func__, ggml_type_name(type_max));
@@ -380,6 +380,7 @@ static ggml_type tensor_type_fallback(quantize_state_impl & qs, const ggml_tenso
            case GGML_TYPE_IQ3_XXS:
            case GGML_TYPE_IQ3_S:   // types on the right: block size 32
            case GGML_TYPE_IQ4_XS:  return_type = GGML_TYPE_IQ4_NL; break;
+            case GGML_TYPE_Q2_0:
            case GGML_TYPE_Q2_K:
            case GGML_TYPE_Q3_K:
            case GGML_TYPE_TQ1_0:
@@ -480,7 +481,7 @@ static ggml_type llama_tensor_get_type_impl(quantize_state_impl & qs, ggml_type
            else if (ftype == LLAMA_FTYPE_MOSTLY_IQ3_XXS) {
                new_type = GGML_TYPE_IQ3_S;
            }
-            else if (ftype == LLAMA_FTYPE_MOSTLY_TQ1_0 || ftype == LLAMA_FTYPE_MOSTLY_TQ2_0) {
+            else if (ftype == LLAMA_FTYPE_MOSTLY_TQ1_0 || ftype == LLAMA_FTYPE_MOSTLY_TQ2_0 || ftype == LLAMA_FTYPE_MOSTLY_Q2_0) {
                new_type = GGML_TYPE_Q4_K;
            }
        }
@@ -800,6 +801,7 @@ ggml_type llama_ftype_get_default_type(llama_ftype ftype) {
        case LLAMA_FTYPE_MOSTLY_BF16: return GGML_TYPE_BF16;
        case LLAMA_FTYPE_ALL_F32:     return GGML_TYPE_F32;
        case LLAMA_FTYPE_MOSTLY_Q1_0: return GGML_TYPE_Q1_0;
+        case LLAMA_FTYPE_MOSTLY_Q2_0: return GGML_TYPE_Q2_0;

        case LLAMA_FTYPE_MOSTLY_MXFP4_MOE: return GGML_TYPE_MXFP4;

@@ -1137,6 +1137,10 @@ struct test_case {
    }

    virtual ggml_tensor * build_graph(ggml_context * ctx) = 0;
+    virtual ggml_tensor * build_graph(ggml_context * ctx, ggml_context * ctx_weights) {
+        GGML_UNUSED(ctx_weights);
+        return build_graph(ctx);
+    }

    virtual double max_nmse_err() {
        return 1e-7;
@@ -1213,6 +1217,7 @@ struct test_case {

    virtual bool run_whole_graph() { return false; }
    virtual std::vector<ggml_tensor *> fusion_test_nodes() { return {}; }
+    virtual bool use_weight_context() { return false; }

    ggml_cgraph * gf = nullptr;
    ggml_cgraph * gb = nullptr;
@@ -1319,20 +1324,28 @@ struct test_case {
            /* .mem_base = */ NULL,
            /* .no_alloc = */ true,
        };
+        const bool use_weights = use_weight_context();
+
        ggml_context * ctx = ggml_init(params);
        GGML_ASSERT(ctx);
+        ggml_context * ctx_weights = use_weights ? ggml_init(params) : nullptr;
+        GGML_ASSERT(!use_weights || ctx_weights);

        gf = ggml_new_graph(ctx);

        // pre-graph sentinel
        add_sentinel(ctx);
+        if (ctx_weights) {
+            add_sentinel(ctx_weights);
+        }

-        ggml_tensor * out = build_graph(ctx);
+        ggml_tensor * out = build_graph(ctx, ctx_weights);
        current_op_name   = op_desc(out);
        check_for_f16_tensor(ctx);

        if (!matches_filter(out, op_names_filter)) {
            //printf("  %s: skipping\n", op_desc(out).c_str());
+            ggml_free(ctx_weights);
            ggml_free(ctx);
            return test_status_t::SKIPPED;
        }
@@ -1355,18 +1368,36 @@ struct test_case {

            print_test_result_locked(output_printer, result);

+            ggml_free(ctx_weights);
            ggml_free(ctx);
            return test_status_t::NOT_SUPPORTED;
        }

        // post-graph sentinel
        add_sentinel(ctx);
+        if (ctx_weights) {
+            add_sentinel(ctx_weights);
+        }
+
+        ggml_backend_buffer_t buf_weights = nullptr;
+        if (ctx_weights) {
+            buf_weights = ggml_backend_alloc_ctx_tensors(ctx_weights, backend1);
+            if (buf_weights == NULL) {
+                printf("failed to allocate weight tensors [%s] ", ggml_backend_name(backend1));
+                ggml_free(ctx_weights);
+                ggml_free(ctx);
+                return test_status_t::FAIL;
+            }
+            ggml_backend_buffer_set_usage(buf_weights, GGML_BACKEND_BUFFER_USAGE_WEIGHTS);
+        }

        // allocate
        ggml_backend_buffer_t buf = ggml_backend_alloc_ctx_tensors(ctx, backend1);

        if (buf == NULL) {
            printf("failed to allocate tensors [%s] ", ggml_backend_name(backend1));
+            ggml_backend_buffer_free(buf_weights);
+            ggml_free(ctx_weights);
            ggml_free(ctx);
            return test_status_t::FAIL;
        }
@@ -1381,6 +1412,9 @@ struct test_case {

        // randomize tensors
        initialize_tensors(ctx);
+        if (ctx_weights) {
+            initialize_tensors(ctx_weights);
+        }

        // compare
        struct callback_userdata {
@@ -1466,7 +1500,8 @@ struct test_case {
                                                               fused_nodes_to_verify.size());

        ggml_backend_buffer_free(buf);
-
+        ggml_backend_buffer_free(buf_weights);
+        ggml_free(ctx_weights);
        ggml_free(ctx);

        // Create test result
@@ -1490,10 +1525,14 @@ struct test_case {
            /* .mem_base = */ NULL,
            /* .no_alloc = */ true,
        };
+        const bool use_weights = use_weight_context();
+
        ggml_context_ptr ctx(ggml_init(params)); // smart ptr
        GGML_ASSERT(ctx);
+        ggml_context_ptr ctx_weights(use_weights ? ggml_init(params) : nullptr);
+        GGML_ASSERT(!use_weights || ctx_weights);

-        ggml_tensor * out             = build_graph(ctx.get());
+        ggml_tensor * out             = build_graph(ctx.get(), ctx_weights.get());
        current_op_name               = op_desc(out);
        if (!matches_filter(out, op_names_filter)) {
            //printf("  %s: skipping\n", op_desc(out).c_str());
@@ -1510,6 +1549,16 @@ struct test_case {
            return true;
        }

+        ggml_backend_buffer_ptr buf_weights(nullptr);
+        if (ctx_weights) {
+            buf_weights.reset(ggml_backend_alloc_ctx_tensors(ctx_weights.get(), backend));
+            if (buf_weights == NULL) {
+                printf("failed to allocate weight tensors\n");
+                return false;
+            }
+            ggml_backend_buffer_set_usage(buf_weights.get(), GGML_BACKEND_BUFFER_USAGE_WEIGHTS);
+        }
+
        // allocate
        ggml_backend_buffer_ptr buf(ggml_backend_alloc_ctx_tensors(ctx.get(), backend)); // smart ptr

@@ -1520,6 +1569,9 @@ struct test_case {

        // randomize tensors
        initialize_tensors(ctx.get());
+        if (ctx_weights) {
+            initialize_tensors(ctx_weights.get());
+        }

        // build graph
        ggml_cgraph * gf = ggml_new_graph_custom(ctx.get(), graph_nodes, false);
@@ -5848,19 +5900,21 @@ struct test_mul_mat_vec_fusion : public test_case {
    const bool b;        // broadcast b matrix (only for use_id)
    const bool with_bias;
    const bool with_gate;
+    const bool with_lane_scale;
    std::array<int64_t, 2> batch_dims;

    test_mul_mat_vec_fusion(ggml_type type, ggml_glu_op op, int64_t m, int64_t n, int64_t k,
                        bool use_id = false, int n_mats = 1, int n_used = 1, bool b = false, bool with_bias = false, bool with_gate = true,
-                        std::array<int64_t, 2> batch_dims = {4, 2})
-    : type(type), glu_op(op), m(m), n(n), k(k), use_id(use_id), n_mats(n_mats), n_used(n_used), b(b), with_bias(with_bias), with_gate(with_gate), batch_dims(batch_dims) {
+                        bool with_lane_scale = false, std::array<int64_t, 2> batch_dims = {4, 2})
+    : type(type), glu_op(op), m(m), n(n), k(k), use_id(use_id), n_mats(n_mats), n_used(n_used), b(b), with_bias(with_bias),
+        with_gate(with_gate), with_lane_scale(with_lane_scale), batch_dims(batch_dims) {
        if (use_id) {
            GGML_ASSERT(n_used <= n_mats);
        }
    }

    std::string vars() override {
-        return VARS_TO_STR12(type, glu_op, m, n, k, use_id, n_mats, n_used, b, with_bias, with_gate, batch_dims);
+        return VARS_TO_STR13(type, glu_op, m, n, k, use_id, n_mats, n_used, b, with_bias, with_gate, with_lane_scale, batch_dims);
    }

    std::string op_desc(ggml_tensor * t) override {
@@ -5869,6 +5923,7 @@ struct test_mul_mat_vec_fusion : public test_case {
    }

    bool run_whole_graph() override { return true; }
+    bool use_weight_context() override { return use_id && with_lane_scale; }

    ggml_tensor * build_gate(ggml_context * ctx, ggml_tensor * ffn_gate, ggml_tensor * ffn_up) {
        ggml_tensor * out = nullptr;
@@ -5884,7 +5939,26 @@ struct test_mul_mat_vec_fusion : public test_case {
        return out;
    }

+    ggml_tensor * build_lane_scale_dense(ggml_context * ctx, ggml_tensor * out) {
+        ggml_tensor * scale = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, 1);
+        return ggml_mul(ctx, out, scale);
+    }
+
+    ggml_tensor * build_lane_scale_id(ggml_context * ctx, ggml_context * ctx_weights, ggml_tensor * out, ggml_tensor * ids) {
+        GGML_ASSERT(ctx_weights);
+        ggml_tensor * scale = ggml_new_tensor_1d(ctx_weights, GGML_TYPE_F32, n_mats);
+        ggml_tensor * s = ggml_reshape_3d(ctx, scale, 1, n_mats, 1);
+        s = ggml_repeat_4d(ctx, s, 1, n_mats, m, 1);
+        s = ggml_get_rows(ctx, s, ids);
+        return ggml_mul(ctx, out, s);
+    }
+
    ggml_tensor * build_graph(ggml_context * ctx) override {
+        GGML_ASSERT(!use_weight_context());
+        return build_graph(ctx, nullptr);
+    }
+
+    ggml_tensor * build_graph(ggml_context * ctx, ggml_context * ctx_weights) override {
        if (!use_id) {
            const int              channels = batch_dims[0];
            const int              samples  = batch_dims[1];
@@ -5895,19 +5969,34 @@ struct test_mul_mat_vec_fusion : public test_case {
            ggml_tensor * gate = with_gate ? ggml_new_tensor(ctx, type, 4, ne0.data()) : nullptr;
            ggml_tensor * up   = ggml_new_tensor(ctx, type, 4, ne0.data());

-            ggml_tensor * ffn_up = ggml_mul_mat(ctx, up, cur);
-            if (with_bias) {
-                std::array<int64_t, 4> bias_ne = { ffn_up->ne[0], 1, channels, samples };
-                ggml_tensor * up_bias = ggml_new_tensor(ctx, GGML_TYPE_F32, 4, bias_ne.data());
-                ffn_up = ggml_add(ctx, ffn_up, up_bias);
-            }
+            auto build_lane_up = [&]() {
+                ggml_tensor * ffn_up = ggml_mul_mat(ctx, up, cur);
+                if (with_lane_scale) {
+                    ffn_up = build_lane_scale_dense(ctx, ffn_up);
+                }
+                if (with_bias) {
+                    std::array<int64_t, 4> bias_ne = { ffn_up->ne[0], 1, channels, samples };
+                    ggml_tensor * up_bias = ggml_new_tensor(ctx, GGML_TYPE_F32, 4, bias_ne.data());
+                    ffn_up = ggml_add(ctx, ffn_up, up_bias);
+                }
+                return ffn_up;
+            };

-            ggml_tensor * ffn_gate = with_gate ? ggml_mul_mat(ctx, gate, cur) : nullptr;
-            if (with_bias && with_gate) {
-                std::array<int64_t, 4> bias_ne   = { ffn_gate->ne[0], 1, channels, samples };
-                ggml_tensor * gate_bias = ggml_new_tensor(ctx, GGML_TYPE_F32, 4, bias_ne.data());
-                ffn_gate = ggml_add(ctx, ffn_gate, gate_bias);
-            }
+            auto build_lane_gate = [&]() {
+                ggml_tensor * ffn_gate = ggml_mul_mat(ctx, gate, cur);
+                if (with_lane_scale) {
+                    ffn_gate = build_lane_scale_dense(ctx, ffn_gate);
+                }
+                if (with_bias) {
+                    std::array<int64_t, 4> bias_ne   = { ffn_gate->ne[0], 1, channels, samples };
+                    ggml_tensor * gate_bias = ggml_new_tensor(ctx, GGML_TYPE_F32, 4, bias_ne.data());
+                    ffn_gate = ggml_add(ctx, ffn_gate, gate_bias);
+                }
+                return ffn_gate;
+            };
+
+            ggml_tensor * ffn_up = build_lane_up();
+            ggml_tensor * ffn_gate = with_gate ? build_lane_gate() : nullptr;

            ggml_tensor * out = with_gate ? build_gate(ctx, ffn_gate, ffn_up) : ffn_up;

@@ -5929,17 +6018,32 @@ struct test_mul_mat_vec_fusion : public test_case {
            ggml_tensor * cur = ggml_new_tensor_3d(ctx, GGML_TYPE_F32, k, this->b ? 1 : n_used, m);
            ggml_set_name(cur, "cur");

-            ggml_tensor * ffn_up = ggml_mul_mat_id(ctx, ups, cur, ids);
-            if (with_bias) {
-                ggml_tensor * up_bias_param = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, ffn_up->ne[0], n_mats);
-                ffn_up = ggml_add_id(ctx, ffn_up, up_bias_param, ids);
-            }
+            auto build_lane_up = [&]() {
+                ggml_tensor * ffn_up = ggml_mul_mat_id(ctx, ups, cur, ids);
+                if (with_lane_scale) {
+                    ffn_up = build_lane_scale_id(ctx, ctx_weights, ffn_up, ids);
+                }
+                if (with_bias) {
+                    ggml_tensor * up_bias_param = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, ffn_up->ne[0], n_mats);
+                    ffn_up = ggml_add_id(ctx, ffn_up, up_bias_param, ids);
+                }
+                return ffn_up;
+            };

-            ggml_tensor * ffn_gate = with_gate? ggml_mul_mat_id(ctx, gates, cur, ids) : nullptr;
-            if (with_bias && with_gate) {
-                ggml_tensor * gate_bias_param = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, ffn_gate->ne[0], n_mats);
-                ffn_gate = ggml_add_id(ctx, ffn_gate, gate_bias_param, ids);
-            }
+            auto build_lane_gate = [&]() {
+                ggml_tensor * ffn_gate = ggml_mul_mat_id(ctx, gates, cur, ids);
+                if (with_lane_scale) {
+                    ffn_gate = build_lane_scale_id(ctx, ctx_weights, ffn_gate, ids);
+                }
+                if (with_bias) {
+                    ggml_tensor * gate_bias_param = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, ffn_gate->ne[0], n_mats);
+                    ffn_gate = ggml_add_id(ctx, ffn_gate, gate_bias_param, ids);
+                }
+                return ffn_gate;
+            };
+
+            ggml_tensor * ffn_up = build_lane_up();
+            ggml_tensor * ffn_gate = with_gate ? build_lane_gate() : nullptr;

            ggml_tensor * out = with_gate ? build_gate(ctx, ffn_gate, ffn_up) : ffn_up;

@@ -9202,10 +9306,15 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
                            if (!with_gate && glu_op != GGML_GLU_OP_SWIGLU) {
                                continue;
                            }
-                            test_cases.emplace_back(new test_mul_mat_vec_fusion(type, glu_op, 1, 32, 256,
-                                use_id, 16, 8, b, with_bias, with_gate));
-                            test_cases.emplace_back(new test_mul_mat_vec_fusion(type, glu_op, 1, 32, 256,
-                                use_id, 16, 8, b, with_bias, with_gate, {1, 1}));
+                            for (bool with_lane_scale : {false, true}) {
+                                if (with_lane_scale && type != GGML_TYPE_NVFP4) {
+                                    continue;
+                                }
+                                test_cases.emplace_back(new test_mul_mat_vec_fusion(type, glu_op, 1, 32, 256,
+                                    use_id, 16, 8, b, with_bias, with_gate, with_lane_scale));
+                                test_cases.emplace_back(new test_mul_mat_vec_fusion(type, glu_op, 1, 32, 256,
+                                    use_id, 16, 8, b, with_bias, with_gate, with_lane_scale, {1, 1}));
+                            }
                        }
                    }
                }
@@ -9823,6 +9932,13 @@ static bool test_backend(ggml_backend_t backend, ggml_backend_dev_t dev, test_mo
    }

    if (mode == MODE_GRAD) {
+        test_cases.erase(
+            std::remove_if(test_cases.begin(), test_cases.end(), [](const std::unique_ptr<test_case> & tc) {
+                return tc->run_whole_graph();
+            }),
+            test_cases.end()
+        );
+
        size_t n_ok = 0;
        for (auto & test : test_cases) {
            if (test->eval_grad(backend, op_names_filter, output_printer)) {
@@ -158,6 +158,7 @@ static int test_vec_dot_q(bool verbose) {
                type == GGML_TYPE_Q1_0    ? MAX_QUANTIZATION_TOTAL_ERROR_BINARY :
                type == GGML_TYPE_TQ1_0   ? MAX_QUANTIZATION_TOTAL_ERROR_TERNARY :
                type == GGML_TYPE_TQ2_0   ? MAX_QUANTIZATION_TOTAL_ERROR_TERNARY :
+                type == GGML_TYPE_Q2_0    ? MAX_QUANTIZATION_TOTAL_ERROR_TERNARY :
                type == GGML_TYPE_Q2_K    ? MAX_QUANTIZATION_TOTAL_ERROR_2BITS :
                type == GGML_TYPE_IQ2_S   ? MAX_QUANTIZATION_TOTAL_ERROR_2BITS :
                type == GGML_TYPE_Q3_K    ? MAX_QUANTIZATION_TOTAL_ERROR_3BITS :
@@ -183,7 +184,7 @@ static int test_vec_dot_q(bool verbose) {
                ? MAX_DOT_PRODUCT_ERROR_LOWBIT
                : type == GGML_TYPE_Q1_0
                ? MAX_DOT_PRODUCT_ERROR_BINARY
-                : type == GGML_TYPE_TQ1_0 || type == GGML_TYPE_TQ2_0
+                : type == GGML_TYPE_TQ1_0 || type == GGML_TYPE_TQ2_0 || type == GGML_TYPE_Q2_0
                ? MAX_DOT_PRODUCT_ERROR_TERNARY
                : type == GGML_TYPE_NVFP4
                ? MAX_DOT_PRODUCT_ERROR_FP4
@@ -33,6 +33,7 @@ struct quant_option {

 static const std::vector<quant_option> QUANT_OPTIONS = {
    { "Q1_0",     LLAMA_FTYPE_MOSTLY_Q1_0,     " 1.125 bpw quantization",           },
+    { "Q2_0",     LLAMA_FTYPE_MOSTLY_Q2_0,     " 2.25 bpw quantization (group 64)",  },
    { "Q4_0",     LLAMA_FTYPE_MOSTLY_Q4_0,     " 4.34G, +0.4685 ppl @ Llama-3-8B",  },
    { "Q4_1",     LLAMA_FTYPE_MOSTLY_Q4_1,     " 4.78G, +0.4511 ppl @ Llama-3-8B",  },
    { "MXFP4_MOE",LLAMA_FTYPE_MOSTLY_MXFP4_MOE," MXFP4 MoE",  },
@@ -897,8 +897,10 @@ private:

    server_batch batch;

-    llama_model_ptr model_dft;
-    llama_context_ptr ctx_dft;
+    llama_model   * model_dft = nullptr;
+    llama_context * ctx_dft   = nullptr;
+
+    common_speculative_init_result_ptr spec_init;

    common_context_seq_rm_type ctx_tgt_seq_rm_type = COMMON_CONTEXT_SEQ_RM_TYPE_NO;
    common_context_seq_rm_type ctx_dft_seq_rm_type = COMMON_CONTEXT_SEQ_RM_TYPE_NO;
@@ -939,8 +941,10 @@ private:

    void destroy() {
        spec.reset();
-        ctx_dft.reset();
-        model_dft.reset();
+        spec_init.reset();
+
+        ctx_dft   = nullptr;
+        model_dft = nullptr;

        llama_init.reset();

@@ -1084,30 +1088,15 @@ private:
        // optionally reserve VRAM for the draft / MTP context before fitting the target model
        if (params_base.fit_params) {
            if (has_spec) {
-                common_params params_dft = params_base;
-                bool measure_model_bytes = true;
+                // MTP draft context lives on the target model, only context+compute are new
+                bool measure_model_bytes = has_draft;

-                if (has_draft) {
-                    const auto & params_spec = params_base.speculative.draft;
-                    params_dft.devices               = params_spec.devices;
-                    params_dft.model                 = params_spec.mparams;
-                    params_dft.n_gpu_layers          = params_spec.n_gpu_layers;
-                    params_dft.cache_type_k          = params_spec.cache_type_k;
-                    params_dft.cache_type_v          = params_spec.cache_type_v;
-                    params_dft.tensor_buft_overrides = params_spec.tensor_buft_overrides;
-                } else {
-                    // MTP draft context lives on the target model, only context+compute are new
-                    measure_model_bytes = false;
-                }
-
-                params_dft.n_outputs_max = params_base.n_parallel;
+                common_params params_dft = common_base_params_to_speculative(params_base);

                auto mparams_dft = common_model_params_to_llama(params_dft);
                auto cparams_dft = common_context_params_to_llama(params_dft);
                if (spec_mtp) {
                    cparams_dft.ctx_type = LLAMA_CONTEXT_TYPE_MTP;
-                    cparams_dft.type_k   = params_base.speculative.draft.cache_type_k;
-                    cparams_dft.type_v   = params_base.speculative.draft.cache_type_v;
                }
                cparams_dft.n_rs_seq = 0;

@@ -1175,82 +1164,36 @@ private:

        add_bos_token = llama_vocab_get_add_bos(vocab);

-        if (has_draft) {
-            // TODO speculative: move to common/speculative.cpp?
-            const auto & params_spec = params_base.speculative.draft;
-
-            SRV_TRC("loading draft model '%s'\n", params_spec.mparams.path.c_str());
-
-            auto params_dft = params_base;
-
-            params_dft.devices      = params_spec.devices;
-            params_dft.model        = params_spec.mparams;
-            params_dft.n_gpu_layers = params_spec.n_gpu_layers;
-            params_dft.cache_type_k = params_spec.cache_type_k;
-            params_dft.cache_type_v = params_spec.cache_type_v;
-
-            if (params_spec.cpuparams.n_threads > 0) {
-                params_dft.cpuparams.n_threads       = params_spec.cpuparams.n_threads;
-                params_dft.cpuparams_batch.n_threads = params_spec.cpuparams_batch.n_threads;
-            }
-
-            params_dft.tensor_buft_overrides = params_spec.tensor_buft_overrides;
-
-            auto mparams_dft = common_model_params_to_llama(params_dft);
-
-            // progress callback
-            mparams_dft.progress_callback           = load_progress_callback;
-            mparams_dft.progress_callback_user_data = &load_progress_spec;
-
-            model_dft.reset(llama_model_load_from_file(params_dft.model.path.c_str(), mparams_dft));
-            if (model_dft == nullptr) {
-                SRV_ERR("failed to load draft model, '%s'\n", params_dft.model.path.c_str());
-                return false;
-            }
-
-            auto cparams = common_context_params_to_llama(params_dft);
-
-            if (spec_mtp) {
-                cparams.ctx_type = LLAMA_CONTEXT_TYPE_MTP;
-            }
-
-            // note: for small models maybe we can set this to the maximum possible draft from all speculative types
-            //       the extra memory for small models is likely negligible?
-            cparams.n_rs_seq  = 0;
-            cparams.ctx_other = ctx_tgt;
-
-            ctx_dft.reset(llama_init_from_model(model_dft.get(), cparams));
-            if (ctx_dft == nullptr) {
-                SRV_ERR("%s", "failed to create draft context\n");
-                return false;
-            }
-
-            params_base.speculative.draft.ctx_tgt = ctx_tgt;
-            params_base.speculative.draft.ctx_dft = ctx_dft.get();
-        } else if (spec_mtp) {
-            // no new model load, so we simply report 0.0 and 1.0 progress
+        if (has_spec) {
+            // spec_mtp doesn't use load a model internally, so we report 0.0 and 1.0 manually
            load_progress_callback(0.0f, &load_progress_spec);
+            load_progress_spec.t_last_load_progress_ms = 0;  // reset so internal cbs aren't delayed

-            SRV_TRC("creating MTP draft context against the target model '%s'\n",
-                    params_base.model.path.c_str());
+            {
+                common_params params_dft = common_base_params_to_speculative(params_base);

-            auto cparams_mtp = common_context_params_to_llama(params_base);
-            cparams_mtp.ctx_type      = LLAMA_CONTEXT_TYPE_MTP;
-            cparams_mtp.type_k        = params_base.speculative.draft.cache_type_k;
-            cparams_mtp.type_v        = params_base.speculative.draft.cache_type_v;
-            cparams_mtp.n_rs_seq      = 0;
-            cparams_mtp.n_outputs_max = params_base.n_parallel;
-            cparams_mtp.ctx_other     = ctx_tgt;
+                // progress callback
+                params_dft.load_progress_callback           = load_progress_callback;
+                params_dft.load_progress_callback_user_data = &load_progress_spec;

-            ctx_dft.reset(llama_init_from_model(model_tgt, cparams_mtp));
-            if (ctx_dft == nullptr) {
-                SRV_ERR("%s", "failed to create MTP context\n");
-                return false;
+                spec_init = common_speculative_init_from_params(params_dft, model_tgt, ctx_tgt);
+                model_dft = spec_init->model();
+                ctx_dft   = spec_init->context();
+
+                if (has_draft && model_dft == nullptr) {
+                    SRV_ERR("failed to load draft model, '%s'\n", params_dft.model.path.c_str());
+                    return false;
+                }
+
+                if (ctx_dft == nullptr) {
+                    SRV_ERR("%s", "failed to create MTP context\n");
+                    return false;
+                }
+
+                params_base.speculative.draft.ctx_tgt = ctx_tgt;
+                params_base.speculative.draft.ctx_dft = ctx_dft;
            }

-            params_base.speculative.draft.ctx_tgt = ctx_tgt;
-            params_base.speculative.draft.ctx_dft = ctx_dft.get();
-
            load_progress_callback(1.0f, &load_progress_spec);
        }

@@ -1343,13 +1286,15 @@ private:
        }

        if (ctx_dft) {
-            ctx_dft_seq_rm_type = common_context_can_seq_rm(ctx_dft.get());
+            ctx_dft_seq_rm_type = common_context_can_seq_rm(ctx_dft);
        }

        if (spec) {
            SRV_TRC("%s", "speculative decoding context initialized\n");
        } else {
-            ctx_dft.reset();
+            spec_init.reset();
+            ctx_dft   = nullptr;
+            model_dft = nullptr;
        }

        for (int i = 0; i < params_base.n_parallel; i++) {
@@ -1357,7 +1302,7 @@ private:

            slot.id      = i;
            slot.ctx_tgt = ctx_tgt;
-            slot.ctx_dft = ctx_dft.get();
+            slot.ctx_dft = ctx_dft;
            slot.spec    = spec.get();
            slot.n_ctx   = n_ctx_slot;

@@ -2362,8 +2307,8 @@ private:
        //       this is not true for SWA models: https://github.com/ggml-org/llama.cpp/pull/24411#issuecomment-4677983225
        cur.update_pos(slot.prompt.n_tokens() - n_tokens_cur, pos_min, pos_max);

-        cur.update_tgt(ctx_tgt,       slot.id, LLAMA_STATE_SEQ_FLAGS_PARTIAL_ONLY);
-        cur.update_dft(ctx_dft.get(), slot.id, LLAMA_STATE_SEQ_FLAGS_PARTIAL_ONLY);
+        cur.update_tgt(ctx_tgt, slot.id, LLAMA_STATE_SEQ_FLAGS_PARTIAL_ONLY);
+        cur.update_dft(ctx_dft, slot.id, LLAMA_STATE_SEQ_FLAGS_PARTIAL_ONLY);
        // stash the draft's speculative state with the checkpoint
        common_speculative_get_state(spec.get(), slot.id, cur.data_spec);

@@ -2899,8 +2844,8 @@ private:
                common_context_seq_add(ctx_tgt, slot.id, n_keep + n_discard, slot.prompt.n_tokens(), -n_discard);

                if (ctx_dft) {
-                    common_context_seq_rm (ctx_dft.get(), slot.id, n_keep            , n_keep + n_discard);
-                    common_context_seq_add(ctx_dft.get(), slot.id, n_keep + n_discard, slot.prompt.tokens.pos_next(), -n_discard);
+                    common_context_seq_rm (ctx_dft, slot.id, n_keep            , n_keep + n_discard);
+                    common_context_seq_add(ctx_dft, slot.id, n_keep + n_discard, slot.prompt.tokens.pos_next(), -n_discard);
                }

                // add generated tokens to cache
@@ -2972,7 +2917,7 @@ private:
                                llama_memory_seq_pos_max(llama_get_memory(ctx_tgt), slot.id));

                        if (use_ckpt_dft) {
-                            slot.spec_ckpt.update_dft(ctx_dft.get(), slot.id, LLAMA_STATE_SEQ_FLAGS_PARTIAL_ONLY);
+                            slot.spec_ckpt.update_dft(ctx_dft, slot.id, LLAMA_STATE_SEQ_FLAGS_PARTIAL_ONLY);
                        }

                        slot.spec_prompt = slot.prompt.tokens.get_text_tokens();
@@ -3009,10 +2954,10 @@ private:

            if (ctx_dft) {
                if (use_ckpt_dft) {
-                    ckpt.load_dft(ctx_dft.get(), slot.id, LLAMA_STATE_SEQ_FLAGS_PARTIAL_ONLY);
+                    ckpt.load_dft(ctx_dft, slot.id, LLAMA_STATE_SEQ_FLAGS_PARTIAL_ONLY);
                }

-                common_context_seq_rm(ctx_dft.get(), slot.id, ckpt.pos_max + 1, -1);
+                common_context_seq_rm(ctx_dft, slot.id, ckpt.pos_max + 1, -1);
            }

            if (!draft.empty()) {
@@ -3021,7 +2966,7 @@ private:
                   (ctx_tgt_seq_rm_type == COMMON_CONTEXT_SEQ_RM_TYPE_RS && draft.size() > llama_n_rs_seq(ctx_tgt));

                const bool use_ckpt_dft =
-                   (ctx_dft_seq_rm_type == COMMON_CONTEXT_SEQ_RM_TYPE_RS && draft.size() > llama_n_rs_seq(ctx_dft.get()));
+                   (ctx_dft_seq_rm_type == COMMON_CONTEXT_SEQ_RM_TYPE_RS && draft.size() > llama_n_rs_seq(ctx_dft));

                if (use_ckpt_tgt) {
                    //const int64_t t_start = ggml_time_us();
@@ -3038,7 +2983,7 @@ private:
                }

                if (use_ckpt_dft) {
-                    ckpt.update_dft(ctx_dft.get(), slot.id, LLAMA_STATE_SEQ_FLAGS_PARTIAL_ONLY);
+                    ckpt.update_dft(ctx_dft, slot.id, LLAMA_STATE_SEQ_FLAGS_PARTIAL_ONLY);
                }
            }
        });
@@ -3219,8 +3164,8 @@ private:
                                            common_context_seq_add(ctx_tgt, slot.id, head_c, head_c + n_match, kv_shift);

                                            if (ctx_dft) {
-                                                common_context_seq_rm (ctx_dft.get(), slot.id, head_p, head_c);
-                                                common_context_seq_add(ctx_dft.get(), slot.id, head_c, head_c + n_match, kv_shift);
+                                                common_context_seq_rm (ctx_dft, slot.id, head_p, head_c);
+                                                common_context_seq_add(ctx_dft, slot.id, head_c, head_c + n_match, kv_shift);
                                            }

                                            for (size_t i = 0; i < n_match; i++) {
@@ -3320,8 +3265,8 @@ private:

                                    if (!do_reset) {
                                        // restore the context checkpoint
-                                        it->load_tgt(ctx_tgt,       slot.id, LLAMA_STATE_SEQ_FLAGS_PARTIAL_ONLY);
-                                        it->load_dft(ctx_dft.get(), slot.id, LLAMA_STATE_SEQ_FLAGS_PARTIAL_ONLY);
+                                        it->load_tgt(ctx_tgt, slot.id, LLAMA_STATE_SEQ_FLAGS_PARTIAL_ONLY);
+                                        it->load_dft(ctx_dft, slot.id, LLAMA_STATE_SEQ_FLAGS_PARTIAL_ONLY);
                                        // restore the draft's speculative state
                                        common_speculative_set_state(spec.get(), slot.id, it->data_spec);

@@ -3395,7 +3340,7 @@ private:

                    common_context_seq_rm(ctx_tgt, slot.id, p0, -1);
                    if (ctx_dft) {
-                        common_context_seq_rm(ctx_dft.get(), slot.id, p0, -1);
+                        common_context_seq_rm(ctx_dft, slot.id, p0, -1);
                    }

                    // If using an alora, there may be uncached tokens that come
@@ -730,6 +730,10 @@ json server_task_result_cmpl_final::to_json_oaicompat_resp_stream() {
        }}
    });

+    if (timings.prompt_n >= 0) {
+        server_sent_events.back().at("data").push_back({"timings", timings.to_json()});
+    }
+
    return server_sent_events;
 }

@@ -1016,6 +1020,7 @@ void server_task_result_cmpl_partial::update(task_result_state & state) {
    thinking_block_started = state.thinking_block_started;
    text_block_started     = state.text_block_started;

+    oai_resp_created       = state.oai_resp_created;
    oai_resp_id            = state.oai_resp_id;
    oai_resp_reasoning_id  = state.oai_resp_reasoning_id;
    oai_resp_message_id    = state.oai_resp_message_id;
@@ -1024,6 +1029,10 @@ void server_task_result_cmpl_partial::update(task_result_state & state) {
    // track if the accumulated message has any reasoning content
    anthropic_has_reasoning = !state.chat_msg.reasoning_content.empty();

+    if (res_type == TASK_RESPONSE_TYPE_OAI_RESP && !state.oai_resp_created && (is_progress || n_decoded == 1)) {
+        state.oai_resp_created = true;
+    }
+
    // Pre-compute state updates based on diffs (for next chunk)
    for (const common_chat_msg_diff & diff : oaicompat_msg_diffs) {
        if (!diff.reasoning_content_delta.empty() && !state.thinking_block_started) {
@@ -1181,7 +1190,7 @@ json server_task_result_cmpl_partial::to_json_oaicompat_chat() {
 json server_task_result_cmpl_partial::to_json_oaicompat_resp() {
    std::vector<json> events;

-    if (n_decoded == 1) {
+    if (!oai_resp_created) {
        events.push_back(json {
            {"event", "response.created"},
            {"data", json {
@@ -1204,6 +1213,18 @@ json server_task_result_cmpl_partial::to_json_oaicompat_resp() {
                }},
            }},
        });
+    } else if (is_progress) {
+        events.push_back(json {
+            {"event", "response.in_progress"},
+            {"data", json {
+                {"type", "response.in_progress"},
+                {"response", json {
+                    {"id",     oai_resp_id},
+                    {"object", "response"},
+                    {"status", "in_progress"},
+                }},
+            }},
+        });
    }

    for (const common_chat_msg_diff & diff : oaicompat_msg_diffs) {
@@ -1302,6 +1323,17 @@ json server_task_result_cmpl_partial::to_json_oaicompat_resp() {
            });
        }
    }
+
+    if (!events.empty()) {
+        json & data = events.back().at("data");
+        if (timings.prompt_n >= 0) {
+            data.push_back({"timings", timings.to_json()});
+        }
+        if (is_progress) {
+            data.push_back({"prompt_progress", progress.to_json()});
+        }
+    }
+
    return events;
 }

@@ -1631,7 +1663,22 @@ server_prompt * server_prompt_cache::alloc(const server_prompt & prompt, size_t
        }
    }

-    // next, remove any cached prompts that are fully contained in the current prompt
+    // calculate checkpoints size to see if it will fit with the prompt
+    size_t checkpoints_size = 0;
+    for (const auto & ckpt : prompt.checkpoints) {
+        checkpoints_size += ckpt.size();
+    }
+
+    const size_t state_size_new = state_size_tgt + state_size_dft + checkpoints_size;
+
+    // skip over-limit entries to avoid disturbing the cache
+    if (limit_size > 0 && state_size_new > limit_size) {
+        SRV_WRN(" - prompt state size %.3f MiB exceeds cache size limit %.3f MiB, skipping\n",
+                state_size_new / (1024.0 * 1024.0), limit_size / (1024.0 * 1024.0));
+        return nullptr;
+    }
+
+    // remove any cached prompts that are fully contained in the current prompt
    for (auto it = states.begin(); it != states.end();) {
        const int len = it->tokens.get_common_prefix(prompt.tokens);

@@ -1644,6 +1691,16 @@ server_prompt * server_prompt_cache::alloc(const server_prompt & prompt, size_t
        }
    }

+    if (limit_size > 0) {
+        // make room before allocating the new vectors to avoid breaching the limit
+        while (!states.empty() && size() + state_size_new > limit_size) {
+            SRV_WRN(" - making room for prompt cache entry, removing oldest entry (size = %.3f MiB)\n",
+                    states.front().size() / (1024.0 * 1024.0));
+
+            states.pop_front();
+        }
+    }
+
    std::vector<uint8_t> state_data_tgt;
    std::vector<uint8_t> state_data_dft;

@@ -1752,12 +1809,7 @@ bool server_prompt_cache::load(server_prompt & prompt, const server_tokens & tok

 void server_prompt_cache::update() {
    if (limit_size > 0) {
-        // always keep at least one state, regardless of the limits
-        while (states.size() > 1 && size() > limit_size) {
-            if (states.empty()) {
-                break;
-            }
-
+        while (!states.empty() && size() > limit_size) {
            SRV_WRN(" - cache size limit reached, removing oldest entry (size = %.3f MiB)\n", states.front().size() / (1024.0 * 1024.0));

            states.pop_front();
@@ -1771,11 +1823,7 @@ void server_prompt_cache::update() {
    const size_t limit_tokens_cur = limit_size > 0 ? std::max<size_t>(limit_tokens, limit_size/size_per_token) : limit_tokens;

    if (limit_tokens > 0) {
-        while (states.size() > 1 && n_tokens() > limit_tokens_cur) {
-            if (states.empty()) {
-                break;
-            }
-
+        while (!states.empty() && n_tokens() > limit_tokens_cur) {
            SRV_WRN(" - cache token limit (%zu, est: %zu) reached, removing oldest entry (size = %.3f MiB)\n",
                    limit_tokens, limit_tokens_cur, states.front().size() / (1024.0 * 1024.0));

@@ -117,6 +117,7 @@ struct task_result_state {
    bool text_block_started = false;

    // for OpenAI Responses streaming API
+    bool oai_resp_created = false;
    const std::string oai_resp_id;
    const std::string oai_resp_reasoning_id;
    const std::string oai_resp_message_id;
@@ -440,6 +441,7 @@ struct server_task_result_cmpl_partial : server_task_result {
    bool text_block_started     = false;

    // for OpenAI Responses API
+    bool oai_resp_created = false;
    std::string oai_resp_id;
    std::string oai_resp_reasoning_id;
    std::string oai_resp_message_id;
@@ -71,3 +71,44 @@ def test_responses_stream_with_openai_library():
            assert r.response.output[0].id.startswith("msg_")
            assert gathered_text == r.response.output_text
            assert match_regex("(Suddenly)+", r.response.output_text)
+
+
+def test_responses_stream_with_llama_telemetry():
+    global server
+    server.n_ctx = 256
+    server.n_batch = 32
+    server.n_slots = 1
+    server.start()
+
+    saw_progress = False
+    saw_delta_timings = False
+    completed = None
+
+    res = server.make_stream_request("POST", "/responses", data={
+        "input": "This is a test" * 10,
+        "max_output_tokens": 8,
+        "temperature": 0.8,
+        "stream": True,
+        "timings_per_token": True,
+        "return_progress": True,
+    })
+
+    for data in res:
+        if "prompt_progress" in data:
+            assert data["type"] == "response.in_progress"
+            assert data["prompt_progress"]["total"] > 0
+            assert data["prompt_progress"]["processed"] >= data["prompt_progress"]["cache"]
+            saw_progress = True
+        if "timings" in data:
+            assert "prompt_per_second" in data["timings"]
+            assert "predicted_per_second" in data["timings"]
+            if data["type"] == "response.output_text.delta":
+                saw_delta_timings = True
+        if data["type"] == "response.completed":
+            completed = data
+
+    assert saw_progress
+    assert saw_delta_timings
+    assert completed is not None
+    assert "usage" in completed["response"]
+    assert "timings" in completed
Author	SHA1	Message	Date
Pasha Khosravi	bec4772f6a	Add Q2_0 quantization: type definition and CPU backend (#24448 )	2026-07-07 12:05:47 -07:00
Georgi Gerganov	c198af4dc2	spec : fix naming, spacing (#25410 )	2026-07-07 18:52:30 +03:00
Oliver Simons	3899b39ce2	CUDA: Fuse MMVQ post-scale for NVFP4 (#24481 ) * CUDA: Fuse MMVQ for NVFP4 and BS 1 TODO: 1. Add tests to test-backend-ops (did verify correctness manually for one model) 2. Reorder bias/scale once PRs for NVFP4 are merged/landed * Add dense MMVQ fusion as well Perf numbers on B4500. Note qwen35 is FP8->Q8 + ./scripts/compare-llama-bench.py -b master -c osimons/nvfp4_fuse_mmvq --tool llama-bench -i llama-bench.sqlite \| Model \| Test \| t/s master \| t/s osimons/nvfp4_fuse_mmvq \| Speedup \| \|:-------------------------\|:-------------\|-------------:\|------------------------------:\|----------:\| \| qwen35moe 35B.A3B NVFP4 \| tg128@d32768 \| 150.15 \| 156.29 \| 1.04 \| \| qwen35moe 35B.A3B Q4_K_M \| tg128@d32768 \| 157.91 \| 157.64 \| 1.00 \| Perf numbers on DGX Spark + ./scripts/compare-llama-bench.py -b master -c osimons/nvfp4_fuse_mmvq --tool llama-bench -i llama-bench.sqlite \| Model \| Test \| t/s master \| t/s osimons/nvfp4_fuse_mmvq \| Speedup \| \|:-------------------------\|:-------------\|-------------:\|------------------------------:\|----------:\| \| qwen35moe 35B.A3B NVFP4 \| tg128@d32768 \| 58.31 \| 59.69 \| 1.02 \| \| qwen35moe 35B.A3B Q4_K_M \| tg128@d32768 \| 54.94 \| 54.79 \| 1.00 \| * Add tests for the added fusion ops * Cleanup test-backend-ops * Cleanup ggml-cuda/mmvq 1. Unrestrict post-scale fusion 2. Rename names accordingly 3. Remove env variable to disable fusion * Merge old mul_mat patterns into the lane-based approach * Enable fusion for MoE in shared MMVQ * Restrict scale_view_nodes, enroll MM + ADD into lane-matcher * Refactor mmvq loads, still does not help non-nvfp4 kernels * Restrict scale-fusion to NVFP4 This is necessary, as the prolog is quite heavy in GEMV for some quants/model configs, leading to net perf regression. We should really be looking to refactor this such that ratio of prologue/hot-loop/epilogue is better on the hot-loop front: + ./scripts/compare-llama-bench.py -b master -c `c1b9381d32` --tool llama-bench -i llama-bench.sqlite \| CPU \| Model \| Test \| t/s master \| t/s `c1b9381d3` \| Speedup \| \|:----------------------------\|:-------------------------\|:-------------\|-------------:\|----------------:\|----------:\| \| INTEL(R) XEON(R) GOLD 6542Y \| gemma4 26B.A4B NVFP4 \| tg128@d32768 \| 151.70 \| 154.32 \| 1.02 \| \| INTEL(R) XEON(R) GOLD 6542Y \| gemma4 26B.A4B Q4_K_M \| tg128@d32768 \| 187.95 \| 185.73 \| 0.99 \| \| INTEL(R) XEON(R) GOLD 6542Y \| gpt-oss 20B MXFP4 MoE \| tg128@d32768 \| 304.62 \| 300.69 \| 0.99 \| \| INTEL(R) XEON(R) GOLD 6542Y \| qwen35moe 35B.A3B NVFP4 \| tg128@d32768 \| 193.72 \| 211.99 \| 1.09 \| \| INTEL(R) XEON(R) GOLD 6542Y \| qwen35moe 35B.A3B Q4_K_M \| tg128@d32768 \| 217.76 \| 218.15 \| 1.00 * Reorder scale & bias-add to adhere to #24331 * Restrict lane scale to NVFP4 Don't need to test unfused combinations * Cleanup * Merge single-lane mm-fusion helpers * Refactor and clean-up host-side fusion logic * Move gate_bias and scale into the same active-thread guard Latest perf numbers: B6000 build: 5b7d9f272 (9578) + ./scripts/compare-llama-bench.py -b master -c osimons/nvfp4_fuse_mmvq --tool llama-bench -i llama-bench.sqlite \| CPU \| Model \| Test \| t/s master \| t/s osimons/nvfp4_fuse_mmvq \| Speedup \| \|:----------------------------\|:-------------------------\|:-------------\|-------------:\|------------------------------:\|----------:\| \| INTEL(R) XEON(R) GOLD 6542Y \| gemma4 26B.A4B NVFP4 \| tg128@d32768 \| 151.79 \| 154.10 \| 1.02 \| \| INTEL(R) XEON(R) GOLD 6542Y \| gemma4 26B.A4B Q4_K_M \| tg128@d32768 \| 187.90 \| 187.27 \| 1.00 \| \| INTEL(R) XEON(R) GOLD 6542Y \| gpt-oss 20B MXFP4 MoE \| tg128@d32768 \| 303.77 \| 306.56 \| 1.01 \| \| INTEL(R) XEON(R) GOLD 6542Y \| qwen35moe 35B.A3B NVFP4 \| tg128@d32768 \| 193.41 \| 207.99 \| 1.08 \| \| INTEL(R) XEON(R) GOLD 6542Y \| qwen35moe 35B.A3B Q4_K_M \| tg128@d32768 \| 217.60 \| 218.58 \| 1.00 \| DGX Spark build: 5b7d9f272 (9578) + ./scripts/compare-llama-bench.py -b master -c osimons/nvfp4_fuse_mmvq --tool llama-bench -i llama-bench.sqlite \| CPU \| Model \| Test \| t/s master \| t/s osimons/nvfp4_fuse_mmvq \| Speedup \| \|:------\|:-------------------------\|:-------------\|-------------:\|------------------------------:\|----------:\| \| CPU \| gemma4 26B.A4B NVFP4 \| tg128@d32768 \| 34.61 \| 34.84 \| 1.01 \| \| CPU \| gemma4 26B.A4B Q4_K_M \| tg128@d32768 \| 46.95 \| 46.90 \| 1.00 \| \| CPU \| gpt-oss 20B MXFP4 MoE \| tg128@d32768 \| 64.84 \| 64.62 \| 1.00 \| \| CPU \| qwen35moe 35B.A3B NVFP4 \| tg128@d32768 \| 59.63 \| 60.72 \| 1.02 \| \| CPU \| qwen35moe 35B.A3B Q4_K_M \| tg128@d32768 \| 56.53 \| 56.55 \| 1.00 \| PPL values for 5 chunks: this PR model mode ppl uncertainty log /mnt/share/gguf/unsloth/Qwen3.6-35B-A3B-GGUF/Qwen3.6-35B-A3B-UD-Q4_K_M.gguf fusion_enabled 5.2892 0.35389 ppl-value-checks/Qwen3.6-35B-A3B-UD-Q4_K_M.fusion_enabled.log /mnt/share/gguf/unsloth/Qwen3.6-35B-A3B-GGUF/Qwen3.6-35B-A3B-UD-Q4_K_M.gguf fusion_disabled 5.2742 0.35215 ppl-value-checks/Qwen3.6-35B-A3B-UD-Q4_K_M.fusion_disabled.log /mnt/share/gguf/nvidia/Qwen3.6-35B-A3B-2.06GB-per-token-CT/Qwen3.6-35B-A3B-2.06GB-per-token-CT_fp8_q8.gguf fusion_enabled 5.4487 0.36866 ppl-value-checks/Qwen3.6-35B-A3B-2.06GB-per-token-CT_fp8_q8.fusion_enabled.log /mnt/share/gguf/nvidia/Qwen3.6-35B-A3B-2.06GB-per-token-CT/Qwen3.6-35B-A3B-2.06GB-per-token-CT_fp8_q8.gguf fusion_disabled 5.4403 0.36782 ppl-value-checks/Qwen3.6-35B-A3B-2.06GB-per-token-CT_fp8_q8.fusion_disabled.log /mnt/share/gguf/nvidia/Gemma-4-26B-A4B-NVFP4/Gemma-4-26B-A4B-NVFP4_fp8_q8.gguf fusion_enabled 17342.4348 3703.13932 ppl-value-checks/Gemma-4-26B-A4B-NVFP4_fp8_q8.fusion_enabled.log /mnt/share/gguf/nvidia/Gemma-4-26B-A4B-NVFP4/Gemma-4-26B-A4B-NVFP4_fp8_q8.gguf fusion_disabled 18627.0624 3998.42475 ppl-value-checks/Gemma-4-26B-A4B-NVFP4_fp8_q8.fusion_disabled.log /mnt/share/gguf/ggml-org/gpt-oss-20b-GGUF/gpt-oss-20b-mxfp4.gguf fusion_enabled 363.8913 33.14007 ppl-value-checks/gpt-oss-20b-mxfp4.fusion_enabled.log /mnt/share/gguf/ggml-org/gpt-oss-20b-GGUF/gpt-oss-20b-mxfp4.gguf fusion_disabled 363.8913 33.14007 ppl-value-checks/gpt-oss-20b-mxfp4.fusion_disabled.log /mnt/share/gguf/unsloth/gemma-4-26B-A4B-it-GGUF/gemma-4-26B-A4B-it-UD-Q4_K_XL.gguf fusion_enabled 17330.3926 3716.70472 ppl-value-checks/gemma-4-26B-A4B-it-UD-Q4_K_XL.fusion_enabled.log /mnt/share/gguf/unsloth/gemma-4-26B-A4B-it-GGUF/gemma-4-26B-A4B-it-UD-Q4_K_XL.gguf fusion_disabled 17933.9524 3883.17066 ppl-value-checks/gemma-4-26B-A4B-it-UD-Q4_K_XL.fusion_disabled.log master: summary: ppl-value-checks/summary.tsv model mode ppl uncertainty log /mnt/share/gguf/unsloth/Qwen3.6-35B-A3B-GGUF/Qwen3.6-35B-A3B-UD-Q4_K_M.gguf fusion_enabled 5.2892 0.35389 ppl-value-checks/Qwen3.6-35B-A3B-UD-Q4_K_M.fusion_enabled.log /mnt/share/gguf/unsloth/Qwen3.6-35B-A3B-GGUF/Qwen3.6-35B-A3B-UD-Q4_K_M.gguf fusion_disabled 5.2742 0.35215 ppl-value-checks/Qwen3.6-35B-A3B-UD-Q4_K_M.fusion_disabled.log /mnt/share/gguf/nvidia/Qwen3.6-35B-A3B-2.06GB-per-token-CT/Qwen3.6-35B-A3B-2.06GB-per-token-CT_fp8_q8.gguf fusion_enabled 5.4487 0.36866 ppl-value-checks/Qwen3.6-35B-A3B-2.06GB-per-token-CT_fp8_q8.fusion_enabled.log /mnt/share/gguf/nvidia/Qwen3.6-35B-A3B-2.06GB-per-token-CT/Qwen3.6-35B-A3B-2.06GB-per-token-CT_fp8_q8.gguf fusion_disabled 5.4403 0.36782 ppl-value-checks/Qwen3.6-35B-A3B-2.06GB-per-token-CT_fp8_q8.fusion_disabled.log /mnt/share/gguf/nvidia/Gemma-4-26B-A4B-NVFP4/Gemma-4-26B-A4B-NVFP4_fp8_q8.gguf fusion_enabled 17342.4348 3703.13932 ppl-value-checks/Gemma-4-26B-A4B-NVFP4_fp8_q8.fusion_enabled.log /mnt/share/gguf/nvidia/Gemma-4-26B-A4B-NVFP4/Gemma-4-26B-A4B-NVFP4_fp8_q8.gguf fusion_disabled 18627.0624 3998.42475 ppl-value-checks/Gemma-4-26B-A4B-NVFP4_fp8_q8.fusion_disabled.log /mnt/share/gguf/ggml-org/gpt-oss-20b-GGUF/gpt-oss-20b-mxfp4.gguf fusion_enabled 363.8913 33.14007 ppl-value-checks/gpt-oss-20b-mxfp4.fusion_enabled.log /mnt/share/gguf/ggml-org/gpt-oss-20b-GGUF/gpt-oss-20b-mxfp4.gguf fusion_disabled 363.8913 33.14007 ppl-value-checks/gpt-oss-20b-mxfp4.fusion_disabled.log /mnt/share/gguf/unsloth/gemma-4-26B-A4B-it-GGUF/gemma-4-26B-A4B-it-UD-Q4_K_XL.gguf fusion_enabled 17330.3926 3716.70472 ppl-value-checks/gemma-4-26B-A4B-it-UD-Q4_K_XL.fusion_enabled.log /mnt/share/gguf/unsloth/gemma-4-26B-A4B-it-GGUF/gemma-4-26B-A4B-it-UD-Q4_K_XL.gguf fusion_disabled 17933.9524 3883.17066 ppl-value-checks/gemma-4-26B-A4B-it-UD-Q4_K_XL.fusion_disabled.log * Allow views to weights in ggml_can_fuse_subgraph * Remove gate_first from test_mul_mat_vec_fusion * Ditch lane-parsing approach in favor of hard-coded patterns * Apply suggestions from code review Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> * Rename ggml_is_constant_view_src to ggml_is_constant * Finish renaming of `0905129e9d` * Readd descriptive prints for fusion debugging * Add weight-buffer pre-allocation to `test_case` This is required so we correctly test fusion of NVFP4. * Update ggml/src/ggml.c Co-authored-by: Johannes Gäßler <johannesg@5d6.de> * Add 2nd context for weights as suggested by @JohannesGaessler This reflects more natural use of ggml compared to artifically pre-allocating weights into the same context * Exclude fused tests from gradient mode I'm unsure of the current state, but naively every fusion pattern should require its own backpropagation implementation. I don't see these implemented for the CUDA backend, so we can disable tests to avoid triggering GGML_ASSERT for ggml_tensor * build_graph(ggml_context * ctx) override { GGML_ASSERT(!use_weight_context()); return build_graph(ctx, nullptr); } * Apply suggestions from code review Co-authored-by: Johannes Gäßler <johannesg@5d6.de> --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> Co-authored-by: Johannes Gäßler <johannesg@5d6.de>	2026-07-07 17:12:19 +02:00
Alex	f5525f7e7a	server : fix draft model fit vs load inconsistency (#25056 ) * fix: draft model fit vs load inconsistency * refactor(server): unify draft/mtp parameter initialization, model, and context load - moves speculative init to speculative.cpp - changes server_context_impl model_dft and ctx_dft to use raw pointers - fix: don't throttle progress callback when loading draft model - refactor: rename draft model/ctx load method * fix: valign	2026-07-07 17:20:42 +03:00
Thomas LECONTE	5eca4e3cab	server : add timings and progress to /responses API stream (#25348 )	2026-07-07 16:13:03 +02:00
Thiago Padilha	6c487e2f79	server: enforce prompt cache RAM limit (#25070 ) Before this commit, --cache-ram was not a hard limit: - The cache always kept at least one entry, even if that entry exceeded the RAM/token limits. - Old entries were only evicted for the RAM/token limits after saving the new one, which could cause the cache to temporarily exceed the RAM/token limits even if individual entries were below the limit. Now, ensure that the RAM limit is strict with these changes: - Skip saving state to cache if by itself it exceeds the RAM limit. - Evict old entries as necessary to make the new entry fit. Additionally, token-limit cleanup may now evict the last remaining cache entry instead of always preserving one.	2026-07-07 15:24:35 +02:00