opencl: Q6_K GEMM/GEMV fix for ne01 of weights that are not multiples of 128. (#25464 )

* opencl: fix garbled output for Q6_K weights with ne01 % 128 != 0 on Adreno Observed with granite-3.1-3b-a800m-instruct, whose vocab is an odd number. Route Q6_K dense mul_mat with ne01 % 128 != 0 off the noshuffle path: decode (ne1==1) uses the correct flat GEMV and the matching GEMM (ne1>1) falls back to CPU (the flat convert has no verified small-batch GEMM kernel for these shapes). All standard hidden/FFN/vocab dims are multiples of 128 and keep the noshuffle path. * opencl: reserve alignment slack for the SOA subbuffer carve in alloc size set_tensor carves quantized weights into per-component subbuffers (d/q, ql/qh/s/d, ...) whose origins are each rounded up to the device base address alignment. When a component's size is not a multiple of the alignment, the carve extends past ggml_nbytes(tensor) and the last subbuffer overlaps the next tensor in the pool -- e.g. q6_K [1536, 49155]: size_s = 49155*96 ends 32 bytes past a 128-byte boundary, so the d subbuffer ends 96 bytes past the tensor's allocation, and whichever of the two neighboring tensors is uploaded last silently corrupts the other (here: the last vocab rows' block scales). This affects any quant type whose component sizes can be misaligned, on any shape with ne01 not a multiple of the alignment granularity; standard power-of-two dims are unaffected. Implement get_alloc_size for the OpenCL buffer type and reserve the worst-case carve slack (4 aligned gaps; 5 components max, q5_K) for quantized tensors. Costs at most 512 bytes per quantized tensor at the observed 128-byte alignment. * opencl: use lm based q6_k mm when ne1 is not multiple of 128 --------- Co-authored-by: Li He <lih@qti.qualcomm.com>
vulkan: disable FA mask_opt on GCN to improve performance (#24362 )
2026-07-10 06:25:55 +02:00 · 2026-07-08 15:52:21 -07:00 · 2026-07-08 19:01:25 +02:00 · 2026-07-08 09:44:55 -07:00 · 2026-07-08 19:24:34 +03:00 · 2026-07-08 18:15:18 +02:00
11 changed files with 673 additions and 52 deletions
@@ -517,6 +517,10 @@ struct ggml_backend_opencl_context {
    bool has_qcom_subgroup_shuffle = false;  // specifically cl_qcom_subgroup_shuffle
    bool disable_fusion;

+    // ragged moe, use int to directly pass to kernel
+    cl_uint  adreno_use_moe_ragged;
+    cl_uint  adreno_moe_ragged_skip_gran;
+
    bool adreno_has_large_buffer;
    bool adreno_use_large_buffer;
    bool adreno_use_bin_kernels;
@@ -5342,6 +5346,15 @@ static ggml_backend_opencl_context * ggml_cl_init(ggml_backend_dev_t dev) {
    backend_ctx->adreno_use_large_buffer = getenv("GGML_OPENCL_ADRENO_USE_LARGE_BUFFER") != nullptr &&
                                           backend_ctx->gpu_family == GPU_FAMILY::ADRENO;

+    // ragged moe, unspecified or non-zero means enabled, set to 0 to disable
+    static const char * ragged_fp16_env = getenv("GGML_OPENCL_MOE_RAGGED_FP16");
+    backend_ctx->adreno_use_moe_ragged = (ragged_fp16_env == NULL) ? 1 : (atoi(ragged_fp16_env) != 0);
+
+    // ragged moe, tile-skip granularity (columns per skip-group): 8 = quarter (default),
+    // 16 = half (legacy), 32 = disabled. Override with GGML_OPENCL_MOE_RAGGED_GRAN={8,16,32}
+    static const char * ragged_gran_env = getenv("GGML_OPENCL_MOE_RAGGED_GRAN");
+    backend_ctx->adreno_moe_ragged_skip_gran = (ragged_gran_env != NULL) ? atoi(ragged_gran_env) : 8;
+
 #ifdef GGML_OPENCL_USE_ADRENO_BIN_KERNELS
    // try loading adreno binary kernels if enabled
    // if fails to load, builtin kernels will be used
@@ -6229,8 +6242,14 @@ inline bool use_adreno_kernels(const ggml_backend_opencl_context *backend_ctx, c
        threshold_ne0 = 128;
        threshold_ne1 = 128;
    }
-    return tensor->ne[0] >= threshold_ne0 && tensor->ne[1] >= threshold_ne1 &&
+    bool threashold_ok = tensor->ne[0] >= threshold_ne0 && tensor->ne[1] >= threshold_ne1 &&
            tensor->ne[2] == 1 && tensor->ne[3] == 1;
+
+    // q6_K adreno kernels requires ne1 is multiple of 128
+    if (tensor->type == GGML_TYPE_Q6_K) {
+        return threashold_ok && tensor->ne[1] % 128 == 0;
+    }
+    return threashold_ok;
 }

 inline bool use_adreno_moe_kernels(const ggml_backend_opencl_context *backend_ctx, const ggml_tensor *tensor) {
@@ -6260,6 +6279,19 @@ static inline bool use_flat_gemv_for_large_m_q6_K(const ggml_tensor *tensor) {
    // threshold is well above typical hidden/FFN dims, but below typical vocab sizes.
    // q6_K flat gemv is worse for smaller K; 2048 seems to be a reasonable threshold.
    // note that this forces large M weights to use LM GEMM.
+    // The noshuffle (transposed-weight) layout packs 2 rows per 32-bit texel and the
+    // gemv reads it with a ne01/2 texel stride and an exact-cover dispatch of
+    // ceil(ne01/2 / 64)*64 work-items with no store guard; the gemm uses 4-row tiles.
+    // It is therefore only correct for ne01 % 128 == 0: an odd ne01 (e.g. granitemoe
+    // lm_head [1536, 49155] -- odd vocab) truncates the texel stride, misaligning every
+    // odd column of the transposed layout (gross garbage) and dropping the last row;
+    // other non-multiples over-dispatch and write past the end of dst. Route such
+    // tensors to the flat GEMV + regular convert; the matching GEMM (ne1>1) falls back
+    // to CPU (see supports_op). All standard even-vocab/hidden dims are multiples of
+    // 128 and keep the noshuffle path.
+    if ((tensor->ne[1] % 128 != 0) && tensor->ne[2] == 1 && tensor->ne[3] == 1) {
+        return true;
+    }
    return tensor->ne[1] >= 32768 && tensor->ne[0] >= 2048 && tensor->ne[2] == 1 && tensor->ne[3] == 1;
 }

@@ -9763,12 +9795,30 @@ static bool ggml_backend_opencl_buffer_type_supports_backend(ggml_backend_buffer
    UNUSED(buft);
 }

+static size_t ggml_backend_opencl_buffer_type_get_alloc_size(ggml_backend_buffer_type_t buft, const ggml_tensor * tensor) {
+    size_t size = ggml_nbytes(tensor);
+#ifdef GGML_OPENCL_SOA_Q
+    // set_tensor carves quantized weights into per-component subbuffers (d/q,
+    // ql/qh/s/d, ...) whose origins are each rounded up to the device base
+    // alignment. When a component's size is not a multiple of the alignment
+    // (e.g. q6_K [1536,49155]: size_s = 49155*96 leaves a 96-byte gap at 128-byte
+    // alignment), the aligned carve extends past ggml_nbytes and the last
+    // subbuffer would overlap the next tensor in the pool. Reserve the worst-case
+    // carve slack: at most 5 components (q5_K), i.e. 4 aligned gaps.
+    if (ggml_is_quantized(tensor->type)) {
+        ggml_backend_opencl_device_context * dev_ctx = (ggml_backend_opencl_device_context *) buft->device->context;
+        size += 4 * dev_ctx->backend_ctx->alignment;
+    }
+#endif // GGML_OPENCL_SOA_Q
+    return size;
+}
+
 static ggml_backend_buffer_type_i ggml_backend_opencl_buffer_type_interface = {
    /* .get_name         = */ ggml_backend_opencl_buffer_type_get_name,
    /* .alloc_buffer     = */ ggml_backend_opencl_buffer_type_alloc_buffer,
    /* .get_alignment    = */ ggml_backend_opencl_buffer_type_get_alignment,
    /* .get_max_size     = */ ggml_backend_opencl_buffer_type_get_max_size,
-    /* .get_alloc_size   = */ NULL,
+    /* .get_alloc_size   = */ ggml_backend_opencl_buffer_type_get_alloc_size,
    /* .is_host          = */ NULL,
 };

@@ -19338,6 +19388,8 @@ static void ggml_cl_mul_mat_id(ggml_backend_t backend, const ggml_tensor * src0,
                    CL_CHECK(clSetKernelArg(kernel, arg_idx++, sizeof(cl_mem),    &(backend_ctx->prealloc_total_tiles.buffer)));
                    CL_CHECK(clSetKernelArg(kernel, arg_idx++, sizeof(int),       &ne00));
                    CL_CHECK(clSetKernelArg(kernel, arg_idx++, sizeof(int),       &ne01));
+                    CL_CHECK(clSetKernelArg(kernel, arg_idx++, sizeof(cl_uint),   &backend_ctx->adreno_use_moe_ragged));
+                    CL_CHECK(clSetKernelArg(kernel, arg_idx++, sizeof(cl_uint),   &backend_ctx->adreno_moe_ragged_skip_gran));

                    // set thread grid
                    global_size[1] = static_cast<size_t>((ne01 + 63) / 64);
@@ -19564,6 +19616,8 @@ static void ggml_cl_mul_mat_id(ggml_backend_t backend, const ggml_tensor * src0,
                    CL_CHECK(clSetKernelArg(kernel, arg_idx++, sizeof(cl_mem),    &(backend_ctx->prealloc_total_tiles.buffer)));
                    CL_CHECK(clSetKernelArg(kernel, arg_idx++, sizeof(int),       &ne00));
                    CL_CHECK(clSetKernelArg(kernel, arg_idx++, sizeof(int),       &ne01));
+                    CL_CHECK(clSetKernelArg(kernel, arg_idx++, sizeof(cl_uint),   &backend_ctx->adreno_use_moe_ragged));
+                    CL_CHECK(clSetKernelArg(kernel, arg_idx++, sizeof(cl_uint),   &backend_ctx->adreno_moe_ragged_skip_gran));

                    // set thread grid
                    global_size[1] = static_cast<size_t>((ne01 + 63) / 64);
@@ -19740,6 +19794,8 @@ static void ggml_cl_mul_mat_id(ggml_backend_t backend, const ggml_tensor * src0,
                    CL_CHECK(clSetKernelArg(kernel, arg_idx++, sizeof(cl_mem),    &(backend_ctx->prealloc_total_tiles.buffer)));
                    CL_CHECK(clSetKernelArg(kernel, arg_idx++, sizeof(int),       &ne00));
                    CL_CHECK(clSetKernelArg(kernel, arg_idx++, sizeof(int),       &ne01));
+                    CL_CHECK(clSetKernelArg(kernel, arg_idx++, sizeof(cl_uint),   &backend_ctx->adreno_use_moe_ragged));
+                    CL_CHECK(clSetKernelArg(kernel, arg_idx++, sizeof(cl_uint),   &backend_ctx->adreno_moe_ragged_skip_gran));

                    // set thread grid
                    global_size[1] = static_cast<size_t>((ne01 + 63) / 64);
@@ -19917,6 +19973,8 @@ static void ggml_cl_mul_mat_id(ggml_backend_t backend, const ggml_tensor * src0,
                    CL_CHECK(clSetKernelArg(kernel, arg_idx++, sizeof(cl_mem),    &(backend_ctx->prealloc_total_tiles.buffer)));
                    CL_CHECK(clSetKernelArg(kernel, arg_idx++, sizeof(int),       &ne00));
                    CL_CHECK(clSetKernelArg(kernel, arg_idx++, sizeof(int),       &ne01));
+                    CL_CHECK(clSetKernelArg(kernel, arg_idx++, sizeof(cl_uint),       &backend_ctx->adreno_use_moe_ragged));
+                    CL_CHECK(clSetKernelArg(kernel, arg_idx++, sizeof(cl_uint),       &backend_ctx->adreno_moe_ragged_skip_gran));

                    // set thread grid
                    global_size[1] = static_cast<size_t>((ne01 + 63) / 64);
@@ -20174,6 +20232,8 @@ static void ggml_cl_mul_mat_id(ggml_backend_t backend, const ggml_tensor * src0,
                    CL_CHECK(clSetKernelArg(kernel, arg_idx++, sizeof(cl_mem),    &(backend_ctx->prealloc_total_tiles.buffer)));
                    CL_CHECK(clSetKernelArg(kernel, arg_idx++, sizeof(int),       &ne00));
                    CL_CHECK(clSetKernelArg(kernel, arg_idx++, sizeof(int),       &ne01));
+                    CL_CHECK(clSetKernelArg(kernel, arg_idx++, sizeof(cl_uint),   &backend_ctx->adreno_use_moe_ragged));
+                    CL_CHECK(clSetKernelArg(kernel, arg_idx++, sizeof(cl_uint),   &backend_ctx->adreno_moe_ragged_skip_gran));

                    // set thread grid
                    global_size[1] = static_cast<size_t>((ne01 + 63) / 64);
@@ -20352,6 +20412,8 @@ static void ggml_cl_mul_mat_id(ggml_backend_t backend, const ggml_tensor * src0,
                    CL_CHECK(clSetKernelArg(kernel, arg_idx++, sizeof(cl_mem),    &(backend_ctx->prealloc_total_tiles.buffer)));
                    CL_CHECK(clSetKernelArg(kernel, arg_idx++, sizeof(int),       &ne00));
                    CL_CHECK(clSetKernelArg(kernel, arg_idx++, sizeof(int),       &ne01));
+                    CL_CHECK(clSetKernelArg(kernel, arg_idx++, sizeof(cl_uint),   &backend_ctx->adreno_use_moe_ragged));
+                    CL_CHECK(clSetKernelArg(kernel, arg_idx++, sizeof(cl_uint),   &backend_ctx->adreno_moe_ragged_skip_gran));

                    // set thread grid
                    global_size[1] = static_cast<size_t>((ne01 + 63) / 64);
@@ -20527,6 +20589,8 @@ static void ggml_cl_mul_mat_id(ggml_backend_t backend, const ggml_tensor * src0,
                    CL_CHECK(clSetKernelArg(kernel, arg_idx++, sizeof(cl_mem),    &(backend_ctx->prealloc_total_tiles.buffer)));
                    CL_CHECK(clSetKernelArg(kernel, arg_idx++, sizeof(int),       &ne00));
                    CL_CHECK(clSetKernelArg(kernel, arg_idx++, sizeof(int),       &ne01));
+                    CL_CHECK(clSetKernelArg(kernel, arg_idx++, sizeof(cl_uint),   &backend_ctx->adreno_use_moe_ragged));
+                    CL_CHECK(clSetKernelArg(kernel, arg_idx++, sizeof(cl_uint),   &backend_ctx->adreno_moe_ragged_skip_gran));

                    // set thread grid
                    global_size[1] = static_cast<size_t>((ne01 + 63) / 64);
@@ -20710,6 +20774,8 @@ static void ggml_cl_mul_mat_id(ggml_backend_t backend, const ggml_tensor * src0,
                    CL_CHECK(clSetKernelArg(kernel, arg_idx++, sizeof(cl_mem),    &(backend_ctx->prealloc_total_tiles.buffer)));
                    CL_CHECK(clSetKernelArg(kernel, arg_idx++, sizeof(int),       &ne00));
                    CL_CHECK(clSetKernelArg(kernel, arg_idx++, sizeof(int),       &ne01));
+                    CL_CHECK(clSetKernelArg(kernel, arg_idx++, sizeof(cl_uint),   &backend_ctx->adreno_use_moe_ragged));
+                    CL_CHECK(clSetKernelArg(kernel, arg_idx++, sizeof(cl_uint),   &backend_ctx->adreno_moe_ragged_skip_gran));

                    // set thread grid
                    global_size[1] = static_cast<size_t>((ne01 + 63) / 64);
@@ -132,6 +132,46 @@ static inline half8 mxfp4_to_fp16_packed8(ushort2 fp4x8) {
    c_reg.lo += convert_float8(acc.lo); \
    c_reg.hi += convert_float8(acc.hi); \

+// Quarter-tile variant: computes 8 output columns (one skip-group) into a float8
+// accumulator. Same reduction order / flush cadence as dotx16_reduce8, so the
+// non-skipped path is byte-identical; it just lets the caller skip empty
+// 8-column groups at finer granularity. Uses a private half8 `acc8`.
+#define dotx8_reduce4(a_reg, b_lm, c_reg, lm_offset) \
+    acc8.s0 = dot(a_reg.s0123, b_lm[lm_offset + 0]); \
+    acc8.s1 = dot(a_reg.s0123, b_lm[lm_offset + 1]); \
+    acc8.s2 = dot(a_reg.s0123, b_lm[lm_offset + 2]); \
+    acc8.s3 = dot(a_reg.s0123, b_lm[lm_offset + 3]); \
+    acc8.s4 = dot(a_reg.s0123, b_lm[lm_offset + 4]); \
+    acc8.s5 = dot(a_reg.s0123, b_lm[lm_offset + 5]); \
+    acc8.s6 = dot(a_reg.s0123, b_lm[lm_offset + 6]); \
+    acc8.s7 = dot(a_reg.s0123, b_lm[lm_offset + 7]); \
+    acc8.s0 += dot(a_reg.s4567, b_lm[lm_offset + 32]); \
+    acc8.s1 += dot(a_reg.s4567, b_lm[lm_offset + 33]); \
+    acc8.s2 += dot(a_reg.s4567, b_lm[lm_offset + 34]); \
+    acc8.s3 += dot(a_reg.s4567, b_lm[lm_offset + 35]); \
+    acc8.s4 += dot(a_reg.s4567, b_lm[lm_offset + 36]); \
+    acc8.s5 += dot(a_reg.s4567, b_lm[lm_offset + 37]); \
+    acc8.s6 += dot(a_reg.s4567, b_lm[lm_offset + 38]); \
+    acc8.s7 += dot(a_reg.s4567, b_lm[lm_offset + 39]); \
+    c_reg += convert_float8(acc8); \
+    acc8.s0 = dot(a_reg.s89ab, b_lm[lm_offset + 64]); \
+    acc8.s1 = dot(a_reg.s89ab, b_lm[lm_offset + 65]); \
+    acc8.s2 = dot(a_reg.s89ab, b_lm[lm_offset + 66]); \
+    acc8.s3 = dot(a_reg.s89ab, b_lm[lm_offset + 67]); \
+    acc8.s4 = dot(a_reg.s89ab, b_lm[lm_offset + 68]); \
+    acc8.s5 = dot(a_reg.s89ab, b_lm[lm_offset + 69]); \
+    acc8.s6 = dot(a_reg.s89ab, b_lm[lm_offset + 70]); \
+    acc8.s7 = dot(a_reg.s89ab, b_lm[lm_offset + 71]); \
+    acc8.s0 += dot(a_reg.scdef, b_lm[lm_offset + 96]); \
+    acc8.s1 += dot(a_reg.scdef, b_lm[lm_offset + 97]); \
+    acc8.s2 += dot(a_reg.scdef, b_lm[lm_offset + 98]); \
+    acc8.s3 += dot(a_reg.scdef, b_lm[lm_offset + 99]); \
+    acc8.s4 += dot(a_reg.scdef, b_lm[lm_offset + 100]); \
+    acc8.s5 += dot(a_reg.scdef, b_lm[lm_offset + 101]); \
+    acc8.s6 += dot(a_reg.scdef, b_lm[lm_offset + 102]); \
+    acc8.s7 += dot(a_reg.scdef, b_lm[lm_offset + 103]); \
+    c_reg += convert_float8(acc8); \
+

 static inline half e8m0_to_fp16(uchar x) {
    ushort bits;
@@ -157,7 +197,9 @@ kernel void kernel_gemm_moe_mxfp4_f32_ns(
        __write_only image1d_buffer_t dst,
        __global     int *            total_tiles,
        uint ne00,
-        uint ne01
+        uint ne01,
+        uint is_ragged,
+        uint skip_gran
 ) {
    uint block_id_m = get_global_id(1); // m_tile
    uint block_id_n = get_global_id(2); // n_tile
@@ -167,6 +209,28 @@ kernel void kernel_gemm_moe_mxfp4_f32_ns(
        return;
    }

+    // Ragged tile-skip: when is_ragged and the upper 16 token-slots of this tile are all
+    // padding (router 0xFFFFFFFF), skip the second (reg_c.hi) dotx16_reduce8 half -> ~half
+    // the GEMM dot for sparse tiles. Numerically identical (the skipped lanes are padding).
+    // Ragged tile-skip: tokens are packed contiguously per expert (moe_scatter fills
+    // lanes 0..V-1, moe_fill pre-pads the rest), so router padding (0xFFFFFFFF) is always
+    // trailing. Find the valid-token count V and round it UP to the skip granularity
+    // skip_gran (columns per skip-group: 8 = quarter, 16 = half/legacy, 32 = disabled).
+    // A 8-column group g is all-padding iff its first column (8*g) >= n_active, so its
+    // dotx8_reduce4 is skipped. Numerically identical (skipped lanes are padding).
+    uint n_active = TILESIZE_N;
+    if (is_ragged && skip_gran < TILESIZE_N) {
+        uint n_valid = TILESIZE_N;
+        for (uint _t = 0; _t < TILESIZE_N; ++_t) {
+            if (src2[block_id_n * TILESIZE_N + _t] == 0xFFFFFFFFu) { n_valid = _t; break; }
+        }
+        n_active = min((uint)TILESIZE_N, ((n_valid + skip_gran - 1) / skip_gran) * skip_gran);
+    }
+    // Group 0 (cols 0-7) always runs; groups 1-3 skip when fully padding.
+    bool skip_g1 = (8u  >= n_active);
+    bool skip_g2 = (16u >= n_active);
+    bool skip_g3 = (24u >= n_active);
+
    __private half16 reg_a;
    __private float32 reg_c = (float32)(0);
    __local half4 shared_b[128];
@@ -216,9 +280,11 @@ kernel void kernel_gemm_moe_mxfp4_f32_ns(
        sub_group_barrier(CLK_LOCAL_MEM_FENCE);

        // 32 16x16 fp16 dot product with 8 elements reduction for better precision
-        half16 acc;
-        dotx16_reduce8(reg_a, shared_b, reg_c.lo, 0);
-        dotx16_reduce8(reg_a, shared_b, reg_c.hi, 16);
+        half8 acc8;
+        dotx8_reduce4(reg_a, shared_b, reg_c.lo.lo, 0);
+        if (!skip_g1) { dotx8_reduce4(reg_a, shared_b, reg_c.lo.hi, 8); }
+        if (!skip_g2) { dotx8_reduce4(reg_a, shared_b, reg_c.hi.lo, 16); }
+        if (!skip_g3) { dotx8_reduce4(reg_a, shared_b, reg_c.hi.hi, 24); }

        // Repeat for second sub-block
        uint half_step = step + TILESIZE_K;
@@ -244,8 +310,10 @@ kernel void kernel_gemm_moe_mxfp4_f32_ns(
        sub_group_barrier(CLK_LOCAL_MEM_FENCE);

        // 32 16x16 fp16 dot product with 3-levels reduction for better precision
-        dotx16_reduce8(reg_a, shared_b, reg_c.lo, 0);
-        dotx16_reduce8(reg_a, shared_b, reg_c.hi, 16);
+        dotx8_reduce4(reg_a, shared_b, reg_c.lo.lo, 0);
+        if (!skip_g1) { dotx8_reduce4(reg_a, shared_b, reg_c.lo.hi, 8); }
+        if (!skip_g2) { dotx8_reduce4(reg_a, shared_b, reg_c.hi.lo, 16); }
+        if (!skip_g3) { dotx8_reduce4(reg_a, shared_b, reg_c.hi.hi, 24); }
    }

    if ((get_global_id(0) + block_id_m * TILESIZE_M) >= ne01) {
@@ -98,6 +98,46 @@
    c_reg.lo += convert_float8(acc.lo); \
    c_reg.hi += convert_float8(acc.hi); \

+// Quarter-tile variant: computes 8 output columns (one skip-group) into a float8
+// accumulator. Same reduction order / flush cadence as dotx16_reduce8, so the
+// non-skipped path is byte-identical; it just lets the caller skip empty
+// 8-column groups at finer granularity. Uses a private half8 `acc8`.
+#define dotx8_reduce4(a_reg, b_lm, c_reg, lm_offset) \
+    acc8.s0 = dot(a_reg.s0123, b_lm[lm_offset + 0]); \
+    acc8.s1 = dot(a_reg.s0123, b_lm[lm_offset + 1]); \
+    acc8.s2 = dot(a_reg.s0123, b_lm[lm_offset + 2]); \
+    acc8.s3 = dot(a_reg.s0123, b_lm[lm_offset + 3]); \
+    acc8.s4 = dot(a_reg.s0123, b_lm[lm_offset + 4]); \
+    acc8.s5 = dot(a_reg.s0123, b_lm[lm_offset + 5]); \
+    acc8.s6 = dot(a_reg.s0123, b_lm[lm_offset + 6]); \
+    acc8.s7 = dot(a_reg.s0123, b_lm[lm_offset + 7]); \
+    acc8.s0 += dot(a_reg.s4567, b_lm[lm_offset + 32]); \
+    acc8.s1 += dot(a_reg.s4567, b_lm[lm_offset + 33]); \
+    acc8.s2 += dot(a_reg.s4567, b_lm[lm_offset + 34]); \
+    acc8.s3 += dot(a_reg.s4567, b_lm[lm_offset + 35]); \
+    acc8.s4 += dot(a_reg.s4567, b_lm[lm_offset + 36]); \
+    acc8.s5 += dot(a_reg.s4567, b_lm[lm_offset + 37]); \
+    acc8.s6 += dot(a_reg.s4567, b_lm[lm_offset + 38]); \
+    acc8.s7 += dot(a_reg.s4567, b_lm[lm_offset + 39]); \
+    c_reg += convert_float8(acc8); \
+    acc8.s0 = dot(a_reg.s89ab, b_lm[lm_offset + 64]); \
+    acc8.s1 = dot(a_reg.s89ab, b_lm[lm_offset + 65]); \
+    acc8.s2 = dot(a_reg.s89ab, b_lm[lm_offset + 66]); \
+    acc8.s3 = dot(a_reg.s89ab, b_lm[lm_offset + 67]); \
+    acc8.s4 = dot(a_reg.s89ab, b_lm[lm_offset + 68]); \
+    acc8.s5 = dot(a_reg.s89ab, b_lm[lm_offset + 69]); \
+    acc8.s6 = dot(a_reg.s89ab, b_lm[lm_offset + 70]); \
+    acc8.s7 = dot(a_reg.s89ab, b_lm[lm_offset + 71]); \
+    acc8.s0 += dot(a_reg.scdef, b_lm[lm_offset + 96]); \
+    acc8.s1 += dot(a_reg.scdef, b_lm[lm_offset + 97]); \
+    acc8.s2 += dot(a_reg.scdef, b_lm[lm_offset + 98]); \
+    acc8.s3 += dot(a_reg.scdef, b_lm[lm_offset + 99]); \
+    acc8.s4 += dot(a_reg.scdef, b_lm[lm_offset + 100]); \
+    acc8.s5 += dot(a_reg.scdef, b_lm[lm_offset + 101]); \
+    acc8.s6 += dot(a_reg.scdef, b_lm[lm_offset + 102]); \
+    acc8.s7 += dot(a_reg.scdef, b_lm[lm_offset + 103]); \
+    c_reg += convert_float8(acc8); \
+

 __attribute__((qcom_wave_pair_mode(1))) // 1=force single 2=force pair
 kernel void kernel_gemm_moe_q4_0_f32_ns(
@@ -109,7 +149,9 @@ kernel void kernel_gemm_moe_q4_0_f32_ns(
        __write_only image1d_buffer_t dst,
        __global     int *            total_tiles,
        uint ne00,
-        uint ne01
+        uint ne01,
+        uint is_ragged,
+        uint skip_gran
 ) {
    uint block_id_m = get_global_id(1); // m_tile
    uint block_id_n = get_global_id(2); // n_tile
@@ -119,6 +161,28 @@ kernel void kernel_gemm_moe_q4_0_f32_ns(
        return;
    }

+    // Ragged tile-skip: when is_ragged and the upper 16 token-slots of this tile are all
+    // padding (router 0xFFFFFFFF), skip the second (reg_c.hi) dotx16_reduce8 half -> ~half
+    // the GEMM dot for sparse tiles. Numerically identical (the skipped lanes are padding).
+    // Ragged tile-skip: tokens are packed contiguously per expert (moe_scatter fills
+    // lanes 0..V-1, moe_fill pre-pads the rest), so router padding (0xFFFFFFFF) is always
+    // trailing. Find the valid-token count V and round it UP to the skip granularity
+    // skip_gran (columns per skip-group: 8 = quarter, 16 = half/legacy, 32 = disabled).
+    // A 8-column group g is all-padding iff its first column (8*g) >= n_active, so its
+    // dotx8_reduce4 is skipped. Numerically identical (skipped lanes are padding).
+    uint n_active = TILESIZE_N;
+    if (is_ragged && skip_gran < TILESIZE_N) {
+        uint n_valid = TILESIZE_N;
+        for (uint _t = 0; _t < TILESIZE_N; ++_t) {
+            if (src2[block_id_n * TILESIZE_N + _t] == 0xFFFFFFFFu) { n_valid = _t; break; }
+        }
+        n_active = min((uint)TILESIZE_N, ((n_valid + skip_gran - 1) / skip_gran) * skip_gran);
+    }
+    // Group 0 (cols 0-7) always runs; groups 1-3 skip when fully padding.
+    bool skip_g1 = (8u  >= n_active);
+    bool skip_g2 = (16u >= n_active);
+    bool skip_g3 = (24u >= n_active);
+
    __private half16 reg_a;
    __private float32 reg_c = (float32)(0);
    __local half4 shared_b[128];
@@ -167,9 +231,11 @@ kernel void kernel_gemm_moe_q4_0_f32_ns(
        sub_group_barrier(CLK_LOCAL_MEM_FENCE);

        // 32 16x16 fp16 dot product with 8 elements reduction for better precision
-        half16 acc;
-        dotx16_reduce8(reg_a, shared_b, reg_c.lo, 0);
-        dotx16_reduce8(reg_a, shared_b, reg_c.hi, 16);
+        half8 acc8;
+        dotx8_reduce4(reg_a, shared_b, reg_c.lo.lo, 0);
+        if (!skip_g1) { dotx8_reduce4(reg_a, shared_b, reg_c.lo.hi, 8); }
+        if (!skip_g2) { dotx8_reduce4(reg_a, shared_b, reg_c.hi.lo, 16); }
+        if (!skip_g3) { dotx8_reduce4(reg_a, shared_b, reg_c.hi.hi, 24); }

        // Repeat for second sub-block
        uint half_step = step + TILESIZE_K;
@@ -194,8 +260,10 @@ kernel void kernel_gemm_moe_q4_0_f32_ns(
        sub_group_barrier(CLK_LOCAL_MEM_FENCE);

        // 32 16x16 fp16 dot product with 3-levels reduction for better precision
-        dotx16_reduce8(reg_a, shared_b, reg_c.lo, 0);
-        dotx16_reduce8(reg_a, shared_b, reg_c.hi, 16);
+        dotx8_reduce4(reg_a, shared_b, reg_c.lo.lo, 0);
+        if (!skip_g1) { dotx8_reduce4(reg_a, shared_b, reg_c.lo.hi, 8); }
+        if (!skip_g2) { dotx8_reduce4(reg_a, shared_b, reg_c.hi.lo, 16); }
+        if (!skip_g3) { dotx8_reduce4(reg_a, shared_b, reg_c.hi.hi, 24); }
    }

    if ((get_global_id(0) + block_id_m * TILESIZE_M) >= ne01) {
@@ -98,6 +98,46 @@
    c_reg.lo += convert_float8(acc.lo); \
    c_reg.hi += convert_float8(acc.hi); \

+// Quarter-tile variant: computes 8 output columns (one skip-group) into a float8
+// accumulator. Same reduction order / flush cadence as dotx16_reduce8, so the
+// non-skipped path is byte-identical; it just lets the caller skip empty
+// 8-column groups at finer granularity. Uses a private half8 `acc8`.
+#define dotx8_reduce4(a_reg, b_lm, c_reg, lm_offset) \
+    acc8.s0 = dot(a_reg.s0123, b_lm[lm_offset + 0]); \
+    acc8.s1 = dot(a_reg.s0123, b_lm[lm_offset + 1]); \
+    acc8.s2 = dot(a_reg.s0123, b_lm[lm_offset + 2]); \
+    acc8.s3 = dot(a_reg.s0123, b_lm[lm_offset + 3]); \
+    acc8.s4 = dot(a_reg.s0123, b_lm[lm_offset + 4]); \
+    acc8.s5 = dot(a_reg.s0123, b_lm[lm_offset + 5]); \
+    acc8.s6 = dot(a_reg.s0123, b_lm[lm_offset + 6]); \
+    acc8.s7 = dot(a_reg.s0123, b_lm[lm_offset + 7]); \
+    acc8.s0 += dot(a_reg.s4567, b_lm[lm_offset + 32]); \
+    acc8.s1 += dot(a_reg.s4567, b_lm[lm_offset + 33]); \
+    acc8.s2 += dot(a_reg.s4567, b_lm[lm_offset + 34]); \
+    acc8.s3 += dot(a_reg.s4567, b_lm[lm_offset + 35]); \
+    acc8.s4 += dot(a_reg.s4567, b_lm[lm_offset + 36]); \
+    acc8.s5 += dot(a_reg.s4567, b_lm[lm_offset + 37]); \
+    acc8.s6 += dot(a_reg.s4567, b_lm[lm_offset + 38]); \
+    acc8.s7 += dot(a_reg.s4567, b_lm[lm_offset + 39]); \
+    c_reg += convert_float8(acc8); \
+    acc8.s0 = dot(a_reg.s89ab, b_lm[lm_offset + 64]); \
+    acc8.s1 = dot(a_reg.s89ab, b_lm[lm_offset + 65]); \
+    acc8.s2 = dot(a_reg.s89ab, b_lm[lm_offset + 66]); \
+    acc8.s3 = dot(a_reg.s89ab, b_lm[lm_offset + 67]); \
+    acc8.s4 = dot(a_reg.s89ab, b_lm[lm_offset + 68]); \
+    acc8.s5 = dot(a_reg.s89ab, b_lm[lm_offset + 69]); \
+    acc8.s6 = dot(a_reg.s89ab, b_lm[lm_offset + 70]); \
+    acc8.s7 = dot(a_reg.s89ab, b_lm[lm_offset + 71]); \
+    acc8.s0 += dot(a_reg.scdef, b_lm[lm_offset + 96]); \
+    acc8.s1 += dot(a_reg.scdef, b_lm[lm_offset + 97]); \
+    acc8.s2 += dot(a_reg.scdef, b_lm[lm_offset + 98]); \
+    acc8.s3 += dot(a_reg.scdef, b_lm[lm_offset + 99]); \
+    acc8.s4 += dot(a_reg.scdef, b_lm[lm_offset + 100]); \
+    acc8.s5 += dot(a_reg.scdef, b_lm[lm_offset + 101]); \
+    acc8.s6 += dot(a_reg.scdef, b_lm[lm_offset + 102]); \
+    acc8.s7 += dot(a_reg.scdef, b_lm[lm_offset + 103]); \
+    c_reg += convert_float8(acc8); \
+

 __attribute__((qcom_wave_pair_mode(1))) // 1=force single 2=force pair
 kernel void kernel_gemm_moe_q4_1_f32_ns(
@@ -110,7 +150,9 @@ kernel void kernel_gemm_moe_q4_1_f32_ns(
        __write_only image1d_buffer_t dst,
        __global     int *            total_tiles,
        uint ne00,
-        uint ne01
+        uint ne01,
+        uint is_ragged,
+        uint skip_gran
 ) {
    uint block_id_m = get_global_id(1); // m_tile
    uint block_id_n = get_global_id(2); // n_tile
@@ -120,6 +162,28 @@ kernel void kernel_gemm_moe_q4_1_f32_ns(
        return;
    }

+    // Ragged tile-skip: when is_ragged and the upper 16 token-slots of this tile are all
+    // padding (router 0xFFFFFFFF), skip the second (reg_c.hi) dotx16_reduce8 half -> ~half
+    // the GEMM dot for sparse tiles. Numerically identical (the skipped lanes are padding).
+    // Ragged tile-skip: tokens are packed contiguously per expert (moe_scatter fills
+    // lanes 0..V-1, moe_fill pre-pads the rest), so router padding (0xFFFFFFFF) is always
+    // trailing. Find the valid-token count V and round it UP to the skip granularity
+    // skip_gran (columns per skip-group: 8 = quarter, 16 = half/legacy, 32 = disabled).
+    // A 8-column group g is all-padding iff its first column (8*g) >= n_active, so its
+    // dotx8_reduce4 is skipped. Numerically identical (skipped lanes are padding).
+    uint n_active = TILESIZE_N;
+    if (is_ragged && skip_gran < TILESIZE_N) {
+        uint n_valid = TILESIZE_N;
+        for (uint _t = 0; _t < TILESIZE_N; ++_t) {
+            if (src2[block_id_n * TILESIZE_N + _t] == 0xFFFFFFFFu) { n_valid = _t; break; }
+        }
+        n_active = min((uint)TILESIZE_N, ((n_valid + skip_gran - 1) / skip_gran) * skip_gran);
+    }
+    // Group 0 (cols 0-7) always runs; groups 1-3 skip when fully padding.
+    bool skip_g1 = (8u  >= n_active);
+    bool skip_g2 = (16u >= n_active);
+    bool skip_g3 = (24u >= n_active);
+
    __private half16 reg_a;
    __private float32 reg_c = (float32)(0);
    __local half4 shared_b[128];
@@ -169,9 +233,11 @@ kernel void kernel_gemm_moe_q4_1_f32_ns(
        sub_group_barrier(CLK_LOCAL_MEM_FENCE);

        // 32 16x16 fp16 dot product with 8 elements reduction for better precision
-        half16 acc;
-        dotx16_reduce8(reg_a, shared_b, reg_c.lo, 0);
-        dotx16_reduce8(reg_a, shared_b, reg_c.hi, 16);
+        half8 acc8;
+        dotx8_reduce4(reg_a, shared_b, reg_c.lo.lo, 0);
+        if (!skip_g1) { dotx8_reduce4(reg_a, shared_b, reg_c.lo.hi, 8); }
+        if (!skip_g2) { dotx8_reduce4(reg_a, shared_b, reg_c.hi.lo, 16); }
+        if (!skip_g3) { dotx8_reduce4(reg_a, shared_b, reg_c.hi.hi, 24); }

        // Repeat for second sub-block
        uint half_step = step + TILESIZE_K;
@@ -196,8 +262,10 @@ kernel void kernel_gemm_moe_q4_1_f32_ns(
        sub_group_barrier(CLK_LOCAL_MEM_FENCE);

        // 32 16x16 fp16 dot product with 3-levels reduction for better precision
-        dotx16_reduce8(reg_a, shared_b, reg_c.lo, 0);
-        dotx16_reduce8(reg_a, shared_b, reg_c.hi, 16);
+        dotx8_reduce4(reg_a, shared_b, reg_c.lo.lo, 0);
+        if (!skip_g1) { dotx8_reduce4(reg_a, shared_b, reg_c.lo.hi, 8); }
+        if (!skip_g2) { dotx8_reduce4(reg_a, shared_b, reg_c.hi.lo, 16); }
+        if (!skip_g3) { dotx8_reduce4(reg_a, shared_b, reg_c.hi.hi, 24); }
    }

    if ((get_global_id(0) + block_id_m * TILESIZE_M) >= ne01) {
@@ -114,6 +114,46 @@ inline void get_scale_min_k4(
    c_reg.lo += convert_float8(acc.lo); \
    c_reg.hi += convert_float8(acc.hi); \

+// Quarter-tile variant: computes 8 output columns (one skip-group) into a float8
+// accumulator. Same reduction order / flush cadence as dotx16_reduce8, so the
+// non-skipped path is byte-identical; it just lets the caller skip empty
+// 8-column groups at finer granularity. Uses a private half8 `acc8`.
+#define dotx8_reduce4(a_reg, b_lm, c_reg, lm_offset) \
+    acc8.s0 = dot(a_reg.s0123, b_lm[lm_offset + 0]); \
+    acc8.s1 = dot(a_reg.s0123, b_lm[lm_offset + 1]); \
+    acc8.s2 = dot(a_reg.s0123, b_lm[lm_offset + 2]); \
+    acc8.s3 = dot(a_reg.s0123, b_lm[lm_offset + 3]); \
+    acc8.s4 = dot(a_reg.s0123, b_lm[lm_offset + 4]); \
+    acc8.s5 = dot(a_reg.s0123, b_lm[lm_offset + 5]); \
+    acc8.s6 = dot(a_reg.s0123, b_lm[lm_offset + 6]); \
+    acc8.s7 = dot(a_reg.s0123, b_lm[lm_offset + 7]); \
+    acc8.s0 += dot(a_reg.s4567, b_lm[lm_offset + 32]); \
+    acc8.s1 += dot(a_reg.s4567, b_lm[lm_offset + 33]); \
+    acc8.s2 += dot(a_reg.s4567, b_lm[lm_offset + 34]); \
+    acc8.s3 += dot(a_reg.s4567, b_lm[lm_offset + 35]); \
+    acc8.s4 += dot(a_reg.s4567, b_lm[lm_offset + 36]); \
+    acc8.s5 += dot(a_reg.s4567, b_lm[lm_offset + 37]); \
+    acc8.s6 += dot(a_reg.s4567, b_lm[lm_offset + 38]); \
+    acc8.s7 += dot(a_reg.s4567, b_lm[lm_offset + 39]); \
+    c_reg += convert_float8(acc8); \
+    acc8.s0 = dot(a_reg.s89ab, b_lm[lm_offset + 64]); \
+    acc8.s1 = dot(a_reg.s89ab, b_lm[lm_offset + 65]); \
+    acc8.s2 = dot(a_reg.s89ab, b_lm[lm_offset + 66]); \
+    acc8.s3 = dot(a_reg.s89ab, b_lm[lm_offset + 67]); \
+    acc8.s4 = dot(a_reg.s89ab, b_lm[lm_offset + 68]); \
+    acc8.s5 = dot(a_reg.s89ab, b_lm[lm_offset + 69]); \
+    acc8.s6 = dot(a_reg.s89ab, b_lm[lm_offset + 70]); \
+    acc8.s7 = dot(a_reg.s89ab, b_lm[lm_offset + 71]); \
+    acc8.s0 += dot(a_reg.scdef, b_lm[lm_offset + 96]); \
+    acc8.s1 += dot(a_reg.scdef, b_lm[lm_offset + 97]); \
+    acc8.s2 += dot(a_reg.scdef, b_lm[lm_offset + 98]); \
+    acc8.s3 += dot(a_reg.scdef, b_lm[lm_offset + 99]); \
+    acc8.s4 += dot(a_reg.scdef, b_lm[lm_offset + 100]); \
+    acc8.s5 += dot(a_reg.scdef, b_lm[lm_offset + 101]); \
+    acc8.s6 += dot(a_reg.scdef, b_lm[lm_offset + 102]); \
+    acc8.s7 += dot(a_reg.scdef, b_lm[lm_offset + 103]); \
+    c_reg += convert_float8(acc8); \
+

 __attribute__((qcom_wave_pair_mode(1)))
 kernel void kernel_gemm_moe_q4_k_f32_ns(
@@ -127,7 +167,9 @@ kernel void kernel_gemm_moe_q4_k_f32_ns(
        __write_only image1d_buffer_t dst,
        __global     int *            total_tiles,
        uint ne00,
-        uint ne01
+        uint ne01,
+        uint is_ragged,
+        uint skip_gran
 ) {
    uint block_id_m = get_global_id(1); // m_tile
    uint block_id_n = get_global_id(2); // n_tile
@@ -137,6 +179,25 @@ kernel void kernel_gemm_moe_q4_k_f32_ns(
        return;
    }

+    // Ragged tile-skip: tokens are packed contiguously per expert (moe_scatter fills
+    // lanes 0..V-1, moe_fill pre-pads the rest), so router padding (0xFFFFFFFF) is always
+    // trailing. Find the valid-token count V and round it UP to the skip granularity
+    // skip_gran (columns per skip-group: 8 = quarter, 16 = half/legacy, 32 = disabled).
+    // A 8-column group g is all-padding iff its first column (8*g) >= n_active, so its
+    // dotx8_reduce4 is skipped. Numerically identical (skipped lanes are padding).
+    uint n_active = TILESIZE_N;
+    if (is_ragged && skip_gran < TILESIZE_N) {
+        uint n_valid = TILESIZE_N;
+        for (uint _t = 0; _t < TILESIZE_N; ++_t) {
+            if (src2[block_id_n * TILESIZE_N + _t] == 0xFFFFFFFFu) { n_valid = _t; break; }
+        }
+        n_active = min((uint)TILESIZE_N, ((n_valid + skip_gran - 1) / skip_gran) * skip_gran);
+    }
+    // Group 0 (cols 0-7) always runs; groups 1-3 skip when fully padding.
+    bool skip_g1 = (8u  >= n_active);
+    bool skip_g2 = (16u >= n_active);
+    bool skip_g3 = (24u >= n_active);
+
    __private half16 reg_a;
    __private float32 reg_c = (float32)(0);
    __local half4 shared_b[128];
@@ -199,9 +260,11 @@ kernel void kernel_gemm_moe_q4_k_f32_ns(

        sub_group_barrier(CLK_LOCAL_MEM_FENCE);

-        half16 acc;
-        dotx16_reduce8(reg_a, shared_b, reg_c.lo, 0);
-        dotx16_reduce8(reg_a, shared_b, reg_c.hi, 16);
+        half8 acc8;
+        dotx8_reduce4(reg_a, shared_b, reg_c.lo.lo, 0);
+        if (!skip_g1) { dotx8_reduce4(reg_a, shared_b, reg_c.lo.hi, 8); }
+        if (!skip_g2) { dotx8_reduce4(reg_a, shared_b, reg_c.hi.lo, 16); }
+        if (!skip_g3) { dotx8_reduce4(reg_a, shared_b, reg_c.hi.hi, 24); }

        // Second half (next 16 elements, same sub-block scale)
        uint half_step = step + TILESIZE_K;
@@ -221,8 +284,10 @@ kernel void kernel_gemm_moe_q4_k_f32_ns(

        sub_group_barrier(CLK_LOCAL_MEM_FENCE);

-        dotx16_reduce8(reg_a, shared_b, reg_c.lo, 0);
-        dotx16_reduce8(reg_a, shared_b, reg_c.hi, 16);
+        dotx8_reduce4(reg_a, shared_b, reg_c.lo.lo, 0);
+        if (!skip_g1) { dotx8_reduce4(reg_a, shared_b, reg_c.lo.hi, 8); }
+        if (!skip_g2) { dotx8_reduce4(reg_a, shared_b, reg_c.hi.lo, 16); }
+        if (!skip_g3) { dotx8_reduce4(reg_a, shared_b, reg_c.hi.hi, 24); }
    }

    if ((get_global_id(0) + block_id_m * TILESIZE_M) >= ne01) {
@@ -98,6 +98,46 @@
    c_reg.lo += convert_float8(acc.lo); \
    c_reg.hi += convert_float8(acc.hi); \

+// Quarter-tile variant: computes 8 output columns (one skip-group) into a float8
+// accumulator. Same reduction order / flush cadence as dotx16_reduce8, so the
+// non-skipped path is byte-identical; it just lets the caller skip empty
+// 8-column groups at finer granularity. Uses a private half8 `acc8`.
+#define dotx8_reduce4(a_reg, b_lm, c_reg, lm_offset) \
+    acc8.s0 = dot(a_reg.s0123, b_lm[lm_offset + 0]); \
+    acc8.s1 = dot(a_reg.s0123, b_lm[lm_offset + 1]); \
+    acc8.s2 = dot(a_reg.s0123, b_lm[lm_offset + 2]); \
+    acc8.s3 = dot(a_reg.s0123, b_lm[lm_offset + 3]); \
+    acc8.s4 = dot(a_reg.s0123, b_lm[lm_offset + 4]); \
+    acc8.s5 = dot(a_reg.s0123, b_lm[lm_offset + 5]); \
+    acc8.s6 = dot(a_reg.s0123, b_lm[lm_offset + 6]); \
+    acc8.s7 = dot(a_reg.s0123, b_lm[lm_offset + 7]); \
+    acc8.s0 += dot(a_reg.s4567, b_lm[lm_offset + 32]); \
+    acc8.s1 += dot(a_reg.s4567, b_lm[lm_offset + 33]); \
+    acc8.s2 += dot(a_reg.s4567, b_lm[lm_offset + 34]); \
+    acc8.s3 += dot(a_reg.s4567, b_lm[lm_offset + 35]); \
+    acc8.s4 += dot(a_reg.s4567, b_lm[lm_offset + 36]); \
+    acc8.s5 += dot(a_reg.s4567, b_lm[lm_offset + 37]); \
+    acc8.s6 += dot(a_reg.s4567, b_lm[lm_offset + 38]); \
+    acc8.s7 += dot(a_reg.s4567, b_lm[lm_offset + 39]); \
+    c_reg += convert_float8(acc8); \
+    acc8.s0 = dot(a_reg.s89ab, b_lm[lm_offset + 64]); \
+    acc8.s1 = dot(a_reg.s89ab, b_lm[lm_offset + 65]); \
+    acc8.s2 = dot(a_reg.s89ab, b_lm[lm_offset + 66]); \
+    acc8.s3 = dot(a_reg.s89ab, b_lm[lm_offset + 67]); \
+    acc8.s4 = dot(a_reg.s89ab, b_lm[lm_offset + 68]); \
+    acc8.s5 = dot(a_reg.s89ab, b_lm[lm_offset + 69]); \
+    acc8.s6 = dot(a_reg.s89ab, b_lm[lm_offset + 70]); \
+    acc8.s7 = dot(a_reg.s89ab, b_lm[lm_offset + 71]); \
+    acc8.s0 += dot(a_reg.scdef, b_lm[lm_offset + 96]); \
+    acc8.s1 += dot(a_reg.scdef, b_lm[lm_offset + 97]); \
+    acc8.s2 += dot(a_reg.scdef, b_lm[lm_offset + 98]); \
+    acc8.s3 += dot(a_reg.scdef, b_lm[lm_offset + 99]); \
+    acc8.s4 += dot(a_reg.scdef, b_lm[lm_offset + 100]); \
+    acc8.s5 += dot(a_reg.scdef, b_lm[lm_offset + 101]); \
+    acc8.s6 += dot(a_reg.scdef, b_lm[lm_offset + 102]); \
+    acc8.s7 += dot(a_reg.scdef, b_lm[lm_offset + 103]); \
+    c_reg += convert_float8(acc8); \
+

 __attribute__((qcom_wave_pair_mode(1))) // 1=force single 2=force pair
 kernel void kernel_gemm_moe_q5_0_f32_ns(
@@ -110,7 +150,9 @@ kernel void kernel_gemm_moe_q5_0_f32_ns(
        __write_only image1d_buffer_t dst,
        __global     int *            total_tiles,
        uint ne00,
-        uint ne01
+        uint ne01,
+        uint is_ragged,
+        uint skip_gran
 ) {
    uint block_id_m = get_global_id(1); // m_tile
    uint block_id_n = get_global_id(2); // n_tile
@@ -120,6 +162,28 @@ kernel void kernel_gemm_moe_q5_0_f32_ns(
        return;
    }

+    // Ragged tile-skip: when is_ragged and the upper 16 token-slots of this tile are all
+    // padding (router 0xFFFFFFFF), skip the second (reg_c.hi) dotx16_reduce8 half -> ~half
+    // the GEMM dot for sparse tiles. Numerically identical (the skipped lanes are padding).
+    // Ragged tile-skip: tokens are packed contiguously per expert (moe_scatter fills
+    // lanes 0..V-1, moe_fill pre-pads the rest), so router padding (0xFFFFFFFF) is always
+    // trailing. Find the valid-token count V and round it UP to the skip granularity
+    // skip_gran (columns per skip-group: 8 = quarter, 16 = half/legacy, 32 = disabled).
+    // A 8-column group g is all-padding iff its first column (8*g) >= n_active, so its
+    // dotx8_reduce4 is skipped. Numerically identical (skipped lanes are padding).
+    uint n_active = TILESIZE_N;
+    if (is_ragged && skip_gran < TILESIZE_N) {
+        uint n_valid = TILESIZE_N;
+        for (uint _t = 0; _t < TILESIZE_N; ++_t) {
+            if (src2[block_id_n * TILESIZE_N + _t] == 0xFFFFFFFFu) { n_valid = _t; break; }
+        }
+        n_active = min((uint)TILESIZE_N, ((n_valid + skip_gran - 1) / skip_gran) * skip_gran);
+    }
+    // Group 0 (cols 0-7) always runs; groups 1-3 skip when fully padding.
+    bool skip_g1 = (8u  >= n_active);
+    bool skip_g2 = (16u >= n_active);
+    bool skip_g3 = (24u >= n_active);
+
    __private half16 reg_a;
    __private float32 reg_c = (float32)(0);
    __local half4 shared_b[128];
@@ -171,9 +235,11 @@ kernel void kernel_gemm_moe_q5_0_f32_ns(
        sub_group_barrier(CLK_LOCAL_MEM_FENCE);

        // 32 16x16 fp16 dot product with 8 elements reduction for better precision
-        half16 acc;
-        dotx16_reduce8(reg_a, shared_b, reg_c.lo, 0);
-        dotx16_reduce8(reg_a, shared_b, reg_c.hi, 16);
+        half8 acc8;
+        dotx8_reduce4(reg_a, shared_b, reg_c.lo.lo, 0);
+        if (!skip_g1) { dotx8_reduce4(reg_a, shared_b, reg_c.lo.hi, 8); }
+        if (!skip_g2) { dotx8_reduce4(reg_a, shared_b, reg_c.hi.lo, 16); }
+        if (!skip_g3) { dotx8_reduce4(reg_a, shared_b, reg_c.hi.hi, 24); }

        // Repeat for second sub-block
        uint half_step = step + TILESIZE_K;
@@ -198,8 +264,10 @@ kernel void kernel_gemm_moe_q5_0_f32_ns(
        sub_group_barrier(CLK_LOCAL_MEM_FENCE);

        // 32 16x16 fp16 dot product with 3-levels reduction for better precision
-        dotx16_reduce8(reg_a, shared_b, reg_c.lo, 0);
-        dotx16_reduce8(reg_a, shared_b, reg_c.hi, 16);
+        dotx8_reduce4(reg_a, shared_b, reg_c.lo.lo, 0);
+        if (!skip_g1) { dotx8_reduce4(reg_a, shared_b, reg_c.lo.hi, 8); }
+        if (!skip_g2) { dotx8_reduce4(reg_a, shared_b, reg_c.hi.lo, 16); }
+        if (!skip_g3) { dotx8_reduce4(reg_a, shared_b, reg_c.hi.hi, 24); }
    }

    if ((get_global_id(0) + block_id_m * TILESIZE_M) >= ne01) {
@@ -98,6 +98,46 @@
    c_reg.lo += convert_float8(acc.lo); \
    c_reg.hi += convert_float8(acc.hi); \

+// Quarter-tile variant: computes 8 output columns (one skip-group) into a float8
+// accumulator. Same reduction order / flush cadence as dotx16_reduce8, so the
+// non-skipped path is byte-identical; it just lets the caller skip empty
+// 8-column groups at finer granularity. Uses a private half8 `acc8`.
+#define dotx8_reduce4(a_reg, b_lm, c_reg, lm_offset) \
+    acc8.s0 = dot(a_reg.s0123, b_lm[lm_offset + 0]); \
+    acc8.s1 = dot(a_reg.s0123, b_lm[lm_offset + 1]); \
+    acc8.s2 = dot(a_reg.s0123, b_lm[lm_offset + 2]); \
+    acc8.s3 = dot(a_reg.s0123, b_lm[lm_offset + 3]); \
+    acc8.s4 = dot(a_reg.s0123, b_lm[lm_offset + 4]); \
+    acc8.s5 = dot(a_reg.s0123, b_lm[lm_offset + 5]); \
+    acc8.s6 = dot(a_reg.s0123, b_lm[lm_offset + 6]); \
+    acc8.s7 = dot(a_reg.s0123, b_lm[lm_offset + 7]); \
+    acc8.s0 += dot(a_reg.s4567, b_lm[lm_offset + 32]); \
+    acc8.s1 += dot(a_reg.s4567, b_lm[lm_offset + 33]); \
+    acc8.s2 += dot(a_reg.s4567, b_lm[lm_offset + 34]); \
+    acc8.s3 += dot(a_reg.s4567, b_lm[lm_offset + 35]); \
+    acc8.s4 += dot(a_reg.s4567, b_lm[lm_offset + 36]); \
+    acc8.s5 += dot(a_reg.s4567, b_lm[lm_offset + 37]); \
+    acc8.s6 += dot(a_reg.s4567, b_lm[lm_offset + 38]); \
+    acc8.s7 += dot(a_reg.s4567, b_lm[lm_offset + 39]); \
+    c_reg += convert_float8(acc8); \
+    acc8.s0 = dot(a_reg.s89ab, b_lm[lm_offset + 64]); \
+    acc8.s1 = dot(a_reg.s89ab, b_lm[lm_offset + 65]); \
+    acc8.s2 = dot(a_reg.s89ab, b_lm[lm_offset + 66]); \
+    acc8.s3 = dot(a_reg.s89ab, b_lm[lm_offset + 67]); \
+    acc8.s4 = dot(a_reg.s89ab, b_lm[lm_offset + 68]); \
+    acc8.s5 = dot(a_reg.s89ab, b_lm[lm_offset + 69]); \
+    acc8.s6 = dot(a_reg.s89ab, b_lm[lm_offset + 70]); \
+    acc8.s7 = dot(a_reg.s89ab, b_lm[lm_offset + 71]); \
+    acc8.s0 += dot(a_reg.scdef, b_lm[lm_offset + 96]); \
+    acc8.s1 += dot(a_reg.scdef, b_lm[lm_offset + 97]); \
+    acc8.s2 += dot(a_reg.scdef, b_lm[lm_offset + 98]); \
+    acc8.s3 += dot(a_reg.scdef, b_lm[lm_offset + 99]); \
+    acc8.s4 += dot(a_reg.scdef, b_lm[lm_offset + 100]); \
+    acc8.s5 += dot(a_reg.scdef, b_lm[lm_offset + 101]); \
+    acc8.s6 += dot(a_reg.scdef, b_lm[lm_offset + 102]); \
+    acc8.s7 += dot(a_reg.scdef, b_lm[lm_offset + 103]); \
+    c_reg += convert_float8(acc8); \
+

 __attribute__((qcom_wave_pair_mode(1))) // 1=force single 2=force pair
 kernel void kernel_gemm_moe_q5_1_f32_ns(
@@ -111,7 +151,9 @@ kernel void kernel_gemm_moe_q5_1_f32_ns(
        __write_only image1d_buffer_t dst,
        __global     int *            total_tiles,
        uint ne00,
-        uint ne01
+        uint ne01,
+        uint is_ragged,
+        uint skip_gran
 ) {
    uint block_id_m = get_global_id(1); // m_tile
    uint block_id_n = get_global_id(2); // n_tile
@@ -121,6 +163,28 @@ kernel void kernel_gemm_moe_q5_1_f32_ns(
        return;
    }

+    // Ragged tile-skip: when is_ragged and the upper 16 token-slots of this tile are all
+    // padding (router 0xFFFFFFFF), skip the second (reg_c.hi) dotx16_reduce8 half -> ~half
+    // the GEMM dot for sparse tiles. Numerically identical (the skipped lanes are padding).
+    // Ragged tile-skip: tokens are packed contiguously per expert (moe_scatter fills
+    // lanes 0..V-1, moe_fill pre-pads the rest), so router padding (0xFFFFFFFF) is always
+    // trailing. Find the valid-token count V and round it UP to the skip granularity
+    // skip_gran (columns per skip-group: 8 = quarter, 16 = half/legacy, 32 = disabled).
+    // A 8-column group g is all-padding iff its first column (8*g) >= n_active, so its
+    // dotx8_reduce4 is skipped. Numerically identical (skipped lanes are padding).
+    uint n_active = TILESIZE_N;
+    if (is_ragged && skip_gran < TILESIZE_N) {
+        uint n_valid = TILESIZE_N;
+        for (uint _t = 0; _t < TILESIZE_N; ++_t) {
+            if (src2[block_id_n * TILESIZE_N + _t] == 0xFFFFFFFFu) { n_valid = _t; break; }
+        }
+        n_active = min((uint)TILESIZE_N, ((n_valid + skip_gran - 1) / skip_gran) * skip_gran);
+    }
+    // Group 0 (cols 0-7) always runs; groups 1-3 skip when fully padding.
+    bool skip_g1 = (8u  >= n_active);
+    bool skip_g2 = (16u >= n_active);
+    bool skip_g3 = (24u >= n_active);
+
    __private half16 reg_a;
    __private float32 reg_c = (float32)(0);
    __local half4 shared_b[128];
@@ -173,9 +237,11 @@ kernel void kernel_gemm_moe_q5_1_f32_ns(
        sub_group_barrier(CLK_LOCAL_MEM_FENCE);

        // 32 16x16 fp16 dot product with 8 elements reduction for better precision
-        half16 acc;
-        dotx16_reduce8(reg_a, shared_b, reg_c.lo, 0);
-        dotx16_reduce8(reg_a, shared_b, reg_c.hi, 16);
+        half8 acc8;
+        dotx8_reduce4(reg_a, shared_b, reg_c.lo.lo, 0);
+        if (!skip_g1) { dotx8_reduce4(reg_a, shared_b, reg_c.lo.hi, 8); }
+        if (!skip_g2) { dotx8_reduce4(reg_a, shared_b, reg_c.hi.lo, 16); }
+        if (!skip_g3) { dotx8_reduce4(reg_a, shared_b, reg_c.hi.hi, 24); }

        // Repeat for second sub-block
        uint half_step = step + TILESIZE_K;
@@ -200,8 +266,10 @@ kernel void kernel_gemm_moe_q5_1_f32_ns(
        sub_group_barrier(CLK_LOCAL_MEM_FENCE);

        // 32 16x16 fp16 dot product with 3-levels reduction for better precision
-        dotx16_reduce8(reg_a, shared_b, reg_c.lo, 0);
-        dotx16_reduce8(reg_a, shared_b, reg_c.hi, 16);
+        dotx8_reduce4(reg_a, shared_b, reg_c.lo.lo, 0);
+        if (!skip_g1) { dotx8_reduce4(reg_a, shared_b, reg_c.lo.hi, 8); }
+        if (!skip_g2) { dotx8_reduce4(reg_a, shared_b, reg_c.hi.lo, 16); }
+        if (!skip_g3) { dotx8_reduce4(reg_a, shared_b, reg_c.hi.hi, 24); }
    }

    if ((get_global_id(0) + block_id_m * TILESIZE_M) >= ne01) {
@@ -114,6 +114,46 @@ inline void get_scale_min_k4(
    c_reg.lo += convert_float8(acc.lo); \
    c_reg.hi += convert_float8(acc.hi); \

+// Quarter-tile variant: computes 8 output columns (one skip-group) into a float8
+// accumulator. Same reduction order / flush cadence as dotx16_reduce8, so the
+// non-skipped path is byte-identical; it just lets the caller skip empty
+// 8-column groups at finer granularity. Uses a private half8 `acc8`.
+#define dotx8_reduce4(a_reg, b_lm, c_reg, lm_offset) \
+    acc8.s0 = dot(a_reg.s0123, b_lm[lm_offset + 0]); \
+    acc8.s1 = dot(a_reg.s0123, b_lm[lm_offset + 1]); \
+    acc8.s2 = dot(a_reg.s0123, b_lm[lm_offset + 2]); \
+    acc8.s3 = dot(a_reg.s0123, b_lm[lm_offset + 3]); \
+    acc8.s4 = dot(a_reg.s0123, b_lm[lm_offset + 4]); \
+    acc8.s5 = dot(a_reg.s0123, b_lm[lm_offset + 5]); \
+    acc8.s6 = dot(a_reg.s0123, b_lm[lm_offset + 6]); \
+    acc8.s7 = dot(a_reg.s0123, b_lm[lm_offset + 7]); \
+    acc8.s0 += dot(a_reg.s4567, b_lm[lm_offset + 32]); \
+    acc8.s1 += dot(a_reg.s4567, b_lm[lm_offset + 33]); \
+    acc8.s2 += dot(a_reg.s4567, b_lm[lm_offset + 34]); \
+    acc8.s3 += dot(a_reg.s4567, b_lm[lm_offset + 35]); \
+    acc8.s4 += dot(a_reg.s4567, b_lm[lm_offset + 36]); \
+    acc8.s5 += dot(a_reg.s4567, b_lm[lm_offset + 37]); \
+    acc8.s6 += dot(a_reg.s4567, b_lm[lm_offset + 38]); \
+    acc8.s7 += dot(a_reg.s4567, b_lm[lm_offset + 39]); \
+    c_reg += convert_float8(acc8); \
+    acc8.s0 = dot(a_reg.s89ab, b_lm[lm_offset + 64]); \
+    acc8.s1 = dot(a_reg.s89ab, b_lm[lm_offset + 65]); \
+    acc8.s2 = dot(a_reg.s89ab, b_lm[lm_offset + 66]); \
+    acc8.s3 = dot(a_reg.s89ab, b_lm[lm_offset + 67]); \
+    acc8.s4 = dot(a_reg.s89ab, b_lm[lm_offset + 68]); \
+    acc8.s5 = dot(a_reg.s89ab, b_lm[lm_offset + 69]); \
+    acc8.s6 = dot(a_reg.s89ab, b_lm[lm_offset + 70]); \
+    acc8.s7 = dot(a_reg.s89ab, b_lm[lm_offset + 71]); \
+    acc8.s0 += dot(a_reg.scdef, b_lm[lm_offset + 96]); \
+    acc8.s1 += dot(a_reg.scdef, b_lm[lm_offset + 97]); \
+    acc8.s2 += dot(a_reg.scdef, b_lm[lm_offset + 98]); \
+    acc8.s3 += dot(a_reg.scdef, b_lm[lm_offset + 99]); \
+    acc8.s4 += dot(a_reg.scdef, b_lm[lm_offset + 100]); \
+    acc8.s5 += dot(a_reg.scdef, b_lm[lm_offset + 101]); \
+    acc8.s6 += dot(a_reg.scdef, b_lm[lm_offset + 102]); \
+    acc8.s7 += dot(a_reg.scdef, b_lm[lm_offset + 103]); \
+    c_reg += convert_float8(acc8); \
+

 __attribute__((qcom_wave_pair_mode(1)))
 kernel void kernel_gemm_moe_q5_k_f32_ns(
@@ -128,7 +168,9 @@ kernel void kernel_gemm_moe_q5_k_f32_ns(
        __write_only image1d_buffer_t dst,
        __global     int *            total_tiles,
        uint ne00,
-        uint ne01
+        uint ne01,
+        uint is_ragged,
+        uint skip_gran
 ) {
    uint block_id_m = get_global_id(1); // m_tile
    uint block_id_n = get_global_id(2); // n_tile
@@ -138,6 +180,28 @@ kernel void kernel_gemm_moe_q5_k_f32_ns(
        return;
    }

+    // Ragged tile-skip: when is_ragged and the upper 16 token-slots of this tile are all
+    // padding (router 0xFFFFFFFF), skip the second (reg_c.hi) dotx16_reduce8 half -> ~half
+    // the GEMM dot for sparse tiles. Numerically identical (the skipped lanes are padding).
+    // Ragged tile-skip: tokens are packed contiguously per expert (moe_scatter fills
+    // lanes 0..V-1, moe_fill pre-pads the rest), so router padding (0xFFFFFFFF) is always
+    // trailing. Find the valid-token count V and round it UP to the skip granularity
+    // skip_gran (columns per skip-group: 8 = quarter, 16 = half/legacy, 32 = disabled).
+    // A 8-column group g is all-padding iff its first column (8*g) >= n_active, so its
+    // dotx8_reduce4 is skipped. Numerically identical (skipped lanes are padding).
+    uint n_active = TILESIZE_N;
+    if (is_ragged && skip_gran < TILESIZE_N) {
+        uint n_valid = TILESIZE_N;
+        for (uint _t = 0; _t < TILESIZE_N; ++_t) {
+            if (src2[block_id_n * TILESIZE_N + _t] == 0xFFFFFFFFu) { n_valid = _t; break; }
+        }
+        n_active = min((uint)TILESIZE_N, ((n_valid + skip_gran - 1) / skip_gran) * skip_gran);
+    }
+    // Group 0 (cols 0-7) always runs; groups 1-3 skip when fully padding.
+    bool skip_g1 = (8u  >= n_active);
+    bool skip_g2 = (16u >= n_active);
+    bool skip_g3 = (24u >= n_active);
+
    __private half16 reg_a;
    __private float32 reg_c = (float32)(0);
    __local half4 shared_b[128];
@@ -204,9 +268,11 @@ kernel void kernel_gemm_moe_q5_k_f32_ns(

        sub_group_barrier(CLK_LOCAL_MEM_FENCE);

-        half16 acc;
-        dotx16_reduce8(reg_a, shared_b, reg_c.lo, 0);
-        dotx16_reduce8(reg_a, shared_b, reg_c.hi, 16);
+        half8 acc8;
+        dotx8_reduce4(reg_a, shared_b, reg_c.lo.lo, 0);
+        if (!skip_g1) { dotx8_reduce4(reg_a, shared_b, reg_c.lo.hi, 8); }
+        if (!skip_g2) { dotx8_reduce4(reg_a, shared_b, reg_c.hi.lo, 16); }
+        if (!skip_g3) { dotx8_reduce4(reg_a, shared_b, reg_c.hi.hi, 24); }

        // Second half
        uint half_step = step + TILESIZE_K;
@@ -226,8 +292,10 @@ kernel void kernel_gemm_moe_q5_k_f32_ns(

        sub_group_barrier(CLK_LOCAL_MEM_FENCE);

-        dotx16_reduce8(reg_a, shared_b, reg_c.lo, 0);
-        dotx16_reduce8(reg_a, shared_b, reg_c.hi, 16);
+        dotx8_reduce4(reg_a, shared_b, reg_c.lo.lo, 0);
+        if (!skip_g1) { dotx8_reduce4(reg_a, shared_b, reg_c.lo.hi, 8); }
+        if (!skip_g2) { dotx8_reduce4(reg_a, shared_b, reg_c.hi.lo, 16); }
+        if (!skip_g3) { dotx8_reduce4(reg_a, shared_b, reg_c.hi.hi, 24); }
    }

    if ((get_global_id(0) + block_id_m * TILESIZE_M) >= ne01) {
@@ -98,6 +98,46 @@
    c_reg.lo += convert_float8(acc.lo); \
    c_reg.hi += convert_float8(acc.hi); \

+// Quarter-tile variant: computes 8 output columns (one skip-group) into a float8
+// accumulator. Same reduction order / flush cadence as dotx16_reduce8, so the
+// non-skipped path is byte-identical; it just lets the caller skip empty
+// 8-column groups at finer granularity. Uses a private half8 `acc8`.
+#define dotx8_reduce4(a_reg, b_lm, c_reg, lm_offset) \
+    acc8.s0 = dot(a_reg.s0123, b_lm[lm_offset + 0]); \
+    acc8.s1 = dot(a_reg.s0123, b_lm[lm_offset + 1]); \
+    acc8.s2 = dot(a_reg.s0123, b_lm[lm_offset + 2]); \
+    acc8.s3 = dot(a_reg.s0123, b_lm[lm_offset + 3]); \
+    acc8.s4 = dot(a_reg.s0123, b_lm[lm_offset + 4]); \
+    acc8.s5 = dot(a_reg.s0123, b_lm[lm_offset + 5]); \
+    acc8.s6 = dot(a_reg.s0123, b_lm[lm_offset + 6]); \
+    acc8.s7 = dot(a_reg.s0123, b_lm[lm_offset + 7]); \
+    acc8.s0 += dot(a_reg.s4567, b_lm[lm_offset + 32]); \
+    acc8.s1 += dot(a_reg.s4567, b_lm[lm_offset + 33]); \
+    acc8.s2 += dot(a_reg.s4567, b_lm[lm_offset + 34]); \
+    acc8.s3 += dot(a_reg.s4567, b_lm[lm_offset + 35]); \
+    acc8.s4 += dot(a_reg.s4567, b_lm[lm_offset + 36]); \
+    acc8.s5 += dot(a_reg.s4567, b_lm[lm_offset + 37]); \
+    acc8.s6 += dot(a_reg.s4567, b_lm[lm_offset + 38]); \
+    acc8.s7 += dot(a_reg.s4567, b_lm[lm_offset + 39]); \
+    c_reg += convert_float8(acc8); \
+    acc8.s0 = dot(a_reg.s89ab, b_lm[lm_offset + 64]); \
+    acc8.s1 = dot(a_reg.s89ab, b_lm[lm_offset + 65]); \
+    acc8.s2 = dot(a_reg.s89ab, b_lm[lm_offset + 66]); \
+    acc8.s3 = dot(a_reg.s89ab, b_lm[lm_offset + 67]); \
+    acc8.s4 = dot(a_reg.s89ab, b_lm[lm_offset + 68]); \
+    acc8.s5 = dot(a_reg.s89ab, b_lm[lm_offset + 69]); \
+    acc8.s6 = dot(a_reg.s89ab, b_lm[lm_offset + 70]); \
+    acc8.s7 = dot(a_reg.s89ab, b_lm[lm_offset + 71]); \
+    acc8.s0 += dot(a_reg.scdef, b_lm[lm_offset + 96]); \
+    acc8.s1 += dot(a_reg.scdef, b_lm[lm_offset + 97]); \
+    acc8.s2 += dot(a_reg.scdef, b_lm[lm_offset + 98]); \
+    acc8.s3 += dot(a_reg.scdef, b_lm[lm_offset + 99]); \
+    acc8.s4 += dot(a_reg.scdef, b_lm[lm_offset + 100]); \
+    acc8.s5 += dot(a_reg.scdef, b_lm[lm_offset + 101]); \
+    acc8.s6 += dot(a_reg.scdef, b_lm[lm_offset + 102]); \
+    acc8.s7 += dot(a_reg.scdef, b_lm[lm_offset + 103]); \
+    c_reg += convert_float8(acc8); \
+

 __attribute__((qcom_wave_pair_mode(1)))
 kernel void kernel_gemm_moe_q6_k_f32_ns(
@@ -111,7 +151,9 @@ kernel void kernel_gemm_moe_q6_k_f32_ns(
        __write_only image1d_buffer_t dst,
        __global     int *            total_tiles,
        uint ne00,
-        uint ne01
+        uint ne01,
+        uint is_ragged,
+        uint skip_gran
 ) {
    uint block_id_m = get_global_id(1); // m_tile
    uint block_id_n = get_global_id(2); // n_tile
@@ -121,6 +163,28 @@ kernel void kernel_gemm_moe_q6_k_f32_ns(
        return;
    }

+    // Ragged tile-skip: when is_ragged and the upper 16 token-slots of this tile are all
+    // padding (router 0xFFFFFFFF), skip the second (reg_c.hi) dotx16_reduce8 half -> ~half
+    // the GEMM dot for sparse tiles. Numerically identical (the skipped lanes are padding).
+    // Ragged tile-skip: tokens are packed contiguously per expert (moe_scatter fills
+    // lanes 0..V-1, moe_fill pre-pads the rest), so router padding (0xFFFFFFFF) is always
+    // trailing. Find the valid-token count V and round it UP to the skip granularity
+    // skip_gran (columns per skip-group: 8 = quarter, 16 = half/legacy, 32 = disabled).
+    // A 8-column group g is all-padding iff its first column (8*g) >= n_active, so its
+    // dotx8_reduce4 is skipped. Numerically identical (skipped lanes are padding).
+    uint n_active = TILESIZE_N;
+    if (is_ragged && skip_gran < TILESIZE_N) {
+        uint n_valid = TILESIZE_N;
+        for (uint _t = 0; _t < TILESIZE_N; ++_t) {
+            if (src2[block_id_n * TILESIZE_N + _t] == 0xFFFFFFFFu) { n_valid = _t; break; }
+        }
+        n_active = min((uint)TILESIZE_N, ((n_valid + skip_gran - 1) / skip_gran) * skip_gran);
+    }
+    // Group 0 (cols 0-7) always runs; groups 1-3 skip when fully padding.
+    bool skip_g1 = (8u  >= n_active);
+    bool skip_g2 = (16u >= n_active);
+    bool skip_g3 = (24u >= n_active);
+
    __private half16 reg_a;
    __private float32 reg_c = (float32)(0);
    __local half4 shared_b[128];
@@ -183,9 +247,11 @@ kernel void kernel_gemm_moe_q6_k_f32_ns(

        sub_group_barrier(CLK_LOCAL_MEM_FENCE);

-        half16 acc;
-        dotx16_reduce8(reg_a, shared_b, reg_c.lo, 0);
-        dotx16_reduce8(reg_a, shared_b, reg_c.hi, 16);
+        half8 acc8;
+        dotx8_reduce4(reg_a, shared_b, reg_c.lo.lo, 0);
+        if (!skip_g1) { dotx8_reduce4(reg_a, shared_b, reg_c.lo.hi, 8); }
+        if (!skip_g2) { dotx8_reduce4(reg_a, shared_b, reg_c.hi.lo, 16); }
+        if (!skip_g3) { dotx8_reduce4(reg_a, shared_b, reg_c.hi.hi, 24); }

        // Second half
        uint half_step = step + TILESIZE_K;
@@ -205,8 +271,10 @@ kernel void kernel_gemm_moe_q6_k_f32_ns(

        sub_group_barrier(CLK_LOCAL_MEM_FENCE);

-        dotx16_reduce8(reg_a, shared_b, reg_c.lo, 0);
-        dotx16_reduce8(reg_a, shared_b, reg_c.hi, 16);
+        dotx8_reduce4(reg_a, shared_b, reg_c.lo.lo, 0);
+        if (!skip_g1) { dotx8_reduce4(reg_a, shared_b, reg_c.lo.hi, 8); }
+        if (!skip_g2) { dotx8_reduce4(reg_a, shared_b, reg_c.hi.lo, 16); }
+        if (!skip_g3) { dotx8_reduce4(reg_a, shared_b, reg_c.hi.hi, 24); }
    }

    if ((get_global_id(0) + block_id_m * TILESIZE_M) >= ne01) {
@@ -10310,7 +10310,8 @@ static void ggml_vk_flash_attn(ggml_backend_vk_context * ctx, vk_context& subctx
    }

    // Only use mask opt when the mask is fairly large. This hasn't been tuned extensively.
-    bool use_mask_opt = mask && nem1 >= 32 && nem0 * nem1 > 32768 && nem0 >= tuning_params.block_cols * 16;
+    bool use_mask_opt = mask && nem1 >= 32 && nem0 * nem1 > 32768 && nem0 >= tuning_params.block_cols * 16
+                        && (ctx->device->architecture != vk_device_architecture::AMD_GCN || HSK > 256 || HSV > 256);
    vk_fa_pipeline_state fa_pipeline_state = get_fa_pipeline_state(ctx->device, tuning_params, HSK, HSV, aligned, f32acc,
                                                                   mask != nullptr, use_mask_opt, logit_softcap != 0, k->type, v->type);

@@ -16308,7 +16309,18 @@ static ggml_status ggml_backend_vk_graph_compute(ggml_backend_t backend, ggml_cg
    uint32_t submit_count = 0;
    uint64_t batch_flops = 0;
    uint64_t total_flops = 0;
-    uint64_t flops_per_submit = std::min(uint64_t(200'000'000'000), ctx->last_total_flops / 40u);
+    uint64_t flops_cap = 200'000'000'000ULL;
+
+    // On weaker AMD GPUs larger submissions can hit a driver timeout, submit more often to avoid this
+    if (ctx->device->vendor_id == VK_VENDOR_ID_AMD && ctx->device->shader_core_count > 0) {
+        if (ctx->device->architecture == AMD_GCN && ctx->device->shader_core_count < 32) {
+            flops_cap = 500'000'000ULL * ctx->device->shader_core_count;
+        } else if (ctx->device->architecture != AMD_GCN && ctx->device->shader_core_count < 24) {
+            flops_cap = 2'000'000'000ULL * ctx->device->shader_core_count;
+        }
+    }
+    uint64_t flops_per_submit = std::min(flops_cap, ctx->last_total_flops / 40u);
+
    for (int i = 0; i < cgraph->n_nodes; i++) {
        if (first_node_in_batch) {
            submit_node_idx = i;
@@ -379,6 +379,8 @@ bool llama_batch_allocr::init(
                    LLAMA_LOG_ERROR("%s: sequence %d positions are decreasing (not allowed)\n", __func__, seq_id);
                    return false;
                }
+
+                cur_seq_pos[seq_id] = pos;
            }
        }
    }
Author	SHA1	Message	Date
Hongqiang Wang	92366df30d	opencl: Q6_K GEMM/GEMV fix for ne01 of weights that are not multiples of 128. (#25464 ) * opencl: fix garbled output for Q6_K weights with ne01 % 128 != 0 on Adreno Observed with granite-3.1-3b-a800m-instruct, whose vocab is an odd number. Route Q6_K dense mul_mat with ne01 % 128 != 0 off the noshuffle path: decode (ne1==1) uses the correct flat GEMV and the matching GEMM (ne1>1) falls back to CPU (the flat convert has no verified small-batch GEMM kernel for these shapes). All standard hidden/FFN/vocab dims are multiples of 128 and keep the noshuffle path. * opencl: reserve alignment slack for the SOA subbuffer carve in alloc size set_tensor carves quantized weights into per-component subbuffers (d/q, ql/qh/s/d, ...) whose origins are each rounded up to the device base address alignment. When a component's size is not a multiple of the alignment, the carve extends past ggml_nbytes(tensor) and the last subbuffer overlaps the next tensor in the pool -- e.g. q6_K [1536, 49155]: size_s = 4915596 ends 32 bytes past a 128-byte boundary, so the d subbuffer ends 96 bytes past the tensor's allocation, and whichever of the two neighboring tensors is uploaded last silently corrupts the other (here: the last vocab rows' block scales). This affects any quant type whose component sizes can be misaligned, on any shape with ne01 not a multiple of the alignment granularity; standard power-of-two dims are unaffected. Implement get_alloc_size for the OpenCL buffer type and reserve the worst-case carve slack (4 aligned gaps; 5 components max, q5_K) for quantized tensors. Costs at most 512 bytes per quantized tensor at the observed 128-byte alignment. opencl: use lm based q6_k mm when ne1 is not multiple of 128 --------- Co-authored-by: Li He <lih@qti.qualcomm.com>	2026-07-08 15:52:21 -07:00
Ruben Ortlam	a646006f09	vulkan: disable FA mask_opt on GCN to improve performance (#24362 ) * vulkan: disable FA mask_opt on GCN to improve performance * reenable mask opt over attention head size 256	2026-07-08 19:01:25 +02:00
Hongqiang Wang	167d057604	opencl: ragged-tile MoE prefill FP16 GEMM optimization (skip padded expert tiles) (#25433 ) * opencl: ragged-tile MoE prefill GEMM (skip padded expert tiles) The MoE prefill GEMM groups tokens into TILESIZE_N=32 per-expert tiles; at low tokens-per-expert most tiles are mostly padding. When a tile's upper 16 slots are all padding (router index 0xFFFFFFFF), skip the second dotx16_reduce8 half. Numerically identical (skipped lanes are padding). Applied to all eight _f32_ns MoE GEMMs; default on, opt out with GGML_OPENCL_MOE_RAGGED_FP16=0. opencl: quarter-granularity ragged MoE tile-skip (8-col skip-groups) Replace the two half-tile dotx16_reduce8 calls in the 8 _f32_ns MoE GEMMs with four dotx8_reduce4 (8-column) calls, skipping each empty trailing skip-group independently. Padding is always trailing, so the kernel rounds the valid count up to the skip granularity and skips fully-padding groups. Byte-identical to the non-skipped path. New env GGML_OPENCL_MOE_RAGGED_GRAN={8,16,32} (quarter/half/ off); default quarter. opencl: move ragged moe env var in cl_init --------- Co-authored-by: Li He <lih@qti.qualcomm.com>	2026-07-08 09:44:55 -07:00
Aman Gupta	1ee093937f	llama-batch: fix allowed decreasing pos in a seq (#25449 )	2026-07-08 19:24:34 +03:00
Ruben Ortlam	0bbc87b163	vulkan: for small AMD GPUs, reduce submission threshold based on CU count (#25240 )	2026-07-08 18:15:18 +02:00