diff --git a/common/arg.cpp b/common/arg.cpp
index 1f70d0ad45..a5a72970d6 100644
--- a/common/arg.cpp
+++ b/common/arg.cpp
@@ -4065,6 +4065,9 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         {"--spec-draft-n-max"}, "N",
         string_format("number of tokens to draft for speculative decoding (default: %d)", params.speculative.draft.n_max),
         [](common_params & params, int value) {
+            if (value < 0) {
+                throw std::invalid_argument("invalid value");
+            }
             params.speculative.draft.n_max = value;
         }
     ).set_spec().set_examples({LLAMA_EXAMPLE_SPECULATIVE, LLAMA_EXAMPLE_LOOKUP, LLAMA_EXAMPLE_SERVER, LLAMA_EXAMPLE_CLI}).set_env("LLAMA_ARG_SPEC_DRAFT_N_MAX"));
diff --git a/tools/mtmd/clip.cpp b/tools/mtmd/clip.cpp
index c2d5059e02..c08b41d590 100644
--- a/tools/mtmd/clip.cpp
+++ b/tools/mtmd/clip.cpp
@@ -1948,9 +1948,12 @@ struct clip_model_loader {
                 if (hparams.image_max_pixels < hparams.image_min_pixels) {
                     throw std::runtime_error(string_format("%s: image_max_pixels (%d) is less than image_min_pixels (%d)\n", __func__, hparams.image_max_pixels, hparams.image_min_pixels));
                 }
-                if (hparams.n_merge < 0 || hparams.n_merge >= 65536) {
+                if (hparams.n_merge <= 0 || hparams.n_merge >= 65536) {
                     throw std::runtime_error(string_format("%s: n_merge (%d) must be greater than 0 and less than 65536\n", __func__, hparams.n_merge));
                 }
+                if (hparams.attn_window_size > 4096) {
+                    throw std::runtime_error(string_format("%s: attn_window_size (%d) is too large (max 4096)\n", __func__, hparams.attn_window_size));
+                }
             }
 
             LOG_INF("%s: projector:          %s\n", __func__, proj_type.c_str());
@@ -1985,8 +1988,6 @@ struct clip_model_loader {
                     LOG_INF("%s: preproc_tiles:      %d - %d\n", __func__, hparams.preproc_min_tiles, hparams.preproc_max_tiles);
                 }
             } else if (is_audio) {
-                GGML_ASSERT(hparams.attn_window_size <= 4096); // avoid int32_t overflow in attn_dists/mask buffers
-
                 LOG_INF("\n--- audio hparams ---\n");
                 LOG_INF("%s: n_mel_bins:         %d\n", __func__, hparams.n_mel_bins);
                 LOG_INF("%s: proj_stack_factor:  %d\n", __func__, hparams.proj_stack_factor);
diff --git a/tools/mtmd/mtmd-helper-common.h b/tools/mtmd/mtmd-helper-common.h
index d027e4c9cb..f907346c7b 100644
--- a/tools/mtmd/mtmd-helper-common.h
+++ b/tools/mtmd/mtmd-helper-common.h
@@ -117,7 +117,7 @@ struct decode_embd_batch {
         for (int32_t i = 0; i < batch.n_tokens; i++) {
             const size_t idx = (size_t) i;
             const size_t n_tokens = (size_t) batch.n_tokens;
-            pos[idx                                  ] = rel_pos[i].t;
+            pos[idx                    ] = rel_pos[i].t;
             pos[idx + n_tokens         ] = rel_pos[i].y;
             pos[idx + n_tokens * 2     ] = rel_pos[i].x;
             pos[idx + n_tokens * 3     ] = rel_pos[i].z;
@@ -136,6 +136,7 @@ struct decode_embd_batch {
         for (int i = 0; i < batch.n_tokens; i++) {
             const size_t idx = (size_t) i;
             const size_t n_tokens = (size_t) batch.n_tokens;
+            pos[idx                    ] = pos_0 + i;
             pos[idx + n_tokens         ] = pos_0 + i;
             pos[idx + n_tokens * 2     ] = pos_0 + i;
             pos[idx + n_tokens * 3     ] = pos_0 + i;