localcode 0.3.41__tar.gz → 0.3.43__tar.gz

{localcode-0.3.41/src/localcode.egg-info → localcode-0.3.43}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: localcode
-Version: 0.3.41
+Version: 0.3.43
 Summary: High-performance AI coding on consumer hardware.
 Author: LocalCode contributors
 License-Expression: Apache-2.0
@@ -134,6 +134,7 @@ On launch, LocalCode recommends the best model for **your Mac's RAM** — there'
 | Gemma 4 12B | 7.4 GB (Q4) | 12B (dense) | 16 GB | gemma4-iswa |
 | Gemma 4 26B-A4B | 11.2 GB (Q3) | 3.8B (8/128 experts) | 24 GB | gemma4-iswa |
 | Qwen 3.6 35B-A3B | 10.7 GB (Q2) | 3.0B (8+1/256) | 24 GB | qwen35moe |
+| Qwen 3.8 27B | 17.9 GB (Q4) | 27B (dense) | 32 GB | qwen35 |
 | DiffusionGemma 26B-A4B † | 15.7 GB (Q4) | 4B (diffusion MoE) | 32 GB | diffusion_gemma |
 | North-Mini-Code 30B-A3B † | 17.9 GB (Q4) | 3B (30B MoE) | 36 GB | cohere2_moe |
 | Gemma 4 12B (full) | 23.8 GB (BF16) | 12B (dense) | 48 GB | gemma4-iswa |

{localcode-0.3.41 → localcode-0.3.43}/README.md

@@ -87,6 +87,7 @@ On launch, LocalCode recommends the best model for **your Mac's RAM** — there'
 | Gemma 4 12B | 7.4 GB (Q4) | 12B (dense) | 16 GB | gemma4-iswa |
 | Gemma 4 26B-A4B | 11.2 GB (Q3) | 3.8B (8/128 experts) | 24 GB | gemma4-iswa |
 | Qwen 3.6 35B-A3B | 10.7 GB (Q2) | 3.0B (8+1/256) | 24 GB | qwen35moe |
+| Qwen 3.8 27B | 17.9 GB (Q4) | 27B (dense) | 32 GB | qwen35 |
 | DiffusionGemma 26B-A4B † | 15.7 GB (Q4) | 4B (diffusion MoE) | 32 GB | diffusion_gemma |
 | North-Mini-Code 30B-A3B † | 17.9 GB (Q4) | 3B (30B MoE) | 36 GB | cohere2_moe |
 | Gemma 4 12B (full) | 23.8 GB (BF16) | 12B (dense) | 48 GB | gemma4-iswa |

{localcode-0.3.41 → localcode-0.3.43}/llama-cpp-turboquant/src/llama-arch.cpp

@@ -1064,6 +1064,11 @@ static std::set<llm_tensor> llm_get_tensor_names(llm_arch arch) {
                 LLM_TENSOR_SSM_ALPHA,
                 LLM_TENSOR_SSM_NORM,
                 LLM_TENSOR_SSM_OUT,
+                // Multi-Token-Prediction (nextn) tensors on the trailing MTP layer.
+                LLM_TENSOR_NEXTN_EH_PROJ,
+                LLM_TENSOR_NEXTN_ENORM,
+                LLM_TENSOR_NEXTN_HNORM,
+                LLM_TENSOR_NEXTN_SHARED_HEAD_NORM,
             };
         case LLM_ARCH_QWEN35MOE:
             return {
@@ -2770,12 +2775,16 @@ static const std::map<llm_tensor, llm_tensor_info> LLM_TENSOR_INFOS = {
     {LLM_TENSOR_INDEXER_ATTN_Q_B,           {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
     // NextN/MTP tensors are currently ignored (reserved for future MTP support)
     // These tensors only exist in the last layer(s) and are treated as output tensors
-    {LLM_TENSOR_NEXTN_EH_PROJ,              {LLM_TENSOR_LAYER_OUTPUT, GGML_OP_MUL_MAT}},
-    {LLM_TENSOR_NEXTN_EMBED_TOKENS,         {LLM_TENSOR_LAYER_OUTPUT, GGML_OP_GET_ROWS}},
-    {LLM_TENSOR_NEXTN_ENORM,                {LLM_TENSOR_LAYER_OUTPUT, GGML_OP_GET_ROWS}},
-    {LLM_TENSOR_NEXTN_HNORM,                {LLM_TENSOR_LAYER_OUTPUT, GGML_OP_MUL}},
-    {LLM_TENSOR_NEXTN_SHARED_HEAD_HEAD,     {LLM_TENSOR_LAYER_OUTPUT, GGML_OP_MUL_MAT}},
-    {LLM_TENSOR_NEXTN_SHARED_HEAD_NORM,     {LLM_TENSOR_LAYER_OUTPUT, GGML_OP_MUL}},
+    // NextN/MTP tensors are per-layer (they live on the trailing MTP block),
+    // so they must be REPEATING — an OUTPUT classification forbids the layer
+    // number their `blk.%d.nextn.*` names require. Loaded-but-unused: the
+    // forward pass skips the MTP layer, so this is claim-only.
+    {LLM_TENSOR_NEXTN_EH_PROJ,              {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
+    {LLM_TENSOR_NEXTN_EMBED_TOKENS,         {LLM_TENSOR_LAYER_REPEATING, GGML_OP_GET_ROWS}},
+    {LLM_TENSOR_NEXTN_ENORM,                {LLM_TENSOR_LAYER_REPEATING, GGML_OP_GET_ROWS}},
+    {LLM_TENSOR_NEXTN_HNORM,                {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
+    {LLM_TENSOR_NEXTN_SHARED_HEAD_HEAD,     {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
+    {LLM_TENSOR_NEXTN_SHARED_HEAD_NORM,     {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
     // Nemotron 3 Super
     {LLM_TENSOR_FFN_LATENT_DOWN,            {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
     {LLM_TENSOR_FFN_LATENT_UP,              {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},

{localcode-0.3.41 → localcode-0.3.43}/llama-cpp-turboquant/src/llama-model.cpp

@@ -2437,16 +2437,26 @@ void llama_model::load_hparams(llama_model_loader & ml) {
                 ml.get_key(LLM_KV_SSM_TIME_STEP_RANK, hparams.ssm_dt_rank);
                 ml.get_key(LLM_KV_SSM_GROUP_COUNT,    hparams.ssm_n_group);
 
-                // Mark recurrent layers (linear attention layers)
+                // Multi-Token-Prediction (nextn) layers are trailing prediction
+                // heads, not part of the transformer stack. Exclude them from
+                // recurrent marking, tensor loading and the forward pass.
+                // Defaults to 0, so models without an MTP layer are unaffected.
+                ml.get_key(LLM_KV_NEXTN_PREDICT_LAYERS, hparams.nextn_predict_layers, false);
+                const uint32_t n_transformer = hparams.n_layer - hparams.nextn_predict_layers;
+                hparams.n_layer_kv_from_start = n_transformer;
+
+                // Mark recurrent (linear-attention) layers over the real
+                // transformer layers only; an MTP layer is neither recurrent nor
+                // a full-attention layer (it carries its own nextn tensors).
                 {
                     uint32_t full_attn_interval = 4;
                     ml.get_key(LLM_KV_FULL_ATTENTION_INTERVAL, full_attn_interval, false);
                     for (uint32_t i = 0; i < hparams.n_layer; ++i) {
-                        hparams.recurrent_layer_arr[i] = ((i + 1) % full_attn_interval != 0);
+                        hparams.recurrent_layer_arr[i] = (i < n_transformer) && ((i + 1) % full_attn_interval != 0);
                     }
                 }
 
-                switch (hparams.n_layer) {
+                switch (n_transformer) {
                     case 24: type = hparams.n_embd == 1024 ? LLM_TYPE_0_8B : LLM_TYPE_2B; break;
                     case 32: type = hparams.n_embd == 2560 ? LLM_TYPE_4B : LLM_TYPE_9B; break;
                     case 64: type = LLM_TYPE_27B; break;
@@ -7487,6 +7497,20 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                         layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd,   n_ff}, 0);
                         layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  n_ff, n_embd}, 0);
                         layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff}, 0);
+
+                        // A trailing Multi-Token-Prediction (nextn) layer is a full
+                        // attention+FFN block (loaded above) PLUS prediction-head
+                        // tensors. Claim those too so the loader's tensor count
+                        // matches the GGUF; the forward pass skips this layer, so
+                        // they are loaded-but-unused. No-op when nextn_predict_layers
+                        // is 0 (every model without an MTP head).
+                        if (hparams.nextn_predict_layers > 0 &&
+                            (uint32_t) i >= hparams.n_layer - hparams.nextn_predict_layers) {
+                            layer.nextn.eh_proj          = create_tensor(tn(LLM_TENSOR_NEXTN_EH_PROJ,          "weight", i), { 2 * n_embd, n_embd }, 0);
+                            layer.nextn.enorm            = create_tensor(tn(LLM_TENSOR_NEXTN_ENORM,            "weight", i), { n_embd }, 0);
+                            layer.nextn.hnorm            = create_tensor(tn(LLM_TENSOR_NEXTN_HNORM,            "weight", i), { n_embd }, 0);
+                            layer.nextn.shared_head_norm = create_tensor(tn(LLM_TENSOR_NEXTN_SHARED_HEAD_NORM, "weight", i), { n_embd }, 0);
+                        }
                     }
                 } break;
             case LLM_ARCH_MIMO2:

{localcode-0.3.41 → localcode-0.3.43}/llama-cpp-turboquant/src/models/qwen35.cpp

@@ -23,7 +23,11 @@ llm_build_qwen35::llm_build_qwen35(const llama_model & model, const llm_graph_pa
     ggml_tensor * inp_pos     = build_inp_pos();
     ggml_tensor * inp_out_ids = build_inp_out_ids();
 
-    for (int il = 0; il < n_layer; ++il) {
+    // Trailing Multi-Token-Prediction (nextn) layers are prediction heads, not
+    // part of the decode stack — process only the real transformer layers.
+    // n_transformer == n_layer for models without an MTP layer.
+    const int n_transformer = n_layer - (int) hparams.nextn_predict_layers;
+    for (int il = 0; il < n_transformer; ++il) {
         ggml_tensor * inpSA = inpL;
 
         cur = build_norm(inpL, model.layers[il].attn_norm, nullptr, LLM_NORM_RMS, il);
@@ -40,7 +44,7 @@ llm_build_qwen35::llm_build_qwen35(const llama_model & model, const llm_graph_pa
             cur = build_layer_attn(inp->get_attn(), cur, inp_pos, sections, il);
         }
 
-        if (il == n_layer - 1 && inp_out_ids) {
+        if (il == n_transformer - 1 && inp_out_ids) {
             cur   = ggml_get_rows(ctx0, cur, inp_out_ids);
             inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
         }

{localcode-0.3.41 → localcode-0.3.43}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "localcode"
-version = "0.3.41"
+version = "0.3.43"
 description = "High-performance AI coding on consumer hardware."
 readme = "README.md"
 requires-python = ">=3.10"

{localcode-0.3.41 → localcode-0.3.43}/src/localcode/__init__.py

@@ -8,4 +8,4 @@ from importlib.metadata import PackageNotFoundError, version as _pkg_version
 try:
     __version__ = _pkg_version("localcode")
 except PackageNotFoundError:  # not installed (e.g. raw source checkout)
-    __version__ = "0.3.41"
+    __version__ = "0.3.43"

{localcode-0.3.41 → localcode-0.3.43}/src/localcode/models_catalog.py

@@ -323,6 +323,21 @@ CHOICES: list[ModelChoice] = [
         mmproj_size_gb=0.9,
         mmproj_hf_filename="mmproj-F16.gguf",
     ),
+    ModelChoice(
+        key="qwen38",
+        name="Qwen 3.8 27B (Q4)",
+        hf_repo="unsloth/Qwen3.8-27B-GGUF",
+        filename="Qwen3.8-27B-UD-Q4_K_XL.gguf",
+        size_gb=17.9,
+        active_params="27B dense (hybrid attention + Mamba-2 SSM, 1 MTP layer)",
+        architecture="qwen35",
+        license="Apache 2.0",
+        humaneval_pass_at_1=None,
+        notes="Dense 27B — every parameter active each token (no MoE routing), so it's heavier per token than the 35B-A3B MoE but denser in capability. Hybrid attention + Mamba-2 SSM with a multi-token-prediction (MTP) layer; the TurboQuant fork loads the transformer stack and skips the MTP block at inference. UD-Q4_K_XL keeps attention/embedding above 4-bit. Needs ~32 GB unified memory to fit with context; comfortable on 48-64 GB Apple Silicon. Pair with the F16 mmproj for image input.",
+        mmproj_filename="mmproj-Qwen3.8-27B-F16.gguf",
+        mmproj_size_gb=0.9,
+        mmproj_hf_filename="mmproj-F16.gguf",
+    ),
 ]
 
 
@@ -559,6 +574,25 @@ MODEL_GROUPS: list[ModelGroup] = [
         mmproj_size_gb=0.9,
         mmproj_hf_filename="mmproj-F16.gguf",
     ),
+    ModelGroup(
+        key="qwen-3.8-27b",
+        display_name="Qwen 3.8 27B",
+        maker="Alibaba",
+        hf_repo="unsloth/Qwen3.8-27B-GGUF",
+        family="qwen",
+        architecture="qwen35",
+        license="Apache 2.0",
+        notes=(
+            "Alibaba's dense 27B — all params active per token (no MoE routing). "
+            "Hybrid attention + Mamba-2 SSM with a multi-token-prediction layer; "
+            "the TurboQuant fork serves the transformer stack and skips the MTP "
+            "block. Native multimodal: pair any quant with the F16 mmproj for "
+            "image input. Tool calling via Qwen/Hermes format."
+        ),
+        mmproj_filename="mmproj-Qwen3.8-27B-F16.gguf",
+        mmproj_size_gb=0.9,
+        mmproj_hf_filename="mmproj-F16.gguf",
+    ),
     ModelGroup(
         key="north-mini-code-1.0",
         display_name="North-Mini-Code 1.0 30B-A3B",

{localcode-0.3.41 → localcode-0.3.43}/src/localcode/tui/screens/setup.py

@@ -130,34 +130,57 @@ class SetupScreen(Screen):
         # screen-resize hook on this class still compiles.
         return
 
-    def _render_steps(self) -> str:
-        # Substitute the actually-selected model's size into the "Download model"
-        # label so the user sees the real number, not a hardcoded ~10GB.
-        # If a partial download exists on disk (from a prior aborted attempt),
-        # surface that as "Resume download" so the user knows we're picking up
-        # where they left off — not starting over.
+    def _model_state(self) -> tuple[str, float | None]:
+        """Disk state of the chosen model: ('present'|'partial'|'absent', size_gb).
+
+        - present: the GGUF is fully on disk (≥99% of catalog size) — this
+          launch only *loads* it, no download happens.
+        - partial: a `.incomplete` sidecar or an undersized final exists — a
+          download will *resume*.
+        - absent: nothing on disk — a fresh download is needed.
+        size_gb is None only when there's no selectable model.
+        """
         try:
             from ...models_catalog import current as current_choice
+            # Reuse the canonical completeness check so we agree with the rest
+            # of the app on what "downloaded" means. It compares against DECIMAL
+            # GB (size_gb * 1000**3); an earlier inline check here used binary
+            # GiB (1024**3), which over-counts by ~7% and made every fully
+            # present model read as "partial" → the bogus "Resume model /
+            # one-time download" caption on a model already on disk.
+            from ...bootstrap import _is_complete_download
             cfg = getattr(self.app, "config", None)
             chosen = current_choice(cfg) if cfg is not None else None
-            if chosen is not None:
-                verb = "Download"
-                # huggingface_hub writes `.incomplete` sidecars; our urllib
-                # fallback pre-allocates the final filename. Detect both.
-                partial = chosen.local_path
-                incomplete = partial.with_name(partial.name + ".incomplete")
-                has_partial = (
-                    (partial.exists() and partial.stat().st_size > 0
-                     and partial.stat().st_size < int(chosen.size_gb * 1024 ** 3 * 0.99))
-                    or incomplete.exists()
-                )
-                if has_partial:
-                    verb = "Resume"
-                model_label = f"{verb} model (~{chosen.size_gb:.1f} GB)"
-            else:
-                model_label = "Download model"
+            if chosen is None:
+                return ("absent", None)
+            final = chosen.local_path
+            if _is_complete_download(final, chosen):
+                return ("present", chosen.size_gb)
+            # huggingface_hub writes `.incomplete` sidecars; our urllib
+            # fallback pre-allocates the final filename. Either → a resumable
+            # partial. Otherwise nothing is on disk yet.
+            incomplete = final.with_name(final.name + ".incomplete")
+            has_partial = (
+                (final.exists() and final.stat().st_size > 0) or incomplete.exists()
+            )
+            return ("partial" if has_partial else "absent", chosen.size_gb)
         except Exception:
+            return ("absent", None)
+
+    def _render_steps(self) -> str:
+        # Substitute the actually-selected model's size into the model step
+        # label, and pick a verb that matches what's actually happening: a
+        # fully-present model is only *loaded* (not downloaded), a partial is
+        # *resumed*, and a missing one is *downloaded*.
+        state, size_gb = self._model_state()
+        if size_gb is None:
             model_label = "Download model"
+        elif state == "present":
+            model_label = f"Model ready (~{size_gb:.1f} GB)"
+        elif state == "partial":
+            model_label = f"Resume download (~{size_gb:.1f} GB)"
+        else:
+            model_label = f"Download model (~{size_gb:.1f} GB)"
 
         lines = ["[bold]Setup[/]\n"]
         for i, (key, label) in enumerate(self.STEPS):
@@ -321,6 +344,22 @@ class SetupScreen(Screen):
             steps = self._render_steps()
             self.query_one("#setup-status", Static).update(status)
             self.query_one("#setup-steps", Static).update(steps)
+            # Only show the download-framing captions when a download is
+            # actually going to happen. A fully-present model is just being
+            # loaded into memory — "one-time download" / "resumes where it
+            # left off" would be a lie there.
+            if self._model_state()[0] == "present":
+                self.query_one("#setup-onetime-note", Static).update(
+                    "[dim italic]loading into memory · first launch after boot is slower[/]"
+                )
+                self.query_one("#setup-quit-hint", Static).update("")
+            else:
+                self.query_one("#setup-onetime-note", Static).update(
+                    "[dim italic]one-time download · cached for future launches[/]"
+                )
+                self.query_one("#setup-quit-hint", Static).update(
+                    "[dim]Esc to quit — the download resumes where it left off next launch.[/]"
+                )
             self.refresh()
         except Exception:
             pass

{localcode-0.3.41 → localcode-0.3.43/src/localcode.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: localcode
-Version: 0.3.41
+Version: 0.3.43
 Summary: High-performance AI coding on consumer hardware.
 Author: LocalCode contributors
 License-Expression: Apache-2.0
@@ -134,6 +134,7 @@ On launch, LocalCode recommends the best model for **your Mac's RAM** — there'
 | Gemma 4 12B | 7.4 GB (Q4) | 12B (dense) | 16 GB | gemma4-iswa |
 | Gemma 4 26B-A4B | 11.2 GB (Q3) | 3.8B (8/128 experts) | 24 GB | gemma4-iswa |
 | Qwen 3.6 35B-A3B | 10.7 GB (Q2) | 3.0B (8+1/256) | 24 GB | qwen35moe |
+| Qwen 3.8 27B | 17.9 GB (Q4) | 27B (dense) | 32 GB | qwen35 |
 | DiffusionGemma 26B-A4B † | 15.7 GB (Q4) | 4B (diffusion MoE) | 32 GB | diffusion_gemma |
 | North-Mini-Code 30B-A3B † | 17.9 GB (Q4) | 3B (30B MoE) | 36 GB | cohere2_moe |
 | Gemma 4 12B (full) | 23.8 GB (BF16) | 12B (dense) | 48 GB | gemma4-iswa |

{localcode-0.3.41 → localcode-0.3.43}/tests/test_comprehensive_models_matrix.py

@@ -91,7 +91,7 @@ def test_every_choice_is_wellformed(choice):
     assert choice.filename and choice.filename.endswith(".gguf")
     assert choice.size_gb and choice.size_gb > 0
     assert choice.architecture in {
-        "gemma4-iswa", "qwen35moe", "diffusion_gemma", "cohere2_moe",
+        "gemma4-iswa", "qwen35moe", "qwen35", "diffusion_gemma", "cohere2_moe",
     }, f"unknown arch {choice.architecture!r}"
     # by_filename must round-trip to a choice carrying the same architecture
     # (this is the dispatch key the runtime keys off).

{localcode-0.3.41 → localcode-0.3.43}/tests/test_models_catalog.py

@@ -78,6 +78,7 @@ _EXPECTED_VISION = {
     "qwen": True,
     "gemma-q8": True,
     "qwen-q8": True,
+    "qwen38": True,
     "diffusiongemma": False,
     "north-mini-code": False,
 }