VectorCipher
diff --git a/‎studio/backend/core/inference/llama_cpp.py‎
Lines changed: 666 additions & 49 deletions b/‎studio/backend/core/inference/llama_cpp.py‎
Lines changed: 666 additions & 49 deletions
diff --git a/‎studio/backend/core/inference/llama_server_args.py‎
Lines changed: 26 additions & 1 deletion b/‎studio/backend/core/inference/llama_server_args.py‎
Lines changed: 26 additions & 1 deletion
diff --git a/‎studio/backend/main.py‎
Lines changed: 14 additions & 0 deletions b/‎studio/backend/main.py‎
Lines changed: 14 additions & 0 deletions
diff --git a/‎studio/backend/models/inference.py‎
Lines changed: 143 additions & 1 deletion b/‎studio/backend/models/inference.py‎
Lines changed: 143 additions & 1 deletion
@@ -186,12 +186,25 @@ def is_managed_flag(flag: str) -> bool:
 _TENSOR_SPLIT_FLAGS: frozenset[str] = frozenset({"-ts", "--tensor-split"})
 _SPLIT_SHADOWING_FLAGS: frozenset[str] = _SPLIT_MODE_FLAGS | _TENSOR_SPLIT_FLAGS
 
+# GPU-offload flags. Stripped only when the GPU Memory mode owns offload
+# (manual emits --fit / --gpu-layers / --n-cpu-moe); in auto, a user's
+# inherited -ngl is respected (the offload_overridden path), so this group is
+# opt-in, not default. Layer flags are shared with llama_cpp's override
+# detection; the MoE flags are strip-only (manual's --n-cpu-moe slider owns them).
+_LAYER_OFFLOAD_FLAGS: frozenset[str] = frozenset(
+    {"-ngl", "--gpu-layers", "--n-gpu-layers", "-fit", "--fit"}
+)
+_MOE_OFFLOAD_FLAGS: frozenset[str] = frozenset({"-ncmoe", "--n-cpu-moe", "-cmoe", "--cpu-moe"})
+_OFFLOAD_SHADOWING_FLAGS: frozenset[str] = _LAYER_OFFLOAD_FLAGS | _MOE_OFFLOAD_FLAGS
+
 _SHADOWING_FLAGS: frozenset[str] = (
     _CONTEXT_FLAGS | _CACHE_FLAGS | _SPEC_FLAGS | _TEMPLATE_FLAGS | _SPLIT_SHADOWING_FLAGS
 )
 
 # Shadowing flags that take no value -- strip the flag only, not the next token.
-_BOOLEAN_SHADOWING_FLAGS: frozenset[str] = frozenset({"--spec-default", "--jinja", "--no-jinja"})
+_BOOLEAN_SHADOWING_FLAGS: frozenset[str] = frozenset(
+    {"--spec-default", "--jinja", "--no-jinja", "-cmoe", "--cpu-moe"}
+)
 
 
 def parse_ctx_override(args: Optional[Iterable[str]]) -> Optional[int]:
@@ -424,6 +437,8 @@ def strip_shadowing_flags(
     strip_spec: bool = True,
     strip_template: bool = True,
     strip_split_mode: bool = True,
+    strip_tensor_split: bool = False,
+    strip_offload: bool = False,
 ) -> list[str]:
     """Strip flags that shadow first-class Unsloth settings.
 
@@ -432,6 +447,12 @@ def strip_shadowing_flags(
     (same for cache / spec / template / split-mode). Each ``strip_*``
     toggle controls one group; the route only strips groups whose
     first-class field the caller actually supplied.
+
+    ``strip_split_mode`` removes both ``--split-mode`` and the coupled
+    ``--tensor-split`` (the Tensor Parallelism toggle owns the whole split).
+    ``strip_tensor_split`` removes ``--tensor-split`` *alone*, so manual mode can
+    replace an inherited per-GPU ratio while leaving the user's ``--split-mode``
+    row/none/layer choice intact.
     """
     shadowing: set[str] = set()
     if strip_context:
@@ -444,6 +465,10 @@ def strip_shadowing_flags(
         shadowing |= _TEMPLATE_FLAGS
     if strip_split_mode:
         shadowing |= _SPLIT_SHADOWING_FLAGS
+    if strip_tensor_split:
+        shadowing |= _TENSOR_SPLIT_FLAGS
+    if strip_offload:
+        shadowing |= _OFFLOAD_SHADOWING_FLAGS
 
     tokens = [str(a) for a in (args or [])]
     out: list[str] = []
 
@@ -1167,9 +1167,23 @@ def _get_cached_system_gpu_info(logger) -> dict[str, Any]:
             enriched_dev["vram_utilization_pct"] = util.get("vram_utilization_pct")
             enriched_devices.append(enriched_dev)
 
+        # Whether GGUF loads accept an explicit gpu_ids pick: /load and
+        # /validate 400 picks on XPU hosts (no visibility mask speaks torch-xpu
+        # ordinals) and on Vulkan-only builds (--device pins ggml's own
+        # ordinals), so the picker must not offer them.
+        try:
+            from core.inference.llama_cpp import LlamaCppBackend
+            from utils.hardware import DeviceType, get_device
+            gpu_ids_supported = (
+                get_device() != DeviceType.XPU and not LlamaCppBackend._is_vulkan_backend()
+            )
+        except Exception as e:
+            logger.debug(f"Could not resolve gpu_ids support: {e}")
+            gpu_ids_supported = True
         gpu_info = {
             "available": visibility_info.get("available", False),
             "devices": enriched_devices,
+            "gguf_gpu_ids_supported": gpu_ids_supported,
         }
         _system_gpu_cache = (time.monotonic(), gpu_info)
         return gpu_info
 
@@ -64,7 +64,7 @@ def normalize_blank_chat_template_override(cls, value: Optional[str]) -> Optiona
     )
     gpu_ids: Optional[List[int]] = Field(
         None,
-        description = "Physical GPU indices to use, for example [0, 1]. Omit or pass [] to use automatic selection. Explicit gpu_ids are unsupported when the parent CUDA_VISIBLE_DEVICES uses UUID/MIG entries. Not supported for GGUF models.",
+        description = "Physical GPU indices to use, for example [0, 1]. Omit or pass [] to use automatic selection. Explicit gpu_ids are unsupported when the parent CUDA_VISIBLE_DEVICES uses UUID/MIG entries. For GGUF models the picked devices are pinned via CUDA/HIP_VISIBLE_DEVICES.",
     )
     speculative_type: Optional[str] = Field(
         None,
@@ -100,6 +100,66 @@ def normalize_blank_chat_template_override(cls, value: Optional[str]) -> Optiona
             "No effect on a single GPU. Ignored for non-GGUF models."
         ),
     )
+    gpu_memory_mode: Literal["auto", "manual"] = Field(
+        "auto",
+        description = (
+            "GPU memory strategy for GGUF models. 'auto' (default): Unsloth "
+            "selects GPUs and caps context to fit VRAM. 'manual': you own the "
+            "offload. Leave gpu_layers at -1 (Auto) to hand memory management to "
+            "llama.cpp's --fit (no device masking, no context auto-reduce, no "
+            "gpu-layer/tensor-split planning); set gpu_layers >= 0 to pin layers "
+            "and n_cpu_moe yourself (--fit off), with tensor_parallel still "
+            "applying (split by free VRAM unless tensor_split is set, no planner). "
+            "Ignored for non-GGUF."
+        ),
+    )
+    gpu_layers: int = Field(
+        -1,
+        ge = -1,
+        description = (
+            "Manual mode only: number of layers to offload to the GPU "
+            "(--gpu-layers, with --fit off). A value >= the model's layer count "
+            "offloads all of them. -1 = Auto: hand layer + context sizing to "
+            "llama.cpp's --fit. Ignored unless gpu_memory_mode is 'manual'."
+        ),
+    )
+    n_cpu_moe: int = Field(
+        0,
+        ge = 0,
+        description = (
+            "Manual mode only: keep the first N MoE expert layers on the CPU "
+            "(--n-cpu-moe) to save VRAM on MoE models. 0 = none, N = number of "
+            "MoE layers offloaded (the backend offsets past any leading dense "
+            "layers). Ignored unless gpu_memory_mode is 'manual' with gpu_layers >= 0."
+        ),
+    )
+    tensor_split: Optional[List[float]] = Field(
+        None,
+        description = (
+            "Manual mode only: relative share of the model per GPU (--tensor-split), "
+            "in the order of the GPUs in use, e.g. [2, 1] for 2:1. Omit it to let "
+            "llama.cpp use its default, which splits by free VRAM. Any list given is "
+            "passed through as-is, so send [1, 1] to force an even split. Ignored "
+            "unless gpu_memory_mode is 'manual' with gpu_layers >= 0."
+        ),
+    )
+
+    @field_validator("tensor_split")
+    @classmethod
+    def _reject_degenerate_tensor_split(cls, value: Optional[List[float]]) -> Optional[List[float]]:
+        # A negative / non-finite / all-zero split is silently dropped at launch
+        # (stored as None) yet still compared raw in the reload dedupe, so an
+        # identical Apply reloads forever. Reject it up front; [] = no split.
+        if not value:
+            return value
+        import math
+
+        if any((not math.isfinite(v)) or v < 0 for v in value):
+            raise ValueError("tensor_split entries must be finite and non-negative")
+        if sum(value) <= 0:
+            raise ValueError("tensor_split must have a positive total")
+        return value
+
     llama_extra_args: Optional[List[str]] = Field(
         None,
         description = (
@@ -133,6 +193,14 @@ class ValidateModelRequest(BaseModel):
     max_seq_length: int = Field(0, ge = 0, le = 1048576)
     load_in_4bit: bool = Field(True)
     gpu_ids: Optional[List[int]] = Field(None)
+    gpu_memory_mode: Literal["auto", "manual"] = Field(
+        "auto",
+        description = (
+            "GGUF GPU-memory strategy intended for the follow-up load. Manual "
+            "placement bypasses the training coexistence estimate: Auto layers "
+            "delegate fitting to llama.cpp, while explicit layers are user-owned."
+        ),
+    )
     include_context_length: bool = Field(
         False,
         description = "Also read the native context length from the local GGUF header. "
@@ -188,6 +256,16 @@ class ValidateModelResponse(BaseModel):
         description = "Native training context length, read from the GGUF header when the file "
         "is already downloaded locally; None for non-GGUF, gated, or not-yet-downloaded models.",
     )
+    layer_count: Optional[int] = Field(
+        None,
+        description = "Total layer count (GGUF block_count), the manual gpu-layers ceiling, read "
+        "from the header alongside context_length; None when not read.",
+    )
+    moe_layer_count: Optional[int] = Field(
+        None,
+        description = "MoE expert-layer count (the manual --n-cpu-moe ceiling), read from the GGUF "
+        "header alongside context_length; 0 for dense models, None when not read.",
+    )
     # Additive fields; the consuming consent dialog ships in a follow-up frontend PR.
     requires_transformers_upgrade: bool = Field(
         False,
@@ -333,6 +411,34 @@ class LoadResponse(BaseModel):
         False,
         description = "Whether tensor-parallel split (--split-mode tensor) is active.",
     )
+    gpu_memory_mode: Literal["auto", "manual"] = Field(
+        "auto",
+        description = "Active GPU memory strategy ('auto' or 'manual').",
+    )
+    gpu_layers: int = Field(
+        -1,
+        description = "Manual mode: requested --gpu-layers value (-1 = Auto/--fit, or when not manual).",
+    )
+    n_cpu_moe: int = Field(
+        0,
+        description = "Manual mode: MoE expert layers pinned to CPU (--n-cpu-moe); 0 = none.",
+    )
+    tensor_split: Optional[List[float]] = Field(
+        None,
+        description = "Manual mode: relative model share per GPU (--tensor-split); None = default (split by free VRAM).",
+    )
+    n_layers: Optional[int] = Field(
+        None,
+        description = "Model's layer count (GGUF block_count), for the manual gpu-layers ceiling.",
+    )
+    n_moe_layers: int = Field(
+        0,
+        description = "Model's MoE expert-layer count (the n_cpu_moe ceiling); 0 if not an MoE model.",
+    )
+    gpu_ids: Optional[List[int]] = Field(
+        None,
+        description = "Physical GPU indices the model is pinned to, or None for automatic selection.",
+    )
 
 
 class UnloadResponse(BaseModel):
@@ -461,6 +567,42 @@ class InferenceStatusResponse(BaseModel):
         False,
         description = "Whether tensor-parallel split (--split-mode tensor) is active.",
     )
+    gpu_memory_mode: Literal["auto", "manual"] = Field(
+        "auto",
+        description = "Active GPU memory strategy ('auto' or 'manual').",
+    )
+    gpu_layers: int = Field(
+        -1,
+        description = "Manual mode: requested --gpu-layers value (-1 = Auto/--fit, or when not manual).",
+    )
+    n_cpu_moe: int = Field(
+        0,
+        description = "Manual mode: MoE expert layers pinned to CPU (--n-cpu-moe); 0 = none.",
+    )
+    tensor_split: Optional[List[float]] = Field(
+        None,
+        description = "Manual mode: relative model share per GPU (--tensor-split); None = default (split by free VRAM).",
+    )
+    requested_context_length: Optional[int] = Field(
+        None,
+        description = (
+            "The n_ctx the active GGUF load was invoked with (0 = Auto). Lets the "
+            "UI re-seed a Manual + Auto-layers context pin on hydration, where "
+            "context_length only exposes the resolved value. None for non-GGUF."
+        ),
+    )
+    n_layers: Optional[int] = Field(
+        None,
+        description = "Model's layer count (GGUF block_count), for the manual gpu-layers ceiling.",
+    )
+    n_moe_layers: int = Field(
+        0,
+        description = "Model's MoE expert-layer count (the n_cpu_moe ceiling); 0 if not an MoE model.",
+    )
+    gpu_ids: Optional[List[int]] = Field(
+        None,
+        description = "Physical GPU indices the model is pinned to, or None for automatic selection.",
+    )
     llama_cpp_supports_mtp: bool = Field(
         True,
         description = (