Add support for seed models

Author: lapy

backend/architecture_profiles.py

@@ -172,6 +172,66 @@ def __init__(self) -> None:
         super().__init__(names=("qwen", "qwen2", "qwen3", "qwen2moe", "qwen3moe"))
 
 
+@register_profile
+class SeedProfile(StandardDecoderProfile):
+    """
+    Profile for Seed OSS models (Seed-OSS-36B, etc.).
+    Seed models are Llama-based but may use different metadata keys.
+    Seed-OSS-36B has 64 transformer layers (not 32).
+    """
+    def __init__(self) -> None:
+        super().__init__(names=("seed", "seed-oss", "seedoss"))
+
+    def _calculate_layers(
+        self, metadata: Dict[str, Any], base_block_count: int
+    ) -> LayerConfig:
+        # Seed models might use 'llama' prefix or 'seed' prefix
+        # Check multiple candidate keys, prioritizing seed-specific keys
+        candidate_keys = [
+            "seed.block_count",
+            "seed.n_layer",
+            "seed.n_layers",
+            "llama.block_count",
+            "llama.n_layer",
+            "llama.n_layers",
+            "general.block_count",
+            "general.n_layer",
+        ]
+        
+        metadata_block_count = _get_first_valid_int(
+            metadata, candidate_keys, default=None
+        )
+        
+        # Use the maximum of metadata value or tensor-based count
+        # This is important because Seed models might have incorrect metadata
+        # but correct tensor counts
+        if metadata_block_count is not None:
+            block_count = max(metadata_block_count, base_block_count) if base_block_count > 0 else metadata_block_count
+            if metadata_block_count < base_block_count:
+                logger.info(
+                    f"Seed profile: metadata shows {metadata_block_count} layers but tensor count suggests {base_block_count}, "
+                    f"using tensor count ({base_block_count})"
+                )
+        else:
+            block_count = base_block_count if base_block_count > 0 else 0
+        
+        # Special handling: If we detect around 32 layers but the model size suggests 64,
+        # it might be a Seed model with incorrect metadata
+        if block_count > 0 and block_count < 40:
+            # Check if tensor count suggests more (Seed-OSS-36B should have ~64)
+            if base_block_count > block_count * 1.5:
+                logger.warning(
+                    f"Seed profile: Detected potential mismatch - metadata={block_count}, "
+                    f"tensor_count={base_block_count}. Using tensor_count."
+                )
+                block_count = base_block_count
+        
+        # Standard decoder: blocks + output head
+        effective = (block_count + 1) if block_count > 0 else 0
+        
+        return LayerConfig(block_count=block_count, effective_layer_count=effective)
+
+
 @register_profile
 class LlamaLikeProfile(StandardDecoderProfile):
     """LLaMA, Mistral, Mixtral, Gemma, Phi, etc."""

backend/gguf_reader.py

@@ -707,26 +707,32 @@ def _extract_layer_count(metadata: Dict[str, Any]) -> int:
     """
     # Try different possible keys for layer count
     layer_keys = [
-        'llama.block_count',  # Most common for Llama models
+        'seed.block_count',  # Seed OSS models
+        'seed.n_layer',
+        'seed.layer_count',
+        'llama.block_count',  # Most common for Llama models (Seed models may use this)
         'glm4moe.block_count',  # GLM4 MoE architecture
         'glm4.block_count',  # GLM4 architecture (non-MoE)
         'glm.block_count',  # GLM architecture (non-MoE)
         'qwen3.block_count',  # Qwen3 architecture
         'qwen3moe.block_count',  # Qwen3 MoE architecture
         'qwen.block_count',  # Qwen architecture
         'general.block_count',
+        'seed.layer_count',  # Seed OSS models
         'llama.layer_count',
         'glm4moe.layer_count',
         'glm4.layer_count',  # GLM4 architecture (non-MoE)
         'glm.layer_count',  # GLM architecture (non-MoE)
         'general.layer_count',
         'qwen.layer_count',
         'qwen3.layer_count',
+        'seed.n_layer',  # Seed OSS models
         'llama.n_layer',
         'glm4moe.n_layer',
         'glm4.n_layer',  # GLM4 architecture (non-MoE)
         'glm.n_layer',  # GLM architecture (non-MoE)
         'general.n_layer',
+        'seed.num_layers',  # Seed OSS models
         'llama.num_layers',
         'glm4moe.num_layers',
         'glm4.num_layers',  # GLM4 architecture (non-MoE)

frontend/src/views/ModelConfig.vue

@@ -1226,8 +1226,9 @@ const applyPreset = async (presetName, skipPreview = false) => {
       const defaults = getDefaultConfig()
       const newConfig = { ...defaults, ...smartConfig }
 
-      // Calculate changes
-      const changes = calculateChanges(newConfig, config.value)
+      // Calculate changes - merge defaults with current config to get accurate "before" values
+      const currentConfig = { ...defaults, ...config.value }
+      const changes = calculateChanges(newConfig, currentConfig)
       
       // Calculate impact (only if changes detected)
       let impact = null
@@ -1376,8 +1377,9 @@ const generateAutoConfig = async (skipPreview = false) => {
     const newConfig = { ...defaults, ...smartConfig }
     console.log('Merged new config for application:', newConfig)
 
-    // Calculate changes
-    const changes = calculateChanges(newConfig, config.value)
+    // Calculate changes - merge defaults with current config to get accurate "before" values
+    const currentConfig = { ...defaults, ...config.value }
+    const changes = calculateChanges(newConfig, currentConfig)
     console.log(`Detected ${changes.length} configuration changes`, changes)
     
     // Calculate impact (don't fail if this errors) - only if we have changes