@simulatte/doppler 0.1.8 → 0.1.9

package/src/config/presets/kernel-paths/registry.json

@@ -92,6 +92,13 @@
       "statusReason": "default",
       "notes": "Gemma 3 Q4K dequant default: subgroup GEMV + online attention + tuned lm_head multicol, F32 activations."
     },
+    {
+      "id": "gemma3-q4k-dequant-f32a-small-attn",
+      "file": "gemma3-q4k-dequant-f32a-small-attn.json",
+      "status": "experimental",
+      "statusReason": "diagnostic-probe",
+      "notes": "Diagnostic: same as gemma3-q4k-dequant-f32a-online but uses attention_small_f16kv.wgsl for prefill to isolate streaming attention bug."
+    },
     {
       "id": "gemma3-q4k-dequant-f32w-f32a-online",
       "file": "gemma3-q4k-dequant-f32w-f32a-online.json",

package/src/config/presets/models/gemma3.json

@@ -34,7 +34,8 @@
     },
     "rope": {
       "ropeTheta": 1000000,
-      "ropeLocalTheta": 10000
+      "ropeLocalTheta": 10000,
+      "ropeLocalScalingFactor": 1.0
     },
     "chatTemplate": {
       "type": "gemma",

package/src/config/presets/models/gemma4.json

@@ -0,0 +1,61 @@
+{
+  "id": "gemma4",
+  "name": "Gemma 4",
+  "extends": "gemma3",
+  "modelType": "mixtral",
+
+  "inference": {
+    "attention": {
+      "slidingWindow": 1024
+    },
+    "rope": {
+      "ropeTheta": 1000000,
+      "ropeLocalTheta": 10000,
+      "ropeScalingType": "yarn",
+      "ropeScalingFactor": 8.0,
+      "yarnBetaFast": 4.0,
+      "yarnBetaSlow": 1.0,
+      "yarnOriginalMaxPos": 32768
+    },
+    "moe": {
+      "kernelProfileId": "mixtral-moe-v1",
+      "numExperts": 8,
+      "topK": 2,
+      "numSharedExperts": 0,
+      "routerDtype": "f32",
+      "supportedActivationDtypes": ["f16", "f32"],
+      "preferredActivationDtype": "f32",
+      "tensorPattern": "mixtral"
+    },
+    "kernelPaths": {
+      "q4k": {
+        "default": "gemma3-q4k-dequant-f32a-online",
+        "f16": "gemma3-q4k-dequant-f16a-online",
+        "f16a": "gemma3-q4k-dequant-f16a-online",
+        "f32": "gemma3-q4k-dequant-f32a-online"
+      }
+    }
+  },
+
+  "tensorPatterns": {
+    "ffn": {
+      "gate": ["layers.{layer}.block_sparse_moe.experts.{expert}.w1.weight"],
+      "up": ["layers.{layer}.block_sparse_moe.experts.{expert}.w3.weight"],
+      "down": ["layers.{layer}.block_sparse_moe.experts.{expert}.w2.weight"]
+    }
+  },
+
+  "detection": {
+    "architecturePatterns": [
+      "gemma4",
+      "Gemma4ForCausalLM",
+      "Gemma4ForConditionalGeneration",
+      "gemma-4"
+    ],
+    "modelTypePatterns": [
+      "gemma4",
+      "gemma4_text",
+      "gemma4_moe"
+    ]
+  }
+}

package/src/config/presets/models/granite-docling.json

@@ -0,0 +1,70 @@
+{
+  "id": "granite-docling",
+  "name": "Granite-Docling (Document OCR VLM)",
+  "extends": "transformer",
+  "modelType": "ocr",
+
+  "_notes": "Stabilized successor to SmolDocling-256M-preview. Requires full multimodal pipeline: SigLIP vision encoder, Idefics3-style image-token merge, pixel-shuffle connector, SmolLM2 decoder, DocTags output parsing. This preset covers the decoder config only — vision encoder and connector are separate pipeline stages not yet implemented in Doppler.",
+
+  "architecture": {
+    "headDim": 64,
+    "ropeTheta": 10000,
+    "visionEncoder": {
+      "type": "siglip_b16",
+      "patchSize": 16,
+      "imageSize": 512,
+      "hiddenSize": 768,
+      "numLayers": 12,
+      "numHeads": 12,
+      "parameterCount": 93000000,
+      "_note": "SigLIP base patch-16/512 backbone. Requires dedicated vision encoder pipeline in Doppler."
+    },
+    "connector": {
+      "type": "mlp_pixel_shuffle",
+      "downsampleFactor": 2,
+      "_note": "Idefics3/SmolVLM-style projection. Maps vision tokens to decoder embedding space."
+    }
+  },
+
+  "inference": {
+    "attention": {
+      "queryKeyNorm": false,
+      "causal": true
+    },
+    "normalization": {
+      "rmsNormWeightOffset": false,
+      "rmsNormEps": 1e-5
+    },
+    "ffn": {
+      "activation": "silu"
+    },
+    "output": {
+      "scaleEmbeddings": false,
+      "tieWordEmbeddings": true
+    },
+    "chatTemplate": {
+      "enabled": false
+    },
+    "kernelPaths": {
+      "q4k": {
+        "f16": "granite-docling-q4k-dequant-f32a",
+        "f32": "granite-docling-q4k-dequant-f32a"
+      },
+      "f16": {
+        "f16": "granite-docling-f16-f32a",
+        "f32": "granite-docling-f16-f32a"
+      }
+    }
+  },
+
+  "tokenizer": {
+    "bosToken": "<|endoftext|>",
+    "eosTokens": ["<|endoftext|>", "<|im_end|>"],
+    "addBosToken": true
+  },
+
+  "detection": {
+    "architecturePatterns": ["granite-docling", "GraniteDocling", "smoldocling", "SmolDocling", "SmolVLM"],
+    "modelTypePatterns": ["granite-docling", "smoldocling", "smolvlm"]
+  }
+}

package/src/config/presets/models/lfm2.json

@@ -39,11 +39,16 @@
       "period": null,
       "offset": null,
       "layerTypes": null
+    },
+    "chatTemplate": {
+      "type": "chatml",
+      "enabled": true
     }
   },
 
   "tokenizer": {
-    "addBosToken": false,
+    "bosTokenId": 1,
+    "addBosToken": true,
     "addEosToken": false
   },
 

package/src/config/presets/models/qwen3_vl.json

@@ -0,0 +1,40 @@
+{
+  "id": "qwen3_vl",
+  "name": "Qwen 3 VL",
+  "extends": "qwen3",
+
+  "architecture": {
+    "ropeTheta": 5000000
+  },
+
+  "inference": {
+    "normalization": {
+      "rmsNormWeightOffset": false
+    },
+    "rope": {
+      "ropeTheta": 5000000,
+      "mropeInterleaved": true,
+      "mropeSection": [24, 20, 20],
+      "partialRotaryFactor": null
+    }
+  },
+
+  "vision": {
+    "patchSize": 16,
+    "spatialMergeSize": 2,
+    "temporalPatchSize": 2,
+    "eps": 1e-6,
+    "minPixels": 3136,
+    "maxPixels": 1003520,
+    "projectorType": "spatial_merge",
+    "normalization": {
+      "mean": [0.48145466, 0.4578275, 0.40821073],
+      "std": [0.26862954, 0.26130258, 0.27577711]
+    }
+  },
+
+  "detection": {
+    "architecturePatterns": ["qwen3_vl", "Qwen3VLForConditionalGeneration"],
+    "modelTypePatterns": ["qwen3_vl"]
+  }
+}

package/src/config/presets/runtime/experiments/bench/gemma3-bench-q4k.json

@@ -1,7 +1,8 @@
 {
   "id": "experiments/bench/gemma3-bench-q4k",
   "name": "gemma3-bench-q4k",
-  "intent": "investigate",
+  "description": "Benchmark run for Gemma 3 1B Q4K — calibration-mode throughput measurement.",
+  "intent": "calibrate",
   "stability": "experimental",
   "owner": "doppler-core",
   "createdAtUtc": "2026-02-25T00:00:00Z",

package/src/config/presets/runtime/experiments/verify/lfm2-verify.json

@@ -0,0 +1,46 @@
+{
+  "id": "experiments/verify/lfm2-verify",
+  "name": "lfm2-verify",
+  "intent": "verify",
+  "stability": "experimental",
+  "owner": "doppler-core",
+  "createdAtUtc": "2026-03-16T00:00:00Z",
+  "extends": "modes/bench",
+  "model": "lfm2-5-1-2b-instruct-q4k-ehf16-af32",
+  "runtime": {
+    "shared": {
+      "tooling": {
+        "intent": "verify"
+      },
+      "debug": {
+        "logLevel": {
+          "defaultLogLevel": "warn"
+        },
+        "trace": {
+          "enabled": false
+        },
+        "profiler": {
+          "enabled": false
+        }
+      }
+    },
+    "inference": {
+      "prompt": {
+        "messages": [
+          {
+            "role": "user",
+            "content": "What color is the sky on a clear day?"
+          }
+        ]
+      },
+      "batching": {
+        "maxTokens": 32
+      },
+      "sampling": {
+        "temperature": 0,
+        "topK": 1,
+        "topP": 1
+      }
+    }
+  }
+}

package/src/config/presets/runtime/experiments/verify/translategemma-verify.json

@@ -0,0 +1,39 @@
+{
+  "id": "experiments/verify/translategemma-verify",
+  "name": "translategemma-verify",
+  "intent": "verify",
+  "stability": "experimental",
+  "owner": "doppler-core",
+  "createdAtUtc": "2026-03-16T00:00:00Z",
+  "extends": "modes/bench",
+  "model": "translategemma-4b-it-q4k-ehf16-af32",
+  "runtime": {
+    "shared": {
+      "tooling": {
+        "intent": "verify"
+      },
+      "debug": {
+        "logLevel": {
+          "defaultLogLevel": "warn"
+        },
+        "trace": {
+          "enabled": false
+        },
+        "profiler": {
+          "enabled": false
+        }
+      }
+    },
+    "inference": {
+      "prompt": "Hello from Doppler.",
+      "batching": {
+        "maxTokens": 32
+      },
+      "sampling": {
+        "temperature": 0,
+        "topK": 1,
+        "topP": 1
+      }
+    }
+  }
+}

package/src/config/presets/runtime/modes/trace-layers.json

@@ -6,6 +6,7 @@
   "stability": "canonical",
   "owner": "doppler-core",
   "createdAtUtc": "2026-02-25T00:00:00Z",
+  "extends": "default",
   "runtime": {
     "shared": {
       "tooling": {

package/src/config/presets/runtime/tiers/gemma4-16gb.json

@@ -0,0 +1,69 @@
+{
+  "id": "tiers/gemma4-16gb",
+  "name": "Gemma 4 — 16 GB tier (constrained)",
+  "description": "Gemma 4 MoE runtime tier for 16 GB GPU memory. Aggressively constrained: short context, minimal expert cache, hard budget enforcement. Fail-closed if budget is not met.",
+  "intent": "investigate",
+  "stability": "experimental",
+  "owner": "doppler-core",
+  "createdAtUtc": "2026-03-17T00:00:00Z",
+  "extends": "default",
+  "runtime": {
+    "shared": {
+      "bufferPool": {
+        "budget": {
+          "maxTotalBytes": 13958643712,
+          "highWatermarkRatio": 0.85,
+          "emergencyTrimTargetRatio": 0.7,
+          "hardFailOnBudgetExceeded": true
+        }
+      }
+    },
+    "loading": {
+      "expertCache": {
+        "defaultSizeBytes": 1073741824,
+        "maxBufferPercentage": 0.15,
+        "evictionHighWatermark": 0.8,
+        "emergencyTrimToRatio": 0.65
+      },
+      "prefetch": {
+        "enabled": true,
+        "layersAhead": 1,
+        "maxShards": 4
+      },
+      "memoryManagement": {
+        "flushIntervalLayers": 1,
+        "flushThresholdBytes": 134217728
+      }
+    },
+    "inference": {
+      "kvcache": {
+        "layout": "contiguous",
+        "maxSeqLen": 2048,
+        "kvDtype": "f16",
+        "pageSize": 128,
+        "tiering": {
+          "mode": "off"
+        }
+      },
+      "moe": {
+        "routing": {
+          "routerDtype": "f32"
+        },
+        "cache": {
+          "dequantCacheMaxEntries": 2
+        }
+      },
+      "compute": {
+        "activationDtype": "f32"
+      },
+      "batching": {
+        "maxTokens": 512
+      },
+      "session": {
+        "kvcache": {
+          "kvDtype": "f16"
+        }
+      }
+    }
+  }
+}

package/src/config/presets/runtime/tiers/gemma4-24gb.json

@@ -0,0 +1,66 @@
+{
+  "id": "tiers/gemma4-24gb",
+  "name": "Gemma 4 — 24 GB tier",
+  "description": "Gemma 4 MoE runtime tier for 24 GB GPU memory. Moderate expert cache, contiguous KV, reduced context length.",
+  "intent": "investigate",
+  "stability": "experimental",
+  "owner": "doppler-core",
+  "createdAtUtc": "2026-03-17T00:00:00Z",
+  "extends": "default",
+  "runtime": {
+    "shared": {
+      "bufferPool": {
+        "budget": {
+          "maxTotalBytes": 21474836480,
+          "highWatermarkRatio": 0.9,
+          "emergencyTrimTargetRatio": 0.75,
+          "hardFailOnBudgetExceeded": true
+        }
+      }
+    },
+    "loading": {
+      "expertCache": {
+        "defaultSizeBytes": 3221225472,
+        "maxBufferPercentage": 0.2,
+        "evictionHighWatermark": 0.85,
+        "emergencyTrimToRatio": 0.7
+      },
+      "prefetch": {
+        "enabled": true,
+        "layersAhead": 1,
+        "maxShards": 8
+      },
+      "memoryManagement": {
+        "flushIntervalLayers": 2,
+        "flushThresholdBytes": 268435456
+      }
+    },
+    "inference": {
+      "kvcache": {
+        "layout": "contiguous",
+        "maxSeqLen": 4096,
+        "kvDtype": "f16",
+        "pageSize": 256,
+        "tiering": {
+          "mode": "off"
+        }
+      },
+      "moe": {
+        "routing": {
+          "routerDtype": "f32"
+        },
+        "cache": {
+          "dequantCacheMaxEntries": 64
+        }
+      },
+      "compute": {
+        "activationDtype": "f32"
+      },
+      "session": {
+        "kvcache": {
+          "kvDtype": "f16"
+        }
+      }
+    }
+  }
+}

package/src/config/presets/runtime/tiers/gemma4-32gb.json

@@ -0,0 +1,66 @@
+{
+  "id": "tiers/gemma4-32gb",
+  "name": "Gemma 4 — 32 GB tier",
+  "description": "Gemma 4 MoE runtime tier for 32 GB GPU memory. Generous expert cache, contiguous KV, full-length context.",
+  "intent": "investigate",
+  "stability": "experimental",
+  "owner": "doppler-core",
+  "createdAtUtc": "2026-03-17T00:00:00Z",
+  "extends": "default",
+  "runtime": {
+    "shared": {
+      "bufferPool": {
+        "budget": {
+          "maxTotalBytes": 30064771072,
+          "highWatermarkRatio": 0.9,
+          "emergencyTrimTargetRatio": 0.75,
+          "hardFailOnBudgetExceeded": true
+        }
+      }
+    },
+    "loading": {
+      "expertCache": {
+        "defaultSizeBytes": 6442450944,
+        "maxBufferPercentage": 0.25,
+        "evictionHighWatermark": 0.9,
+        "emergencyTrimToRatio": 0.75
+      },
+      "prefetch": {
+        "enabled": true,
+        "layersAhead": 2,
+        "maxShards": 16
+      },
+      "memoryManagement": {
+        "flushIntervalLayers": 4,
+        "flushThresholdBytes": 536870912
+      }
+    },
+    "inference": {
+      "kvcache": {
+        "layout": "contiguous",
+        "maxSeqLen": 8192,
+        "kvDtype": "f16",
+        "pageSize": 256,
+        "tiering": {
+          "mode": "off"
+        }
+      },
+      "moe": {
+        "routing": {
+          "routerDtype": "f32"
+        },
+        "cache": {
+          "dequantCacheMaxEntries": 128
+        }
+      },
+      "compute": {
+        "activationDtype": "f32"
+      },
+      "session": {
+        "kvcache": {
+          "kvDtype": "f16"
+        }
+      }
+    }
+  }
+}

package/src/config/runtime.js

@@ -58,4 +58,7 @@ function assertNoDeprecatedRuntimeKeys(overrides) {
   if (inference?.sampling?.maxTokens !== undefined) {
     throw new Error('sampling.maxTokens is removed; use inference.batching.maxTokens');
   }
+  if (inference?.session?.maxNewTokens !== undefined) {
+    throw new Error('inference.session.maxNewTokens is not a supported runtime config key; use inference.batching.maxTokens');
+  }
 }

package/src/config/schema/debug.schema.d.ts

@@ -131,6 +131,44 @@ export interface PipelineDebugConfigSchema {
 /** Default pipeline debug configuration */
 export declare const DEFAULT_PIPELINE_DEBUG_CONFIG: PipelineDebugConfigSchema;
 
+/** Loader debug configuration (Q4K dequant and related probes). */
+export interface LoaderDebugConfigSchema {
+  /** Enable loader debug behavior (default: false) */
+  enabled: boolean;
+  /** Force GPU dequant for Q4K tensors even when CPU fallback is eligible. */
+  forceGpuDequant: boolean;
+  /** Prefer CPU dequant for F32 output when eligible (default: false, GPU is preferred). */
+  preferCpuDequant: boolean;
+  /** Throw when CPU dequant fallback is taken. */
+  failOnCpuDequantPath: boolean;
+  /** Enable dtype-aware GPU-vs-CPU parity checks during Q4K dequant. */
+  runQ4KDequantParity: boolean;
+  /** Number of values to read back for parity checks. */
+  q4kDequantParitySamples: number;
+}
+
+/** Default loader debug configuration. */
+export declare const DEFAULT_LOADER_DEBUG_CONFIG: LoaderDebugConfigSchema;
+
+/** Matmul debug configuration (attention split/shape diagnostics). */
+export interface MatmulDebugConfigSchema {
+  /** Enable matmul debug behavior (default: false) */
+  enabled: boolean;
+  /** Force split (non-fused) Q/K/V projection path for diagnostics. */
+  forceSplitQKV: boolean;
+  /** Validate B tensor layout/buffer bytes for attention projection roles. */
+  validateAttentionWeightBuffer: boolean;
+  /** Throw if validation fails due small B tensor. */
+  failOnSmallAttentionWeightBuffer: boolean;
+  /** Emit attention B-buffer diagnostics. */
+  logAttentionWeightBuffer: boolean;
+  /** Log first-8 projection output values for layer 0 decode (diagnostic). */
+  logProjectionValues: boolean;
+}
+
+/** Default matmul debug configuration. */
+export declare const DEFAULT_MATMUL_DEBUG_CONFIG: MatmulDebugConfigSchema;
+
 /**
  * Profiler configuration.
  */
@@ -241,6 +279,8 @@ export interface DebugConfigSchema {
   logLevel: LogLevelConfigSchema;
   trace: TraceConfigSchema;
   pipeline: PipelineDebugConfigSchema;
+  loader: LoaderDebugConfigSchema;
+  matmul: MatmulDebugConfigSchema;
   probes: ProbeConfigSchema[];
   profiler: ProfilerConfigSchema;
   perfGuards: PerfGuardsConfigSchema;

package/src/config/schema/debug.schema.js

@@ -38,6 +38,32 @@ export const DEFAULT_TRACE_CONFIG = {
   file: null,
 };
 
+// =============================================================================
+// Loader Debug Config
+// =============================================================================
+
+export const DEFAULT_LOADER_DEBUG_CONFIG = {
+  enabled: false,
+  forceGpuDequant: false,
+  preferCpuDequant: false,
+  failOnCpuDequantPath: false,
+  runQ4KDequantParity: false,
+  q4kDequantParitySamples: 256,
+};
+
+// =============================================================================
+// Kernel Debug Config
+// =============================================================================
+
+export const DEFAULT_MATMUL_DEBUG_CONFIG = {
+  enabled: false,
+  forceSplitQKV: false,
+  validateAttentionWeightBuffer: false,
+  failOnSmallAttentionWeightBuffer: false,
+  logAttentionWeightBuffer: false,
+  logProjectionValues: false,
+};
+
 // =============================================================================
 // Kernel Trace Config (kernel-trace.js anomaly detection)
 // =============================================================================
@@ -100,6 +126,8 @@ export const DEFAULT_DEBUG_CONFIG = {
   logLevel: DEFAULT_LOG_LEVEL_CONFIG,
   trace: DEFAULT_TRACE_CONFIG,
   pipeline: DEFAULT_PIPELINE_DEBUG_CONFIG,
+  loader: DEFAULT_LOADER_DEBUG_CONFIG,
+  matmul: DEFAULT_MATMUL_DEBUG_CONFIG,
   probes: [],
   profiler: DEFAULT_PROFILER_CONFIG,
   perfGuards: DEFAULT_PERF_GUARDS_CONFIG,

package/src/config/schema/index.js

@@ -217,6 +217,8 @@ export {
   DEFAULT_LOG_HISTORY_CONFIG,
   DEFAULT_LOG_LEVEL_CONFIG,
   DEFAULT_TRACE_CONFIG,
+  DEFAULT_LOADER_DEBUG_CONFIG,
+  DEFAULT_MATMUL_DEBUG_CONFIG,
   DEFAULT_KERNEL_TRACE_CONFIG,
   DEFAULT_PIPELINE_DEBUG_CONFIG,
   DEFAULT_PROFILER_CONFIG,

package/src/config/schema/inference-defaults.schema.js

@@ -93,7 +93,7 @@ export const DEFAULT_TOKENIZER_DEFAULTS = {
 // =============================================================================
 
 export const DEFAULT_CHAT_TEMPLATE_CONFIG = {
-  enabled: false,
+  enabled: undefined,
 };
 
 export const DEFAULT_KERNEL_PATH_POLICY = {

package/src/config/schema/kernel-path.schema.d.ts

@@ -160,6 +160,7 @@ export type BuiltinKernelPathId =
   | 'gemma3-f16-fused-f32a-online-streamingprefill' // Gemma 3 F16 fused FFN online path with streaming prefill attention
   | 'gemma3-q4k-dequant-f16a-online' // Gemma 3 Q4K dequant online path (F16 activations)
   | 'gemma3-q4k-dequant-f32a-online' // Gemma 3 Q4K dequant online path with F32 activations
+  | 'gemma3-q4k-dequant-f32w-f32a-online' // Gemma 3 Q4K path with F32 projection weights and F32 activations
   | 'gemma3-q4k-dequant-f32a-nosubgroups' // Gemma 3 Q4K dequant path with no subgroup requirement
   | 'gemma3-q4k-dequant-f32a' // Legacy alias for gemma3-q4k-dequant-f32a-nosubgroups
   | 'lfm2-q4k-dequant-f32a-online' // LFM2 Q4K path with F32 activations and fast prefill

package/src/config/schema/memory-limits.schema.js

@@ -40,8 +40,8 @@ export const DEFAULT_SEGMENT_ALLOCATION_CONFIG = {
 // =============================================================================
 
 export const DEFAULT_EMULATED_STORAGE_CONFIG = {
-  vramBudgetBytes: 2 * GB,
-  ramBudgetBytes: 8 * GB,
+  vramBudgetBytes: 4 * GB,
+  ramBudgetBytes: 16 * GB,
 };
 
 // =============================================================================