@modular-prompt/driver 0.13.5 → 0.15.0
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- package/README.md +117 -4
- package/dist/driver-registry/ai-service.d.ts +23 -1
- package/dist/driver-registry/ai-service.d.ts.map +1 -1
- package/dist/driver-registry/ai-service.js +44 -10
- package/dist/driver-registry/ai-service.js.map +1 -1
- package/dist/driver-registry/config-based-factory.d.ts.map +1 -1
- package/dist/driver-registry/config-based-factory.js +16 -0
- package/dist/driver-registry/config-based-factory.js.map +1 -1
- package/dist/driver-registry/factory-helper.d.ts +2 -0
- package/dist/driver-registry/factory-helper.d.ts.map +1 -1
- package/dist/driver-registry/factory-helper.js +21 -3
- package/dist/driver-registry/factory-helper.js.map +1 -1
- package/dist/driver-registry/index.d.ts +2 -2
- package/dist/driver-registry/index.d.ts.map +1 -1
- package/dist/driver-registry/index.js +1 -1
- package/dist/driver-registry/index.js.map +1 -1
- package/dist/driver-registry/registry.d.ts.map +1 -1
- package/dist/driver-registry/registry.js +3 -1
- package/dist/driver-registry/registry.js.map +1 -1
- package/dist/driver-registry/types.d.ts +8 -2
- package/dist/driver-registry/types.d.ts.map +1 -1
- package/dist/index.d.ts +11 -2
- package/dist/index.d.ts.map +1 -1
- package/dist/index.js +11 -1
- package/dist/index.js.map +1 -1
- package/dist/local-inference/adapters.d.ts +66 -0
- package/dist/local-inference/adapters.d.ts.map +1 -0
- package/dist/local-inference/adapters.js +2 -0
- package/dist/local-inference/adapters.js.map +1 -0
- package/dist/local-inference/driver.d.ts +51 -0
- package/dist/local-inference/driver.d.ts.map +1 -0
- package/dist/local-inference/driver.js +309 -0
- package/dist/local-inference/driver.js.map +1 -0
- package/dist/local-inference/index.d.ts +22 -0
- package/dist/local-inference/index.d.ts.map +1 -0
- package/dist/local-inference/index.js +17 -0
- package/dist/local-inference/index.js.map +1 -0
- package/dist/local-inference/process-client.d.ts +50 -0
- package/dist/local-inference/process-client.d.ts.map +1 -0
- package/dist/local-inference/process-client.js +92 -0
- package/dist/local-inference/process-client.js.map +1 -0
- package/dist/local-inference/process-communication.d.ts +41 -0
- package/dist/local-inference/process-communication.d.ts.map +1 -0
- package/dist/{mlx-ml/process → local-inference}/process-communication.js +40 -47
- package/dist/local-inference/process-communication.js.map +1 -0
- package/dist/local-inference/process-port.d.ts +12 -0
- package/dist/local-inference/process-port.d.ts.map +1 -0
- package/dist/local-inference/process-port.js +2 -0
- package/dist/local-inference/process-port.js.map +1 -0
- package/dist/local-inference/prompt-utils.d.ts +6 -0
- package/dist/local-inference/prompt-utils.d.ts.map +1 -0
- package/dist/local-inference/prompt-utils.js +17 -0
- package/dist/local-inference/prompt-utils.js.map +1 -0
- package/dist/local-inference/protocol.d.ts +192 -0
- package/dist/local-inference/protocol.d.ts.map +1 -0
- package/dist/local-inference/protocol.js +2 -0
- package/dist/local-inference/protocol.js.map +1 -0
- package/dist/local-inference/queue-types.d.ts +54 -0
- package/dist/local-inference/queue-types.d.ts.map +1 -0
- package/dist/local-inference/queue-types.js +2 -0
- package/dist/local-inference/queue-types.js.map +1 -0
- package/dist/local-inference/request-queue.d.ts +36 -0
- package/dist/local-inference/request-queue.d.ts.map +1 -0
- package/dist/{mlx-ml/process/queue.js → local-inference/request-queue.js} +89 -56
- package/dist/local-inference/request-queue.js.map +1 -0
- package/dist/local-inference/stream-utils.d.ts +19 -0
- package/dist/local-inference/stream-utils.d.ts.map +1 -0
- package/dist/local-inference/stream-utils.js +76 -0
- package/dist/local-inference/stream-utils.js.map +1 -0
- package/dist/mlx-ml/mlx-cache-support.d.ts +23 -0
- package/dist/mlx-ml/mlx-cache-support.d.ts.map +1 -0
- package/dist/mlx-ml/mlx-cache-support.js +45 -0
- package/dist/mlx-ml/mlx-cache-support.js.map +1 -0
- package/dist/mlx-ml/mlx-driver.d.ts +20 -59
- package/dist/mlx-ml/mlx-driver.d.ts.map +1 -1
- package/dist/mlx-ml/mlx-driver.js +86 -415
- package/dist/mlx-ml/mlx-driver.js.map +1 -1
- package/dist/mlx-ml/mlx-local-inference-adapters.d.ts +3 -0
- package/dist/mlx-ml/mlx-local-inference-adapters.d.ts.map +1 -0
- package/dist/mlx-ml/mlx-local-inference-adapters.js +19 -0
- package/dist/mlx-ml/mlx-local-inference-adapters.js.map +1 -0
- package/dist/mlx-ml/mlx-options.d.ts +19 -0
- package/dist/mlx-ml/mlx-options.d.ts.map +1 -0
- package/dist/mlx-ml/mlx-options.js +30 -0
- package/dist/mlx-ml/mlx-options.js.map +1 -0
- package/dist/mlx-ml/process/index.d.ts +11 -8
- package/dist/mlx-ml/process/index.d.ts.map +1 -1
- package/dist/mlx-ml/process/index.js +75 -52
- package/dist/mlx-ml/process/index.js.map +1 -1
- package/dist/mlx-ml/process/model-specific.d.ts +2 -1
- package/dist/mlx-ml/process/model-specific.d.ts.map +1 -1
- package/dist/mlx-ml/process/model-specific.js.map +1 -1
- package/dist/mlx-ml/process/prompt-builder.d.ts +11 -0
- package/dist/mlx-ml/process/prompt-builder.d.ts.map +1 -0
- package/dist/mlx-ml/process/prompt-builder.js +51 -0
- package/dist/mlx-ml/process/prompt-builder.js.map +1 -0
- package/dist/mlx-ml/process/types.d.ts +15 -183
- package/dist/mlx-ml/process/types.d.ts.map +1 -1
- package/dist/mlx-ml/tool-call-parser/tool-formatter.js +2 -2
- package/dist/mlx-ml/tool-call-parser/tool-formatter.js.map +1 -1
- package/dist/mlx-ml/types.d.ts +2 -45
- package/dist/mlx-ml/types.d.ts.map +1 -1
- package/dist/models-config/index.d.ts +8 -0
- package/dist/models-config/index.d.ts.map +1 -0
- package/dist/models-config/index.js +7 -0
- package/dist/models-config/index.js.map +1 -0
- package/dist/models-config/loader.d.ts +20 -0
- package/dist/models-config/loader.d.ts.map +1 -0
- package/dist/models-config/loader.js +85 -0
- package/dist/models-config/loader.js.map +1 -0
- package/dist/models-config/paths.d.ts +7 -0
- package/dist/models-config/paths.d.ts.map +1 -0
- package/dist/models-config/paths.js +11 -0
- package/dist/models-config/paths.js.map +1 -0
- package/dist/models-config/resolve.d.ts +57 -0
- package/dist/models-config/resolve.d.ts.map +1 -0
- package/dist/models-config/resolve.js +187 -0
- package/dist/models-config/resolve.js.map +1 -0
- package/dist/models-config/types.d.ts +60 -0
- package/dist/models-config/types.d.ts.map +1 -0
- package/dist/models-config/types.js +5 -0
- package/dist/models-config/types.js.map +1 -0
- package/dist/pytorch/process/index.d.ts +35 -0
- package/dist/pytorch/process/index.d.ts.map +1 -0
- package/dist/pytorch/process/index.js +69 -0
- package/dist/pytorch/process/index.js.map +1 -0
- package/dist/pytorch/pytorch-driver.d.ts +35 -0
- package/dist/pytorch/pytorch-driver.d.ts.map +1 -0
- package/dist/pytorch/pytorch-driver.js +48 -0
- package/dist/pytorch/pytorch-driver.js.map +1 -0
- package/dist/pytorch/pytorch-local-inference-adapters.d.ts +3 -0
- package/dist/pytorch/pytorch-local-inference-adapters.d.ts.map +1 -0
- package/dist/pytorch/pytorch-local-inference-adapters.js +19 -0
- package/dist/pytorch/pytorch-local-inference-adapters.js.map +1 -0
- package/dist/pytorch/pytorch-options.d.ts +8 -0
- package/dist/pytorch/pytorch-options.d.ts.map +1 -0
- package/dist/pytorch/pytorch-options.js +21 -0
- package/dist/pytorch/pytorch-options.js.map +1 -0
- package/dist/query-logger.js +1 -1
- package/dist/query-logger.js.map +1 -1
- package/dist/query-utils.d.ts +29 -0
- package/dist/query-utils.d.ts.map +1 -0
- package/dist/query-utils.js +61 -0
- package/dist/query-utils.js.map +1 -0
- package/dist/runtime/check.d.ts +10 -0
- package/dist/runtime/check.d.ts.map +1 -0
- package/dist/runtime/check.js +24 -0
- package/dist/runtime/check.js.map +1 -0
- package/dist/runtime/index.d.ts +4 -0
- package/dist/runtime/index.d.ts.map +1 -0
- package/dist/runtime/index.js +4 -0
- package/dist/runtime/index.js.map +1 -0
- package/dist/runtime/manifest-core.d.mts +27 -0
- package/dist/runtime/manifest-core.d.mts.map +1 -0
- package/dist/runtime/manifest-core.mjs +68 -0
- package/dist/runtime/manifest-core.mjs.map +1 -0
- package/dist/runtime/manifest.d.ts +18 -0
- package/dist/runtime/manifest.d.ts.map +1 -0
- package/dist/runtime/manifest.js +9 -0
- package/dist/runtime/manifest.js.map +1 -0
- package/dist/runtime/paths-core.d.mts +17 -0
- package/dist/runtime/paths-core.d.mts.map +1 -0
- package/dist/runtime/paths-core.mjs +67 -0
- package/dist/runtime/paths-core.mjs.map +1 -0
- package/dist/runtime/paths.d.ts +12 -0
- package/dist/runtime/paths.d.ts.map +1 -0
- package/dist/runtime/paths.js +17 -0
- package/dist/runtime/paths.js.map +1 -0
- package/dist/types.d.ts +20 -1
- package/dist/types.d.ts.map +1 -1
- package/dist/types.js.map +1 -1
- package/package.json +8 -5
- package/scripts/download-model.js +25 -9
- package/scripts/runtime-cli.js +315 -0
- package/skills/driver-usage/SKILL.md +56 -1
- package/src/mlx-ml/python/__main__.py +43 -4
- package/src/mlx-ml/python/handlers/__init__.py +2 -1
- package/src/mlx-ml/python/handlers/cancel.py +53 -0
- package/src/mlx-ml/python/handlers/completion.py +3 -24
- package/src/mlx-ml/python/handlers/{chat.py → generate.py} +36 -106
- package/src/mlx-ml/python/handlers/render.py +40 -0
- package/src/mlx-ml/python/pyproject.toml +3 -2
- package/src/mlx-ml/python/server.py +35 -8
- package/src/mlx-ml/python/utils/template_render.py +80 -0
- package/src/mlx-ml/python/utils/token_utils.py +2 -2
- package/src/mlx-ml/python/uv.lock +549 -454
- package/src/pytorch/python/__main__.py +19 -0
- package/src/pytorch/python/backends/__init__.py +3 -0
- package/src/pytorch/python/backends/base.py +84 -0
- package/src/pytorch/python/backends/transformers_lm.py +127 -0
- package/src/pytorch/python/handlers/__init__.py +6 -0
- package/src/pytorch/python/handlers/cancel.py +53 -0
- package/src/pytorch/python/handlers/capabilities.py +6 -0
- package/src/pytorch/python/handlers/completion.py +15 -0
- package/src/pytorch/python/handlers/format_test.py +70 -0
- package/src/pytorch/python/handlers/generate.py +68 -0
- package/src/pytorch/python/handlers/render.py +40 -0
- package/src/pytorch/python/handlers/tokenize.py +63 -0
- package/src/pytorch/python/pyproject.toml +36 -0
- package/src/pytorch/python/server.py +140 -0
- package/src/pytorch/python/utils/__init__.py +0 -0
- package/src/pytorch/python/utils/chat_template_constraints.py +164 -0
- package/src/pytorch/python/utils/prompt_builder.py +54 -0
- package/src/pytorch/python/utils/template_render.py +80 -0
- package/src/pytorch/python/utils/token_utils.py +376 -0
- package/src/pytorch/python/uv.lock +694 -0
- package/dist/mlx-ml/process/process-communication.d.ts +0 -37
- package/dist/mlx-ml/process/process-communication.d.ts.map +0 -1
- package/dist/mlx-ml/process/process-communication.js.map +0 -1
- package/dist/mlx-ml/process/queue.d.ts +0 -33
- package/dist/mlx-ml/process/queue.d.ts.map +0 -1
- package/dist/mlx-ml/process/queue.js.map +0 -1
- package/scripts/setup-mlx.js +0 -53
|
@@ -0,0 +1,50 @@
|
|
|
1
|
+
/**
|
|
2
|
+
* Local Inference Protocol プロセスクライアント
|
|
3
|
+
*
|
|
4
|
+
* stdio JSON-RPC 通信とリクエストキューを統合する。
|
|
5
|
+
*/
|
|
6
|
+
import { Readable } from 'stream';
|
|
7
|
+
import type { RuntimeProfile } from '../runtime/index.js';
|
|
8
|
+
import type { InferenceCapabilities, InferenceFormatTestResult, InferenceRenderResult, InferenceTokenizeResult, InferenceCachePrefillResult, InferenceMessage, InferenceToolDefinition, InferenceSamplingOptions } from './protocol.js';
|
|
9
|
+
import { type SamplingOptionsMapper } from './request-queue.js';
|
|
10
|
+
export interface InferenceProcessClientConfig {
|
|
11
|
+
modelName: string;
|
|
12
|
+
/** uv --project に渡す Python プロジェクトディレクトリ */
|
|
13
|
+
pythonProjectDir: string;
|
|
14
|
+
/** UV_PROJECT_ENVIRONMENT に設定する venv パス */
|
|
15
|
+
venvPath: string;
|
|
16
|
+
/** 指定時は起動前に runtime venv の存在を確認する */
|
|
17
|
+
runtimeProfile?: RuntimeProfile;
|
|
18
|
+
extraSpawnArgs?: string[];
|
|
19
|
+
extraEnv?: Record<string, string>;
|
|
20
|
+
loggerPrefix?: string;
|
|
21
|
+
mapSamplingOptions?: SamplingOptionsMapper;
|
|
22
|
+
processExitErrorMessage?: (code: number | null, signal: string | null) => string;
|
|
23
|
+
}
|
|
24
|
+
export declare class InferenceProcessClient {
|
|
25
|
+
readonly modelName: string;
|
|
26
|
+
private requestQueue;
|
|
27
|
+
private processComm;
|
|
28
|
+
constructor(config: InferenceProcessClientConfig);
|
|
29
|
+
ensureInitialized(): Promise<void>;
|
|
30
|
+
getCapabilities(): Promise<InferenceCapabilities>;
|
|
31
|
+
formatTest(messages: InferenceMessage[], options?: {
|
|
32
|
+
primer?: string;
|
|
33
|
+
}): Promise<InferenceFormatTestResult>;
|
|
34
|
+
render(messages: InferenceMessage[], options?: InferenceSamplingOptions & {
|
|
35
|
+
primer?: string;
|
|
36
|
+
}, tools?: InferenceToolDefinition[], reasoningEffort?: 'low' | 'medium' | 'high'): Promise<InferenceRenderResult>;
|
|
37
|
+
tokenize(messages: InferenceMessage[], tools?: InferenceToolDefinition[], reasoningEffort?: 'low' | 'medium' | 'high'): Promise<InferenceTokenizeResult>;
|
|
38
|
+
cachePrefill(cachePath: string, messages: InferenceMessage[], baseCachePath?: string, trimToTokens?: number, prefixOffsets?: number[], prefixHashes?: string[], tools?: InferenceToolDefinition[], reasoningEffort?: 'low' | 'medium' | 'high'): Promise<InferenceCachePrefillResult>;
|
|
39
|
+
completion(prompt: string, options?: unknown, images?: string[], maxImageSize?: number): Promise<Readable>;
|
|
40
|
+
generate(prompt: string | number[], options?: unknown, images?: string[], maxImageSize?: number, cachePath?: string, cacheTrimTokens?: number, primer?: string): Promise<Readable>;
|
|
41
|
+
exit(): Promise<void>;
|
|
42
|
+
cancelActiveRequest(): void;
|
|
43
|
+
getStatus(): {
|
|
44
|
+
modelName: string;
|
|
45
|
+
queueLength: number;
|
|
46
|
+
isStreamingActive: boolean;
|
|
47
|
+
isJsonBuffering: boolean;
|
|
48
|
+
};
|
|
49
|
+
}
|
|
50
|
+
//# sourceMappingURL=process-client.d.ts.map
|
|
@@ -0,0 +1 @@
|
|
|
1
|
+
{"version":3,"file":"process-client.d.ts","sourceRoot":"","sources":["../../src/local-inference/process-client.ts"],"names":[],"mappings":"AAAA;;;;GAIG;AAEH,OAAO,EAAE,QAAQ,EAAE,MAAM,QAAQ,CAAC;AAElC,OAAO,KAAK,EAAE,cAAc,EAAE,MAAM,qBAAqB,CAAC;AAE1D,OAAO,KAAK,EACV,qBAAqB,EACrB,yBAAyB,EACzB,qBAAqB,EACrB,uBAAuB,EACvB,2BAA2B,EAC3B,gBAAgB,EAChB,uBAAuB,EACvB,wBAAwB,EACzB,MAAM,eAAe,CAAC;AAKvB,OAAO,EAGL,KAAK,qBAAqB,EAC3B,MAAM,oBAAoB,CAAC;AAE5B,MAAM,WAAW,4BAA4B;IAC3C,SAAS,EAAE,MAAM,CAAC;IAClB,2CAA2C;IAC3C,gBAAgB,EAAE,MAAM,CAAC;IACzB,2CAA2C;IAC3C,QAAQ,EAAE,MAAM,CAAC;IACjB,qCAAqC;IACrC,cAAc,CAAC,EAAE,cAAc,CAAC;IAChC,cAAc,CAAC,EAAE,MAAM,EAAE,CAAC;IAC1B,QAAQ,CAAC,EAAE,MAAM,CAAC,MAAM,EAAE,MAAM,CAAC,CAAC;IAClC,YAAY,CAAC,EAAE,MAAM,CAAC;IACtB,kBAAkB,CAAC,EAAE,qBAAqB,CAAC;IAC3C,uBAAuB,CAAC,EAAE,CAAC,IAAI,EAAE,MAAM,GAAG,IAAI,EAAE,MAAM,EAAE,MAAM,GAAG,IAAI,KAAK,MAAM,CAAC;CAClF;AAED,qBAAa,sBAAsB;IACjC,QAAQ,CAAC,SAAS,EAAE,MAAM,CAAC;IAE3B,OAAO,CAAC,YAAY,CAAwB;IAC5C,OAAO,CAAC,WAAW,CAAuB;gBAE9B,MAAM,EAAE,4BAA4B;IAkD1C,iBAAiB,IAAI,OAAO,CAAC,IAAI,CAAC;IAIlC,eAAe,IAAI,OAAO,CAAC,qBAAqB,CAAC;IAIjD,UAAU,CACd,QAAQ,EAAE,gBAAgB,EAAE,EAC5B,OAAO,CAAC,EAAE;QAAE,MAAM,CAAC,EAAE,MAAM,CAAA;KAAE,GAC5B,OAAO,CAAC,yBAAyB,CAAC;IAI/B,MAAM,CACV,QAAQ,EAAE,gBAAgB,EAAE,EAC5B,OAAO,CAAC,EAAE,wBAAwB,GAAG;QAAE,MAAM,CAAC,EAAE,MAAM,CAAA;KAAE,EACxD,KAAK,CAAC,EAAE,uBAAuB,EAAE,EACjC,eAAe,CAAC,EAAE,KAAK,GAAG,QAAQ,GAAG,MAAM,GAC1C,OAAO,CAAC,qBAAqB,CAAC;IAI3B,QAAQ,CACZ,QAAQ,EAAE,gBAAgB,EAAE,EAC5B,KAAK,CAAC,EAAE,uBAAuB,EAAE,EACjC,eAAe,CAAC,EAAE,KAAK,GAAG,QAAQ,GAAG,MAAM,GAC1C,OAAO,CAAC,uBAAuB,CAAC;IAI7B,YAAY,CAChB,SAAS,EAAE,MAAM,EACjB,QAAQ,EAAE,gBAAgB,EAAE,EAC5B,aAAa,CAAC,EAAE,MAAM,EACtB,YAAY,CAAC,EAAE,MAAM,EACrB,aAAa,CAAC,EAAE,MAAM,EAAE,EACxB,YAAY,CAAC,EAAE,MAAM,EAAE,EACvB,KAAK,CAAC,EAAE,uBAAuB,EAAE,EACjC,eAAe,CAAC,EAAE,KAAK,GAAG,QAAQ,GAAG,MAAM,GAC1C,OAAO,CAAC,2BAA2B,CAAC;IAajC,UAAU,CACd,MAAM,EAAE,MAAM,EACd,OAAO,CAAC,EAAE,OAAO,EACjB,MAAM,CAAC,EAAE,MAAM,EAAE,EACjB,YAAY,CAAC,EAAE,MAAM,GACpB,OAAO,CAAC,QAAQ,CAAC;IAId,QAAQ,CACZ,MAAM,EAAE,MAAM,GAAG,MAAM,EAAE,EACzB,OAAO,CAAC,EAAE,OAAO,EACjB,MAAM,CAAC,EAAE,MAAM,EAAE,EACjB,YAAY,CAAC,EAAE,MAAM,EACrB,SAAS,CAAC,EAAE,MAAM,EAClB,eAAe,CAAC,EAAE,MAAM,EACxB,MAAM,CAAC,EAAE,MAAM,GACd,OAAO,CAAC,QAAQ,CAAC;IAYd,IAAI,IAAI,OAAO,CAAC,IAAI,CAAC;IAI3B,mBAAmB,IAAI,IAAI;IAI3B,SAAS;;;;;;CAQV"}
|
|
@@ -0,0 +1,92 @@
|
|
|
1
|
+
/**
|
|
2
|
+
* Local Inference Protocol プロセスクライアント
|
|
3
|
+
*
|
|
4
|
+
* stdio JSON-RPC 通信とリクエストキューを統合する。
|
|
5
|
+
*/
|
|
6
|
+
import { Logger } from '@modular-prompt/utils';
|
|
7
|
+
import { assertRuntimeReady } from '../runtime/index.js';
|
|
8
|
+
import { ProcessCommunication, } from './process-communication.js';
|
|
9
|
+
import { InferenceRequestQueue, } from './request-queue.js';
|
|
10
|
+
export class InferenceProcessClient {
|
|
11
|
+
modelName;
|
|
12
|
+
requestQueue;
|
|
13
|
+
processComm;
|
|
14
|
+
constructor(config) {
|
|
15
|
+
this.modelName = config.modelName;
|
|
16
|
+
if (config.runtimeProfile) {
|
|
17
|
+
assertRuntimeReady(config.runtimeProfile);
|
|
18
|
+
}
|
|
19
|
+
const logger = new Logger({
|
|
20
|
+
prefix: config.loggerPrefix ?? 'LIP',
|
|
21
|
+
context: 'process',
|
|
22
|
+
});
|
|
23
|
+
// processCallbacks は this.requestQueue 代入前に定義するが、
|
|
24
|
+
// コールバックは子プロセスの stdout/exit など非同期イベント時のみ呼ばれるため安全。
|
|
25
|
+
const processCallbacks = {
|
|
26
|
+
onJsonResponse: (jsonData) => this.requestQueue.handleJsonResponse(jsonData),
|
|
27
|
+
onRequestCompleted: () => this.requestQueue.onRequestCompleted(),
|
|
28
|
+
onProcessExit: (code, signal) => {
|
|
29
|
+
if (code !== 0) {
|
|
30
|
+
const message = config.processExitErrorMessage?.(code, signal) ??
|
|
31
|
+
`Inference process exited unexpectedly (code=${code}, signal=${signal})`;
|
|
32
|
+
logger.error(message);
|
|
33
|
+
this.requestQueue.rejectAll(new Error(message));
|
|
34
|
+
}
|
|
35
|
+
},
|
|
36
|
+
};
|
|
37
|
+
const queueCallbacks = {
|
|
38
|
+
sendToProcess: (data) => this.processComm.sendToProcess(data),
|
|
39
|
+
createNewStream: () => this.processComm.createNewStream(),
|
|
40
|
+
cancelActiveStream: () => this.processComm.cancelActiveStream(),
|
|
41
|
+
};
|
|
42
|
+
this.processComm = new ProcessCommunication({
|
|
43
|
+
pythonProjectDir: config.pythonProjectDir,
|
|
44
|
+
venvPath: config.venvPath,
|
|
45
|
+
modelName: config.modelName,
|
|
46
|
+
extraArgs: config.extraSpawnArgs,
|
|
47
|
+
extraEnv: config.extraEnv,
|
|
48
|
+
loggerPrefix: config.loggerPrefix,
|
|
49
|
+
processExitErrorMessage: config.processExitErrorMessage,
|
|
50
|
+
}, processCallbacks);
|
|
51
|
+
this.requestQueue = new InferenceRequestQueue(queueCallbacks, config.mapSamplingOptions);
|
|
52
|
+
}
|
|
53
|
+
async ensureInitialized() {
|
|
54
|
+
// 互換性のため維持
|
|
55
|
+
}
|
|
56
|
+
async getCapabilities() {
|
|
57
|
+
return this.requestQueue.addCapabilitiesRequest();
|
|
58
|
+
}
|
|
59
|
+
async formatTest(messages, options) {
|
|
60
|
+
return this.requestQueue.addFormatTestRequest(messages, options);
|
|
61
|
+
}
|
|
62
|
+
async render(messages, options, tools, reasoningEffort) {
|
|
63
|
+
return this.requestQueue.addRenderRequest(messages, options, tools, reasoningEffort);
|
|
64
|
+
}
|
|
65
|
+
async tokenize(messages, tools, reasoningEffort) {
|
|
66
|
+
return this.requestQueue.addTokenizeRequest(messages, tools, reasoningEffort);
|
|
67
|
+
}
|
|
68
|
+
async cachePrefill(cachePath, messages, baseCachePath, trimToTokens, prefixOffsets, prefixHashes, tools, reasoningEffort) {
|
|
69
|
+
return this.requestQueue.addCachePrefillRequest(cachePath, messages, baseCachePath, trimToTokens, prefixOffsets, prefixHashes, tools, reasoningEffort);
|
|
70
|
+
}
|
|
71
|
+
async completion(prompt, options, images, maxImageSize) {
|
|
72
|
+
return this.requestQueue.addCompletionRequest(prompt, options, images, maxImageSize);
|
|
73
|
+
}
|
|
74
|
+
async generate(prompt, options, images, maxImageSize, cachePath, cacheTrimTokens, primer) {
|
|
75
|
+
return this.requestQueue.addGenerateRequest(prompt, options, images, maxImageSize, cachePath, cacheTrimTokens, primer);
|
|
76
|
+
}
|
|
77
|
+
async exit() {
|
|
78
|
+
await this.processComm.exit();
|
|
79
|
+
}
|
|
80
|
+
cancelActiveRequest() {
|
|
81
|
+
this.requestQueue.cancelActiveRequest();
|
|
82
|
+
}
|
|
83
|
+
getStatus() {
|
|
84
|
+
return {
|
|
85
|
+
modelName: this.modelName,
|
|
86
|
+
queueLength: this.requestQueue.length,
|
|
87
|
+
isStreamingActive: this.processComm.isStreamingActive(),
|
|
88
|
+
isJsonBuffering: this.processComm.isJsonBuffering(),
|
|
89
|
+
};
|
|
90
|
+
}
|
|
91
|
+
}
|
|
92
|
+
//# sourceMappingURL=process-client.js.map
|
|
@@ -0,0 +1 @@
|
|
|
1
|
+
{"version":3,"file":"process-client.js","sourceRoot":"","sources":["../../src/local-inference/process-client.ts"],"names":[],"mappings":"AAAA;;;;GAIG;AAGH,OAAO,EAAE,MAAM,EAAE,MAAM,uBAAuB,CAAC;AAE/C,OAAO,EAAE,kBAAkB,EAAE,MAAM,qBAAqB,CAAC;AAWzD,OAAO,EACL,oBAAoB,GAErB,MAAM,4BAA4B,CAAC;AACpC,OAAO,EACL,qBAAqB,GAGtB,MAAM,oBAAoB,CAAC;AAiB5B,MAAM,OAAO,sBAAsB;IACxB,SAAS,CAAS;IAEnB,YAAY,CAAwB;IACpC,WAAW,CAAuB;IAE1C,YAAY,MAAoC;QAC9C,IAAI,CAAC,SAAS,GAAG,MAAM,CAAC,SAAS,CAAC;QAElC,IAAI,MAAM,CAAC,cAAc,EAAE,CAAC;YAC1B,kBAAkB,CAAC,MAAM,CAAC,cAAc,CAAC,CAAC;QAC5C,CAAC;QAED,MAAM,MAAM,GAAG,IAAI,MAAM,CAAC;YACxB,MAAM,EAAE,MAAM,CAAC,YAAY,IAAI,KAAK;YACpC,OAAO,EAAE,SAAS;SACnB,CAAC,CAAC;QAEH,kDAAkD;QAClD,kDAAkD;QAClD,MAAM,gBAAgB,GAAkC;YACtD,cAAc,EAAE,CAAC,QAAQ,EAAE,EAAE,CAAC,IAAI,CAAC,YAAY,CAAC,kBAAkB,CAAC,QAAQ,CAAC;YAC5E,kBAAkB,EAAE,GAAG,EAAE,CAAC,IAAI,CAAC,YAAY,CAAC,kBAAkB,EAAE;YAChE,aAAa,EAAE,CAAC,IAAI,EAAE,MAAM,EAAE,EAAE;gBAC9B,IAAI,IAAI,KAAK,CAAC,EAAE,CAAC;oBACf,MAAM,OAAO,GACX,MAAM,CAAC,uBAAuB,EAAE,CAAC,IAAI,EAAE,MAAM,CAAC;wBAC9C,+CAA+C,IAAI,YAAY,MAAM,GAAG,CAAC;oBAC3E,MAAM,CAAC,KAAK,CAAC,OAAO,CAAC,CAAC;oBACtB,IAAI,CAAC,YAAY,CAAC,SAAS,CAAC,IAAI,KAAK,CAAC,OAAO,CAAC,CAAC,CAAC;gBAClD,CAAC;YACH,CAAC;SACF,CAAC;QAEF,MAAM,cAAc,GAA0B;YAC5C,aAAa,EAAE,CAAC,IAAI,EAAE,EAAE,CAAC,IAAI,CAAC,WAAW,CAAC,aAAa,CAAC,IAAI,CAAC;YAC7D,eAAe,EAAE,GAAG,EAAE,CAAC,IAAI,CAAC,WAAW,CAAC,eAAe,EAAE;YACzD,kBAAkB,EAAE,GAAG,EAAE,CAAC,IAAI,CAAC,WAAW,CAAC,kBAAkB,EAAE;SAChE,CAAC;QAEF,IAAI,CAAC,WAAW,GAAG,IAAI,oBAAoB,CACzC;YACE,gBAAgB,EAAE,MAAM,CAAC,gBAAgB;YACzC,QAAQ,EAAE,MAAM,CAAC,QAAQ;YACzB,SAAS,EAAE,MAAM,CAAC,SAAS;YAC3B,SAAS,EAAE,MAAM,CAAC,cAAc;YAChC,QAAQ,EAAE,MAAM,CAAC,QAAQ;YACzB,YAAY,EAAE,MAAM,CAAC,YAAY;YACjC,uBAAuB,EAAE,MAAM,CAAC,uBAAuB;SACxD,EACD,gBAAgB,CACjB,CAAC;QAEF,IAAI,CAAC,YAAY,GAAG,IAAI,qBAAqB,CAAC,cAAc,EAAE,MAAM,CAAC,kBAAkB,CAAC,CAAC;IAC3F,CAAC;IAED,KAAK,CAAC,iBAAiB;QACrB,WAAW;IACb,CAAC;IAED,KAAK,CAAC,eAAe;QACnB,OAAO,IAAI,CAAC,YAAY,CAAC,sBAAsB,EAAE,CAAC;IACpD,CAAC;IAED,KAAK,CAAC,UAAU,CACd,QAA4B,EAC5B,OAA6B;QAE7B,OAAO,IAAI,CAAC,YAAY,CAAC,oBAAoB,CAAC,QAAQ,EAAE,OAAO,CAAC,CAAC;IACnE,CAAC;IAED,KAAK,CAAC,MAAM,CACV,QAA4B,EAC5B,OAAwD,EACxD,KAAiC,EACjC,eAA2C;QAE3C,OAAO,IAAI,CAAC,YAAY,CAAC,gBAAgB,CAAC,QAAQ,EAAE,OAAO,EAAE,KAAK,EAAE,eAAe,CAAC,CAAC;IACvF,CAAC;IAED,KAAK,CAAC,QAAQ,CACZ,QAA4B,EAC5B,KAAiC,EACjC,eAA2C;QAE3C,OAAO,IAAI,CAAC,YAAY,CAAC,kBAAkB,CAAC,QAAQ,EAAE,KAAK,EAAE,eAAe,CAAC,CAAC;IAChF,CAAC;IAED,KAAK,CAAC,YAAY,CAChB,SAAiB,EACjB,QAA4B,EAC5B,aAAsB,EACtB,YAAqB,EACrB,aAAwB,EACxB,YAAuB,EACvB,KAAiC,EACjC,eAA2C;QAE3C,OAAO,IAAI,CAAC,YAAY,CAAC,sBAAsB,CAC7C,SAAS,EACT,QAAQ,EACR,aAAa,EACb,YAAY,EACZ,aAAa,EACb,YAAY,EACZ,KAAK,EACL,eAAe,CAChB,CAAC;IACJ,CAAC;IAED,KAAK,CAAC,UAAU,CACd,MAAc,EACd,OAAiB,EACjB,MAAiB,EACjB,YAAqB;QAErB,OAAO,IAAI,CAAC,YAAY,CAAC,oBAAoB,CAAC,MAAM,EAAE,OAAO,EAAE,MAAM,EAAE,YAAY,CAAC,CAAC;IACvF,CAAC;IAED,KAAK,CAAC,QAAQ,CACZ,MAAyB,EACzB,OAAiB,EACjB,MAAiB,EACjB,YAAqB,EACrB,SAAkB,EAClB,eAAwB,EACxB,MAAe;QAEf,OAAO,IAAI,CAAC,YAAY,CAAC,kBAAkB,CACzC,MAAM,EACN,OAAO,EACP,MAAM,EACN,YAAY,EACZ,SAAS,EACT,eAAe,EACf,MAAM,CACP,CAAC;IACJ,CAAC;IAED,KAAK,CAAC,IAAI;QACR,MAAM,IAAI,CAAC,WAAW,CAAC,IAAI,EAAE,CAAC;IAChC,CAAC;IAED,mBAAmB;QACjB,IAAI,CAAC,YAAY,CAAC,mBAAmB,EAAE,CAAC;IAC1C,CAAC;IAED,SAAS;QACP,OAAO;YACL,SAAS,EAAE,IAAI,CAAC,SAAS;YACzB,WAAW,EAAE,IAAI,CAAC,YAAY,CAAC,MAAM;YACrC,iBAAiB,EAAE,IAAI,CAAC,WAAW,CAAC,iBAAiB,EAAE;YACvD,eAAe,EAAE,IAAI,CAAC,WAAW,CAAC,eAAe,EAAE;SACpD,CAAC;IACJ,CAAC;CACF"}
|
|
@@ -0,0 +1,41 @@
|
|
|
1
|
+
/**
|
|
2
|
+
* Thin Python 推論プロセスとの stdio 通信
|
|
3
|
+
*/
|
|
4
|
+
import { Readable } from 'stream';
|
|
5
|
+
export interface ProcessCommunicationCallbacks {
|
|
6
|
+
onJsonResponse: (jsonData: string) => void;
|
|
7
|
+
onRequestCompleted: () => void;
|
|
8
|
+
onProcessExit: (code: number | null, signal: string | null) => void;
|
|
9
|
+
}
|
|
10
|
+
export interface ProcessCommunicationConfig {
|
|
11
|
+
/** uv --project に渡す Python プロジェクトディレクトリ */
|
|
12
|
+
pythonProjectDir: string;
|
|
13
|
+
/** UV_PROJECT_ENVIRONMENT に設定する venv パス */
|
|
14
|
+
venvPath: string;
|
|
15
|
+
modelName: string;
|
|
16
|
+
/** `__main__.py` と modelName の後に付与する追加引数 */
|
|
17
|
+
extraArgs?: string[];
|
|
18
|
+
/** 子プロセス spawn 時にマージする追加環境変数 */
|
|
19
|
+
extraEnv?: Record<string, string>;
|
|
20
|
+
loggerPrefix?: string;
|
|
21
|
+
loggerContext?: string;
|
|
22
|
+
processExitErrorMessage?: (code: number | null, signal: string | null) => string;
|
|
23
|
+
}
|
|
24
|
+
export declare class ProcessCommunication {
|
|
25
|
+
private process;
|
|
26
|
+
private decoder;
|
|
27
|
+
private currentStream;
|
|
28
|
+
private jsonBuffer;
|
|
29
|
+
private draining;
|
|
30
|
+
private callbacks;
|
|
31
|
+
private readonly exitErrorMessage;
|
|
32
|
+
constructor(config: ProcessCommunicationConfig, callbacks: ProcessCommunicationCallbacks);
|
|
33
|
+
private processData;
|
|
34
|
+
createNewStream(): Readable;
|
|
35
|
+
cancelActiveStream(): void;
|
|
36
|
+
sendToProcess(data: string): void;
|
|
37
|
+
isStreamingActive(): boolean;
|
|
38
|
+
isJsonBuffering(): boolean;
|
|
39
|
+
exit(): Promise<void>;
|
|
40
|
+
}
|
|
41
|
+
//# sourceMappingURL=process-communication.d.ts.map
|
|
@@ -0,0 +1 @@
|
|
|
1
|
+
{"version":3,"file":"process-communication.d.ts","sourceRoot":"","sources":["../../src/local-inference/process-communication.ts"],"names":[],"mappings":"AAAA;;GAEG;AAGH,OAAO,EAAE,QAAQ,EAAE,MAAM,QAAQ,CAAC;AAIlC,MAAM,WAAW,6BAA6B;IAC5C,cAAc,EAAE,CAAC,QAAQ,EAAE,MAAM,KAAK,IAAI,CAAC;IAC3C,kBAAkB,EAAE,MAAM,IAAI,CAAC;IAC/B,aAAa,EAAE,CAAC,IAAI,EAAE,MAAM,GAAG,IAAI,EAAE,MAAM,EAAE,MAAM,GAAG,IAAI,KAAK,IAAI,CAAC;CACrE;AAED,MAAM,WAAW,0BAA0B;IACzC,2CAA2C;IAC3C,gBAAgB,EAAE,MAAM,CAAC;IACzB,2CAA2C;IAC3C,QAAQ,EAAE,MAAM,CAAC;IACjB,SAAS,EAAE,MAAM,CAAC;IAClB,4CAA4C;IAC5C,SAAS,CAAC,EAAE,MAAM,EAAE,CAAC;IACrB,gCAAgC;IAChC,QAAQ,CAAC,EAAE,MAAM,CAAC,MAAM,EAAE,MAAM,CAAC,CAAC;IAClC,YAAY,CAAC,EAAE,MAAM,CAAC;IACtB,aAAa,CAAC,EAAE,MAAM,CAAC;IACvB,uBAAuB,CAAC,EAAE,CAAC,IAAI,EAAE,MAAM,GAAG,IAAI,EAAE,MAAM,EAAE,MAAM,GAAG,IAAI,KAAK,MAAM,CAAC;CAClF;AAED,qBAAa,oBAAoB;IAC/B,OAAO,CAAC,OAAO,CAAiC;IAChD,OAAO,CAAC,OAAO,CAAgB;IAC/B,OAAO,CAAC,aAAa,CAAyB;IAC9C,OAAO,CAAC,UAAU,CAAc;IAChC,OAAO,CAAC,QAAQ,CAAS;IACzB,OAAO,CAAC,SAAS,CAAgC;IACjD,OAAO,CAAC,QAAQ,CAAC,gBAAgB,CAAyD;gBAE9E,MAAM,EAAE,0BAA0B,EAAE,SAAS,EAAE,6BAA6B;IAoDxF,OAAO,CAAC,WAAW;IAwCnB,eAAe,IAAI,QAAQ;IAO3B,kBAAkB,IAAI,IAAI;IAc1B,aAAa,CAAC,IAAI,EAAE,MAAM,GAAG,IAAI;IAIjC,iBAAiB,IAAI,OAAO;IAI5B,eAAe,IAAI,OAAO;IAIpB,IAAI,IAAI,OAAO,CAAC,IAAI,CAAC;CAe5B"}
|
|
@@ -1,52 +1,45 @@
|
|
|
1
1
|
/**
|
|
2
|
-
*
|
|
3
|
-
*
|
|
4
|
-
* Pythonプロセスとの通信、データの送受信、ストリーミング処理を管理
|
|
2
|
+
* Thin Python 推論プロセスとの stdio 通信
|
|
5
3
|
*/
|
|
6
4
|
import { spawn } from 'child_process';
|
|
7
5
|
import { Readable } from 'stream';
|
|
8
6
|
import { StringDecoder } from 'string_decoder';
|
|
9
|
-
import path from "path";
|
|
10
|
-
import { fileURLToPath } from "url";
|
|
11
7
|
import { Logger } from '@modular-prompt/utils';
|
|
12
|
-
const logger = new Logger({ prefix: 'MLX', context: 'process' });
|
|
13
|
-
// Get the mlx-ml/python directory
|
|
14
|
-
// From dist/mlx-ml/process/ -> go up 3 levels to package root, then to src/mlx-ml/python
|
|
15
|
-
const packageRoot = path.resolve(path.dirname(fileURLToPath(import.meta.url)), '..', '..', '..' // dist/mlx-ml/process -> dist/mlx-ml -> dist -> package root
|
|
16
|
-
);
|
|
17
|
-
const mlxDriverDir = path.join(packageRoot, 'src', 'mlx-ml', 'python');
|
|
18
8
|
export class ProcessCommunication {
|
|
19
9
|
process;
|
|
20
10
|
decoder;
|
|
21
11
|
currentStream = null;
|
|
22
12
|
jsonBuffer = '';
|
|
13
|
+
draining = false;
|
|
23
14
|
callbacks;
|
|
24
|
-
|
|
15
|
+
exitErrorMessage;
|
|
16
|
+
constructor(config, callbacks) {
|
|
25
17
|
this.callbacks = callbacks;
|
|
26
18
|
this.decoder = new StringDecoder('utf8');
|
|
19
|
+
this.exitErrorMessage =
|
|
20
|
+
config.processExitErrorMessage ??
|
|
21
|
+
((code, signal) => `Inference process exited unexpectedly (code=${code}, signal=${signal})`);
|
|
22
|
+
const logger = new Logger({
|
|
23
|
+
prefix: config.loggerPrefix ?? 'LIP',
|
|
24
|
+
context: config.loggerContext ?? 'process',
|
|
25
|
+
});
|
|
27
26
|
const args = [
|
|
28
27
|
'--project',
|
|
29
|
-
|
|
28
|
+
config.pythonProjectDir,
|
|
30
29
|
'run',
|
|
31
30
|
'python',
|
|
32
31
|
'__main__.py',
|
|
33
|
-
modelName
|
|
32
|
+
config.modelName,
|
|
33
|
+
...(config.extraArgs ?? []),
|
|
34
34
|
];
|
|
35
|
-
if (options?.textOnly) {
|
|
36
|
-
args.push('--text-only');
|
|
37
|
-
}
|
|
38
|
-
if (options?.drafterModel) {
|
|
39
|
-
args.push('--drafter', options.drafterModel);
|
|
40
|
-
}
|
|
41
|
-
if (options?.draftBlockSize !== undefined) {
|
|
42
|
-
args.push('--draft-block-size', options.draftBlockSize.toString());
|
|
43
|
-
}
|
|
44
35
|
this.process = spawn('uv', args, {
|
|
45
|
-
cwd:
|
|
36
|
+
cwd: config.pythonProjectDir,
|
|
37
|
+
env: {
|
|
38
|
+
...process.env,
|
|
39
|
+
UV_PROJECT_ENVIRONMENT: config.venvPath,
|
|
40
|
+
...config.extraEnv,
|
|
41
|
+
},
|
|
46
42
|
});
|
|
47
|
-
this.setupProcessHandlers();
|
|
48
|
-
}
|
|
49
|
-
setupProcessHandlers() {
|
|
50
43
|
this.process.stderr.on('data', (data) => {
|
|
51
44
|
logger.debug(data.toString());
|
|
52
45
|
});
|
|
@@ -58,50 +51,41 @@ export class ProcessCommunication {
|
|
|
58
51
|
});
|
|
59
52
|
this.process.on('exit', (code, signal) => {
|
|
60
53
|
if (this.currentStream) {
|
|
61
|
-
this.currentStream.destroy(new Error(
|
|
54
|
+
this.currentStream.destroy(new Error(this.exitErrorMessage(code, signal)));
|
|
62
55
|
this.currentStream = null;
|
|
63
56
|
}
|
|
64
57
|
this.callbacks.onProcessExit(code, signal);
|
|
65
58
|
});
|
|
66
59
|
}
|
|
67
|
-
/**
|
|
68
|
-
* stdoutから受信したデータを処理する
|
|
69
|
-
* null文字をレスポンス区切りとして解釈し、
|
|
70
|
-
* 1つのdataチャンクに複数のレスポンスが含まれる場合も正しく処理する
|
|
71
|
-
*/
|
|
72
60
|
processData(data) {
|
|
73
61
|
let remaining = data;
|
|
74
62
|
while (remaining.length > 0) {
|
|
75
63
|
const nullIndex = remaining.indexOf('\0');
|
|
76
64
|
if (nullIndex !== -1) {
|
|
77
|
-
// null文字が見つかった場合、レスポンス終了
|
|
78
65
|
const chunk = this.decoder.write(remaining.slice(0, nullIndex));
|
|
79
66
|
this.decoder = new StringDecoder('utf8');
|
|
80
67
|
if (this.currentStream) {
|
|
81
|
-
// ストリーミングレスポンスの場合
|
|
82
68
|
this.currentStream.push(chunk);
|
|
83
|
-
this.currentStream.push(null);
|
|
69
|
+
this.currentStream.push(null);
|
|
84
70
|
this.currentStream = null;
|
|
85
71
|
}
|
|
86
|
-
else {
|
|
87
|
-
// JSONレスポンスの場合
|
|
72
|
+
else if (!this.draining) {
|
|
88
73
|
this.jsonBuffer += chunk;
|
|
89
74
|
this.callbacks.onJsonResponse(this.jsonBuffer);
|
|
90
75
|
this.jsonBuffer = '';
|
|
91
76
|
}
|
|
92
77
|
this.callbacks.onRequestCompleted();
|
|
93
|
-
|
|
78
|
+
if (this.draining) {
|
|
79
|
+
this.draining = false;
|
|
80
|
+
}
|
|
94
81
|
remaining = remaining.slice(nullIndex + 1);
|
|
95
82
|
}
|
|
96
83
|
else {
|
|
97
|
-
// null文字がない場合、データを蓄積
|
|
98
84
|
const chunk = this.decoder.write(remaining);
|
|
99
85
|
if (this.currentStream) {
|
|
100
|
-
// ストリーミング中
|
|
101
86
|
this.currentStream.push(chunk);
|
|
102
87
|
}
|
|
103
|
-
else {
|
|
104
|
-
// JSONレスポンス蓄積中
|
|
88
|
+
else if (!this.draining) {
|
|
105
89
|
this.jsonBuffer += chunk;
|
|
106
90
|
}
|
|
107
91
|
break;
|
|
@@ -110,10 +94,21 @@ export class ProcessCommunication {
|
|
|
110
94
|
}
|
|
111
95
|
createNewStream() {
|
|
112
96
|
this.currentStream = new Readable({
|
|
113
|
-
read() { }
|
|
97
|
+
read() { },
|
|
114
98
|
});
|
|
115
99
|
return this.currentStream;
|
|
116
100
|
}
|
|
101
|
+
cancelActiveStream() {
|
|
102
|
+
if (!this.currentStream && !this.draining) {
|
|
103
|
+
return;
|
|
104
|
+
}
|
|
105
|
+
if (this.currentStream) {
|
|
106
|
+
this.currentStream.destroy();
|
|
107
|
+
this.currentStream = null;
|
|
108
|
+
}
|
|
109
|
+
this.draining = true;
|
|
110
|
+
this.sendToProcess(JSON.stringify({ method: 'cancel' }) + '\n');
|
|
111
|
+
}
|
|
117
112
|
sendToProcess(data) {
|
|
118
113
|
this.process.stdin.write(data);
|
|
119
114
|
}
|
|
@@ -126,15 +121,13 @@ export class ProcessCommunication {
|
|
|
126
121
|
async exit() {
|
|
127
122
|
return new Promise((resolve) => {
|
|
128
123
|
const timeout = setTimeout(() => {
|
|
129
|
-
// タイムアウト後は強制終了
|
|
130
124
|
this.process.kill('SIGTERM');
|
|
131
125
|
resolve();
|
|
132
|
-
}, 5000);
|
|
126
|
+
}, 5000);
|
|
133
127
|
this.process.once('exit', () => {
|
|
134
128
|
clearTimeout(timeout);
|
|
135
129
|
resolve();
|
|
136
130
|
});
|
|
137
|
-
// stdinを閉じてプロセスに終了を通知
|
|
138
131
|
this.process.stdin.end();
|
|
139
132
|
});
|
|
140
133
|
}
|
|
@@ -0,0 +1 @@
|
|
|
1
|
+
{"version":3,"file":"process-communication.js","sourceRoot":"","sources":["../../src/local-inference/process-communication.ts"],"names":[],"mappings":"AAAA;;GAEG;AAEH,OAAO,EAAkC,KAAK,EAAE,MAAM,eAAe,CAAC;AACtE,OAAO,EAAE,QAAQ,EAAE,MAAM,QAAQ,CAAC;AAClC,OAAO,EAAE,aAAa,EAAE,MAAM,gBAAgB,CAAC;AAC/C,OAAO,EAAE,MAAM,EAAE,MAAM,uBAAuB,CAAC;AAuB/C,MAAM,OAAO,oBAAoB;IACvB,OAAO,CAAiC;IACxC,OAAO,CAAgB;IACvB,aAAa,GAAoB,IAAI,CAAC;IACtC,UAAU,GAAW,EAAE,CAAC;IACxB,QAAQ,GAAG,KAAK,CAAC;IACjB,SAAS,CAAgC;IAChC,gBAAgB,CAAyD;IAE1F,YAAY,MAAkC,EAAE,SAAwC;QACtF,IAAI,CAAC,SAAS,GAAG,SAAS,CAAC;QAC3B,IAAI,CAAC,OAAO,GAAG,IAAI,aAAa,CAAC,MAAM,CAAC,CAAC;QACzC,IAAI,CAAC,gBAAgB;YACnB,MAAM,CAAC,uBAAuB;gBAC9B,CAAC,CAAC,IAAI,EAAE,MAAM,EAAE,EAAE,CAAC,+CAA+C,IAAI,YAAY,MAAM,GAAG,CAAC,CAAC;QAE/F,MAAM,MAAM,GAAG,IAAI,MAAM,CAAC;YACxB,MAAM,EAAE,MAAM,CAAC,YAAY,IAAI,KAAK;YACpC,OAAO,EAAE,MAAM,CAAC,aAAa,IAAI,SAAS;SAC3C,CAAC,CAAC;QAEH,MAAM,IAAI,GAAG;YACX,WAAW;YACX,MAAM,CAAC,gBAAgB;YACvB,KAAK;YACL,QAAQ;YACR,aAAa;YACb,MAAM,CAAC,SAAS;YAChB,GAAG,CAAC,MAAM,CAAC,SAAS,IAAI,EAAE,CAAC;SAC5B,CAAC;QAEF,IAAI,CAAC,OAAO,GAAG,KAAK,CAAC,IAAI,EAAE,IAAI,EAAE;YAC/B,GAAG,EAAE,MAAM,CAAC,gBAAgB;YAC5B,GAAG,EAAE;gBACH,GAAG,OAAO,CAAC,GAAG;gBACd,sBAAsB,EAAE,MAAM,CAAC,QAAQ;gBACvC,GAAG,MAAM,CAAC,QAAQ;aACnB;SACF,CAAC,CAAC;QAEH,IAAI,CAAC,OAAO,CAAC,MAAM,CAAC,EAAE,CAAC,MAAM,EAAE,CAAC,IAAI,EAAE,EAAE;YACtC,MAAM,CAAC,KAAK,CAAC,IAAI,CAAC,QAAQ,EAAE,CAAC,CAAC;QAChC,CAAC,CAAC,CAAC;QAEH,IAAI,CAAC,OAAO,CAAC,MAAM,CAAC,EAAE,CAAC,MAAM,EAAE,CAAC,IAAI,EAAE,EAAE;YACtC,IAAI,CAAC,WAAW,CAAC,IAAI,CAAC,CAAC;QACzB,CAAC,CAAC,CAAC;QAEH,IAAI,CAAC,OAAO,CAAC,EAAE,CAAC,OAAO,EAAE,CAAC,GAAG,EAAE,EAAE;YAC/B,MAAM,CAAC,KAAK,CAAC,sBAAsB,EAAE,GAAG,CAAC,CAAC;QAC5C,CAAC,CAAC,CAAC;QAEH,IAAI,CAAC,OAAO,CAAC,EAAE,CAAC,MAAM,EAAE,CAAC,IAAI,EAAE,MAAM,EAAE,EAAE;YACvC,IAAI,IAAI,CAAC,aAAa,EAAE,CAAC;gBACvB,IAAI,CAAC,aAAa,CAAC,OAAO,CAAC,IAAI,KAAK,CAAC,IAAI,CAAC,gBAAgB,CAAC,IAAI,EAAE,MAAM,CAAC,CAAC,CAAC,CAAC;gBAC3E,IAAI,CAAC,aAAa,GAAG,IAAI,CAAC;YAC5B,CAAC;YACD,IAAI,CAAC,SAAS,CAAC,aAAa,CAAC,IAAI,EAAE,MAAM,CAAC,CAAC;QAC7C,CAAC,CAAC,CAAC;IACL,CAAC;IAEO,WAAW,CAAC,IAAY;QAC9B,IAAI,SAAS,GAAW,IAAI,CAAC;QAE7B,OAAO,SAAS,CAAC,MAAM,GAAG,CAAC,EAAE,CAAC;YAC5B,MAAM,SAAS,GAAG,SAAS,CAAC,OAAO,CAAC,IAAI,CAAC,CAAC;YAE1C,IAAI,SAAS,KAAK,CAAC,CAAC,EAAE,CAAC;gBACrB,MAAM,KAAK,GAAG,IAAI,CAAC,OAAO,CAAC,KAAK,CAAC,SAAS,CAAC,KAAK,CAAC,CAAC,EAAE,SAAS,CAAC,CAAC,CAAC;gBAChE,IAAI,CAAC,OAAO,GAAG,IAAI,aAAa,CAAC,MAAM,CAAC,CAAC;gBAEzC,IAAI,IAAI,CAAC,aAAa,EAAE,CAAC;oBACvB,IAAI,CAAC,aAAa,CAAC,IAAI,CAAC,KAAK,CAAC,CAAC;oBAC/B,IAAI,CAAC,aAAa,CAAC,IAAI,CAAC,IAAI,CAAC,CAAC;oBAC9B,IAAI,CAAC,aAAa,GAAG,IAAI,CAAC;gBAC5B,CAAC;qBAAM,IAAI,CAAC,IAAI,CAAC,QAAQ,EAAE,CAAC;oBAC1B,IAAI,CAAC,UAAU,IAAI,KAAK,CAAC;oBACzB,IAAI,CAAC,SAAS,CAAC,cAAc,CAAC,IAAI,CAAC,UAAU,CAAC,CAAC;oBAC/C,IAAI,CAAC,UAAU,GAAG,EAAE,CAAC;gBACvB,CAAC;gBAED,IAAI,CAAC,SAAS,CAAC,kBAAkB,EAAE,CAAC;gBAEpC,IAAI,IAAI,CAAC,QAAQ,EAAE,CAAC;oBAClB,IAAI,CAAC,QAAQ,GAAG,KAAK,CAAC;gBACxB,CAAC;gBAED,SAAS,GAAG,SAAS,CAAC,KAAK,CAAC,SAAS,GAAG,CAAC,CAAC,CAAC;YAC7C,CAAC;iBAAM,CAAC;gBACN,MAAM,KAAK,GAAG,IAAI,CAAC,OAAO,CAAC,KAAK,CAAC,SAAS,CAAC,CAAC;gBAE5C,IAAI,IAAI,CAAC,aAAa,EAAE,CAAC;oBACvB,IAAI,CAAC,aAAa,CAAC,IAAI,CAAC,KAAK,CAAC,CAAC;gBACjC,CAAC;qBAAM,IAAI,CAAC,IAAI,CAAC,QAAQ,EAAE,CAAC;oBAC1B,IAAI,CAAC,UAAU,IAAI,KAAK,CAAC;gBAC3B,CAAC;gBACD,MAAM;YACR,CAAC;QACH,CAAC;IACH,CAAC;IAED,eAAe;QACb,IAAI,CAAC,aAAa,GAAG,IAAI,QAAQ,CAAC;YAChC,IAAI,KAAI,CAAC;SACV,CAAC,CAAC;QACH,OAAO,IAAI,CAAC,aAAa,CAAC;IAC5B,CAAC;IAED,kBAAkB;QAChB,IAAI,CAAC,IAAI,CAAC,aAAa,IAAI,CAAC,IAAI,CAAC,QAAQ,EAAE,CAAC;YAC1C,OAAO;QACT,CAAC;QAED,IAAI,IAAI,CAAC,aAAa,EAAE,CAAC;YACvB,IAAI,CAAC,aAAa,CAAC,OAAO,EAAE,CAAC;YAC7B,IAAI,CAAC,aAAa,GAAG,IAAI,CAAC;QAC5B,CAAC;QAED,IAAI,CAAC,QAAQ,GAAG,IAAI,CAAC;QACrB,IAAI,CAAC,aAAa,CAAC,IAAI,CAAC,SAAS,CAAC,EAAE,MAAM,EAAE,QAAQ,EAAE,CAAC,GAAG,IAAI,CAAC,CAAC;IAClE,CAAC;IAED,aAAa,CAAC,IAAY;QACxB,IAAI,CAAC,OAAO,CAAC,KAAK,CAAC,KAAK,CAAC,IAAI,CAAC,CAAC;IACjC,CAAC;IAED,iBAAiB;QACf,OAAO,IAAI,CAAC,aAAa,KAAK,IAAI,CAAC;IACrC,CAAC;IAED,eAAe;QACb,OAAO,IAAI,CAAC,UAAU,CAAC,MAAM,GAAG,CAAC,CAAC;IACpC,CAAC;IAED,KAAK,CAAC,IAAI;QACR,OAAO,IAAI,OAAO,CAAC,CAAC,OAAO,EAAE,EAAE;YAC7B,MAAM,OAAO,GAAG,UAAU,CAAC,GAAG,EAAE;gBAC9B,IAAI,CAAC,OAAO,CAAC,IAAI,CAAC,SAAS,CAAC,CAAC;gBAC7B,OAAO,EAAE,CAAC;YACZ,CAAC,EAAE,IAAI,CAAC,CAAC;YAET,IAAI,CAAC,OAAO,CAAC,IAAI,CAAC,MAAM,EAAE,GAAG,EAAE;gBAC7B,YAAY,CAAC,OAAO,CAAC,CAAC;gBACtB,OAAO,EAAE,CAAC;YACZ,CAAC,CAAC,CAAC;YAEH,IAAI,CAAC,OAAO,CAAC,KAAK,CAAC,GAAG,EAAE,CAAC;QAC3B,CAAC,CAAC,CAAC;IACL,CAAC;CACF"}
|
|
@@ -0,0 +1,12 @@
|
|
|
1
|
+
import type { Readable } from 'stream';
|
|
2
|
+
import type { InferenceCapabilities, InferenceMessage, InferenceRenderResult, InferenceToolDefinition } from './protocol.js';
|
|
3
|
+
/** LIP プロセスへの最小ポート(MlxProcess / 将来 PyTorchProcess が実装) */
|
|
4
|
+
export interface InferenceProcessPort {
|
|
5
|
+
ensureInitialized(): Promise<void>;
|
|
6
|
+
getCapabilities(): Promise<InferenceCapabilities>;
|
|
7
|
+
render(messages: InferenceMessage[], options?: Record<string, unknown>, tools?: InferenceToolDefinition[], reasoningEffort?: 'low' | 'medium' | 'high'): Promise<InferenceRenderResult>;
|
|
8
|
+
generate(prompt: string | number[], options?: Record<string, unknown>, images?: string[], maxImageSize?: number, cachePath?: string, cacheTrimTokens?: number, primer?: string): Promise<Readable>;
|
|
9
|
+
cancelActiveRequest(): void;
|
|
10
|
+
exit(): Promise<void>;
|
|
11
|
+
}
|
|
12
|
+
//# sourceMappingURL=process-port.d.ts.map
|
|
@@ -0,0 +1 @@
|
|
|
1
|
+
{"version":3,"file":"process-port.d.ts","sourceRoot":"","sources":["../../src/local-inference/process-port.ts"],"names":[],"mappings":"AAAA,OAAO,KAAK,EAAE,QAAQ,EAAE,MAAM,QAAQ,CAAC;AACvC,OAAO,KAAK,EACV,qBAAqB,EACrB,gBAAgB,EAChB,qBAAqB,EACrB,uBAAuB,EACxB,MAAM,eAAe,CAAC;AAEvB,0DAA0D;AAC1D,MAAM,WAAW,oBAAoB;IACnC,iBAAiB,IAAI,OAAO,CAAC,IAAI,CAAC,CAAC;IACnC,eAAe,IAAI,OAAO,CAAC,qBAAqB,CAAC,CAAC;IAClD,MAAM,CACJ,QAAQ,EAAE,gBAAgB,EAAE,EAC5B,OAAO,CAAC,EAAE,MAAM,CAAC,MAAM,EAAE,OAAO,CAAC,EACjC,KAAK,CAAC,EAAE,uBAAuB,EAAE,EACjC,eAAe,CAAC,EAAE,KAAK,GAAG,QAAQ,GAAG,MAAM,GAC1C,OAAO,CAAC,qBAAqB,CAAC,CAAC;IAClC,QAAQ,CACN,MAAM,EAAE,MAAM,GAAG,MAAM,EAAE,EACzB,OAAO,CAAC,EAAE,MAAM,CAAC,MAAM,EAAE,OAAO,CAAC,EACjC,MAAM,CAAC,EAAE,MAAM,EAAE,EACjB,YAAY,CAAC,EAAE,MAAM,EACrB,SAAS,CAAC,EAAE,MAAM,EAClB,eAAe,CAAC,EAAE,MAAM,EACxB,MAAM,CAAC,EAAE,MAAM,GACd,OAAO,CAAC,QAAQ,CAAC,CAAC;IACrB,mBAAmB,IAAI,IAAI,CAAC;IAC5B,IAAI,IAAI,OAAO,CAAC,IAAI,CAAC,CAAC;CACvB"}
|
|
@@ -0,0 +1 @@
|
|
|
1
|
+
{"version":3,"file":"process-port.js","sourceRoot":"","sources":["../../src/local-inference/process-port.ts"],"names":[],"mappings":""}
|
|
@@ -0,0 +1 @@
|
|
|
1
|
+
{"version":3,"file":"prompt-utils.d.ts","sourceRoot":"","sources":["../../src/local-inference/prompt-utils.ts"],"names":[],"mappings":"AAAA,OAAO,KAAK,EAAE,cAAc,EAAE,MAAM,sBAAsB,CAAC;AAE3D;;GAEG;AACH,wBAAgB,iBAAiB,CAAC,MAAM,EAAE,cAAc,GAAG,OAAO,CAcjE"}
|
|
@@ -0,0 +1,17 @@
|
|
|
1
|
+
/**
|
|
2
|
+
* Check if the prompt contains MessageElement
|
|
3
|
+
*/
|
|
4
|
+
export function hasMessageElement(prompt) {
|
|
5
|
+
const checkElements = (elements) => {
|
|
6
|
+
if (!elements)
|
|
7
|
+
return false;
|
|
8
|
+
return elements.some((element) => {
|
|
9
|
+
const el = element;
|
|
10
|
+
return el?.type === 'message';
|
|
11
|
+
});
|
|
12
|
+
};
|
|
13
|
+
return (checkElements(prompt.instructions) ||
|
|
14
|
+
checkElements(prompt.data) ||
|
|
15
|
+
checkElements(prompt.output));
|
|
16
|
+
}
|
|
17
|
+
//# sourceMappingURL=prompt-utils.js.map
|
|
@@ -0,0 +1 @@
|
|
|
1
|
+
{"version":3,"file":"prompt-utils.js","sourceRoot":"","sources":["../../src/local-inference/prompt-utils.ts"],"names":[],"mappings":"AAEA;;GAEG;AACH,MAAM,UAAU,iBAAiB,CAAC,MAAsB;IACtD,MAAM,aAAa,GAAG,CAAC,QAAmB,EAAW,EAAE;QACrD,IAAI,CAAC,QAAQ;YAAE,OAAO,KAAK,CAAC;QAC5B,OAAO,QAAQ,CAAC,IAAI,CAAC,CAAC,OAAO,EAAE,EAAE;YAC/B,MAAM,EAAE,GAAG,OAA4B,CAAC;YACxC,OAAO,EAAE,EAAE,IAAI,KAAK,SAAS,CAAC;QAChC,CAAC,CAAC,CAAC;IACL,CAAC,CAAC;IAEF,OAAO,CACL,aAAa,CAAC,MAAM,CAAC,YAAY,CAAC;QAClC,aAAa,CAAC,MAAM,CAAC,IAAI,CAAC;QAC1B,aAAa,CAAC,MAAM,CAAC,MAAM,CAAC,CAC7B,CAAC;AACJ,CAAC"}
|
|
@@ -0,0 +1,192 @@
|
|
|
1
|
+
/**
|
|
2
|
+
* Local Inference Protocol (LIP) 型定義
|
|
3
|
+
*
|
|
4
|
+
* Thin Python 推論型ドライバ(MLX / PyTorch 等)が共有する stdio JSON-RPC プロトコル。
|
|
5
|
+
* Python プロセスからの応答は snake_case のまま受け取る。
|
|
6
|
+
*/
|
|
7
|
+
import type { SpecialToken, SpecialTokenPair } from '../formatter/types.js';
|
|
8
|
+
/** LIP で利用可能な JSON-RPC メソッド */
|
|
9
|
+
export type InferenceMethod = 'capabilities' | 'format_test' | 'render' | 'chat' | 'completion' | 'generate' | 'cache_prefill' | 'tokenize';
|
|
10
|
+
/** VLM content part for structured message content */
|
|
11
|
+
export type InferenceContentPart = {
|
|
12
|
+
type: 'text';
|
|
13
|
+
text: string;
|
|
14
|
+
} | {
|
|
15
|
+
type: 'image';
|
|
16
|
+
};
|
|
17
|
+
/** 標準メッセージ(system / user / assistant) */
|
|
18
|
+
export interface InferenceStandardMessage {
|
|
19
|
+
role: 'user' | 'assistant' | 'system';
|
|
20
|
+
content: string | InferenceContentPart[];
|
|
21
|
+
}
|
|
22
|
+
/** tool_calls 付き assistant メッセージ(HuggingFace 互換形式) */
|
|
23
|
+
export interface InferenceAssistantToolCallMessage {
|
|
24
|
+
role: 'assistant';
|
|
25
|
+
content: string;
|
|
26
|
+
tool_calls: Array<{
|
|
27
|
+
id: string;
|
|
28
|
+
type: 'function';
|
|
29
|
+
function: {
|
|
30
|
+
name: string;
|
|
31
|
+
arguments: string;
|
|
32
|
+
};
|
|
33
|
+
}>;
|
|
34
|
+
}
|
|
35
|
+
/** tool result メッセージ(HuggingFace 互換形式) */
|
|
36
|
+
export interface InferenceToolResultMessage {
|
|
37
|
+
role: 'tool';
|
|
38
|
+
content: string;
|
|
39
|
+
tool_call_id: string;
|
|
40
|
+
name: string;
|
|
41
|
+
}
|
|
42
|
+
export type InferenceMessage = InferenceStandardMessage | InferenceAssistantToolCallMessage | InferenceToolResultMessage;
|
|
43
|
+
/** HuggingFace apply_chat_template 互換の tool 定義 */
|
|
44
|
+
export interface InferenceToolDefinition {
|
|
45
|
+
type: 'function';
|
|
46
|
+
function: {
|
|
47
|
+
name: string;
|
|
48
|
+
description?: string;
|
|
49
|
+
parameters?: Record<string, unknown>;
|
|
50
|
+
};
|
|
51
|
+
}
|
|
52
|
+
/** サンプリング等の推論オプション(バックエンド共通の最小集合) */
|
|
53
|
+
export interface InferenceSamplingOptions {
|
|
54
|
+
max_tokens?: number;
|
|
55
|
+
temperature?: number;
|
|
56
|
+
top_p?: number;
|
|
57
|
+
top_k?: number;
|
|
58
|
+
repetition_penalty?: number;
|
|
59
|
+
repetition_context_size?: number;
|
|
60
|
+
trust_remote_code?: boolean;
|
|
61
|
+
}
|
|
62
|
+
export interface InferenceBaseRequest {
|
|
63
|
+
method: InferenceMethod;
|
|
64
|
+
}
|
|
65
|
+
export interface InferenceCapabilitiesRequest extends InferenceBaseRequest {
|
|
66
|
+
method: 'capabilities';
|
|
67
|
+
}
|
|
68
|
+
export interface InferenceFormatTestRequest extends InferenceBaseRequest {
|
|
69
|
+
method: 'format_test';
|
|
70
|
+
messages: InferenceMessage[];
|
|
71
|
+
options?: {
|
|
72
|
+
primer?: string;
|
|
73
|
+
};
|
|
74
|
+
}
|
|
75
|
+
export interface InferenceTokenizeRequest extends InferenceBaseRequest {
|
|
76
|
+
method: 'tokenize';
|
|
77
|
+
messages: InferenceMessage[];
|
|
78
|
+
tools?: InferenceToolDefinition[];
|
|
79
|
+
reasoning_effort?: 'low' | 'medium' | 'high';
|
|
80
|
+
}
|
|
81
|
+
export interface InferenceRenderRequest extends InferenceBaseRequest {
|
|
82
|
+
method: 'render';
|
|
83
|
+
messages: InferenceMessage[];
|
|
84
|
+
options?: InferenceSamplingOptions & {
|
|
85
|
+
primer?: string;
|
|
86
|
+
};
|
|
87
|
+
tools?: InferenceToolDefinition[];
|
|
88
|
+
reasoning_effort?: 'low' | 'medium' | 'high';
|
|
89
|
+
}
|
|
90
|
+
/** @deprecated Python chat ハンドラ廃止。render + generate を使用すること。 */
|
|
91
|
+
export interface InferenceChatRequest extends InferenceBaseRequest {
|
|
92
|
+
method: 'chat';
|
|
93
|
+
messages: InferenceMessage[];
|
|
94
|
+
primer?: string;
|
|
95
|
+
tools?: InferenceToolDefinition[];
|
|
96
|
+
options?: InferenceSamplingOptions;
|
|
97
|
+
images?: string[];
|
|
98
|
+
maxImageSize?: number;
|
|
99
|
+
reasoning_effort?: 'low' | 'medium' | 'high';
|
|
100
|
+
cache_path?: string;
|
|
101
|
+
cache_trim_tokens?: number;
|
|
102
|
+
}
|
|
103
|
+
export interface InferenceRenderResult {
|
|
104
|
+
formatted_prompt: string | null;
|
|
105
|
+
error: string | null;
|
|
106
|
+
}
|
|
107
|
+
export interface InferenceCompletionRequest extends InferenceBaseRequest {
|
|
108
|
+
method: 'completion';
|
|
109
|
+
prompt: string;
|
|
110
|
+
options?: InferenceSamplingOptions;
|
|
111
|
+
images?: string[];
|
|
112
|
+
maxImageSize?: number;
|
|
113
|
+
}
|
|
114
|
+
/** 整形済み prompt(文字列 or トークン ID)のストリーム推論 */
|
|
115
|
+
export interface InferenceGenerateRequest extends InferenceBaseRequest {
|
|
116
|
+
method: 'generate';
|
|
117
|
+
prompt: string | number[];
|
|
118
|
+
primer?: string;
|
|
119
|
+
options?: InferenceSamplingOptions;
|
|
120
|
+
images?: string[];
|
|
121
|
+
maxImageSize?: number;
|
|
122
|
+
cache_path?: string;
|
|
123
|
+
cache_trim_tokens?: number;
|
|
124
|
+
}
|
|
125
|
+
export interface InferenceCachePrefillRequest extends InferenceBaseRequest {
|
|
126
|
+
method: 'cache_prefill';
|
|
127
|
+
cache_path: string;
|
|
128
|
+
messages: InferenceMessage[];
|
|
129
|
+
base_cache_path?: string;
|
|
130
|
+
trim_to_tokens?: number;
|
|
131
|
+
prefix_offsets?: number[];
|
|
132
|
+
prefix_hashes?: string[];
|
|
133
|
+
tools?: InferenceToolDefinition[];
|
|
134
|
+
reasoning_effort?: 'low' | 'medium' | 'high';
|
|
135
|
+
}
|
|
136
|
+
export interface InferenceCachePrefillResult {
|
|
137
|
+
cache_path: string;
|
|
138
|
+
token_count?: number;
|
|
139
|
+
prefix_offsets?: number[];
|
|
140
|
+
prefix_hashes?: string[];
|
|
141
|
+
}
|
|
142
|
+
export type InferenceRequest = InferenceCapabilitiesRequest | InferenceFormatTestRequest | InferenceRenderRequest | InferenceTokenizeRequest | InferenceChatRequest | InferenceCompletionRequest | InferenceGenerateRequest | InferenceCachePrefillRequest;
|
|
143
|
+
/** MLX-LM が認識する tool_parser_type(capabilities 経由で参照)。
|
|
144
|
+
* 現時点では MLX 固有の知識。PyTorch バックエンド追加時(#302 Phase 6)に
|
|
145
|
+
* バックエンド固有モジュールへ移動・一般化する可能性あり。 */
|
|
146
|
+
export type KnownToolParserType = 'json_tools' | 'pythonic' | 'function_gemma' | 'mistral' | 'kimi_k2' | 'longcat' | 'glm47' | 'qwen3_coder' | 'minimax_m2' | 'gemma4' | 'harmony';
|
|
147
|
+
export interface ToolCallFormat {
|
|
148
|
+
tool_parser_type?: string;
|
|
149
|
+
call_start?: string;
|
|
150
|
+
call_end?: string;
|
|
151
|
+
response_start?: string;
|
|
152
|
+
response_end?: string;
|
|
153
|
+
}
|
|
154
|
+
export interface ChatTemplateInfo {
|
|
155
|
+
supported_roles: string[];
|
|
156
|
+
preview?: string;
|
|
157
|
+
constraints: Record<string, unknown>;
|
|
158
|
+
tool_call_format?: ToolCallFormat;
|
|
159
|
+
}
|
|
160
|
+
/**
|
|
161
|
+
* Python プロセスから取得するランタイム能力情報(snake_case)
|
|
162
|
+
*/
|
|
163
|
+
export interface InferenceCapabilities {
|
|
164
|
+
methods: string[];
|
|
165
|
+
special_tokens: Record<string, SpecialToken | SpecialTokenPair>;
|
|
166
|
+
features: {
|
|
167
|
+
apply_chat_template: boolean;
|
|
168
|
+
vocab_size?: number;
|
|
169
|
+
model_max_length?: number;
|
|
170
|
+
chat_template?: ChatTemplateInfo;
|
|
171
|
+
};
|
|
172
|
+
chat_restrictions?: {
|
|
173
|
+
single_system_at_start?: boolean;
|
|
174
|
+
max_system_messages?: number;
|
|
175
|
+
alternating_turns?: boolean;
|
|
176
|
+
requires_user_last?: boolean;
|
|
177
|
+
allow_empty_messages?: boolean;
|
|
178
|
+
};
|
|
179
|
+
model_kind?: 'lm' | 'vlm';
|
|
180
|
+
}
|
|
181
|
+
export interface InferenceFormatTestResult {
|
|
182
|
+
formatted_prompt: string | null;
|
|
183
|
+
template_applied: boolean;
|
|
184
|
+
model_specific_processing: InferenceMessage[] | null;
|
|
185
|
+
error: string | null;
|
|
186
|
+
}
|
|
187
|
+
export interface InferenceTokenizeResult {
|
|
188
|
+
token_ids: number[] | null;
|
|
189
|
+
token_count: number;
|
|
190
|
+
error: string | null;
|
|
191
|
+
}
|
|
192
|
+
//# sourceMappingURL=protocol.d.ts.map
|
|
@@ -0,0 +1 @@
|
|
|
1
|
+
{"version":3,"file":"protocol.d.ts","sourceRoot":"","sources":["../../src/local-inference/protocol.ts"],"names":[],"mappings":"AAAA;;;;;GAKG;AACH,OAAO,KAAK,EAAE,YAAY,EAAE,gBAAgB,EAAE,MAAM,uBAAuB,CAAC;AAE5E,+BAA+B;AAC/B,MAAM,MAAM,eAAe,GACvB,cAAc,GACd,aAAa,GACb,QAAQ,GACR,MAAM,GACN,YAAY,GACZ,UAAU,GACV,eAAe,GACf,UAAU,CAAC;AAEf,sDAAsD;AACtD,MAAM,MAAM,oBAAoB,GAC5B;IAAE,IAAI,EAAE,MAAM,CAAC;IAAC,IAAI,EAAE,MAAM,CAAA;CAAE,GAC9B;IAAE,IAAI,EAAE,OAAO,CAAA;CAAE,CAAC;AAEtB,yCAAyC;AACzC,MAAM,WAAW,wBAAwB;IACvC,IAAI,EAAE,MAAM,GAAG,WAAW,GAAG,QAAQ,CAAC;IACtC,OAAO,EAAE,MAAM,GAAG,oBAAoB,EAAE,CAAC;CAC1C;AAED,sDAAsD;AACtD,MAAM,WAAW,iCAAiC;IAChD,IAAI,EAAE,WAAW,CAAC;IAClB,OAAO,EAAE,MAAM,CAAC;IAChB,UAAU,EAAE,KAAK,CAAC;QAChB,EAAE,EAAE,MAAM,CAAC;QACX,IAAI,EAAE,UAAU,CAAC;QACjB,QAAQ,EAAE;YACR,IAAI,EAAE,MAAM,CAAC;YACb,SAAS,EAAE,MAAM,CAAC;SACnB,CAAC;KACH,CAAC,CAAC;CACJ;AAED,0CAA0C;AAC1C,MAAM,WAAW,0BAA0B;IACzC,IAAI,EAAE,MAAM,CAAC;IACb,OAAO,EAAE,MAAM,CAAC;IAChB,YAAY,EAAE,MAAM,CAAC;IACrB,IAAI,EAAE,MAAM,CAAC;CACd;AAED,MAAM,MAAM,gBAAgB,GACxB,wBAAwB,GACxB,iCAAiC,GACjC,0BAA0B,CAAC;AAE/B,kDAAkD;AAClD,MAAM,WAAW,uBAAuB;IACtC,IAAI,EAAE,UAAU,CAAC;IACjB,QAAQ,EAAE;QACR,IAAI,EAAE,MAAM,CAAC;QACb,WAAW,CAAC,EAAE,MAAM,CAAC;QACrB,UAAU,CAAC,EAAE,MAAM,CAAC,MAAM,EAAE,OAAO,CAAC,CAAC;KACtC,CAAC;CACH;AAED,qCAAqC;AACrC,MAAM,WAAW,wBAAwB;IACvC,UAAU,CAAC,EAAE,MAAM,CAAC;IACpB,WAAW,CAAC,EAAE,MAAM,CAAC;IACrB,KAAK,CAAC,EAAE,MAAM,CAAC;IACf,KAAK,CAAC,EAAE,MAAM,CAAC;IACf,kBAAkB,CAAC,EAAE,MAAM,CAAC;IAC5B,uBAAuB,CAAC,EAAE,MAAM,CAAC;IACjC,iBAAiB,CAAC,EAAE,OAAO,CAAC;CAC7B;AAED,MAAM,WAAW,oBAAoB;IACnC,MAAM,EAAE,eAAe,CAAC;CACzB;AAED,MAAM,WAAW,4BAA6B,SAAQ,oBAAoB;IACxE,MAAM,EAAE,cAAc,CAAC;CACxB;AAED,MAAM,WAAW,0BAA2B,SAAQ,oBAAoB;IACtE,MAAM,EAAE,aAAa,CAAC;IACtB,QAAQ,EAAE,gBAAgB,EAAE,CAAC;IAC7B,OAAO,CAAC,EAAE;QACR,MAAM,CAAC,EAAE,MAAM,CAAC;KACjB,CAAC;CACH;AAED,MAAM,WAAW,wBAAyB,SAAQ,oBAAoB;IACpE,MAAM,EAAE,UAAU,CAAC;IACnB,QAAQ,EAAE,gBAAgB,EAAE,CAAC;IAC7B,KAAK,CAAC,EAAE,uBAAuB,EAAE,CAAC;IAClC,gBAAgB,CAAC,EAAE,KAAK,GAAG,QAAQ,GAAG,MAAM,CAAC;CAC9C;AAED,MAAM,WAAW,sBAAuB,SAAQ,oBAAoB;IAClE,MAAM,EAAE,QAAQ,CAAC;IACjB,QAAQ,EAAE,gBAAgB,EAAE,CAAC;IAC7B,OAAO,CAAC,EAAE,wBAAwB,GAAG;QACnC,MAAM,CAAC,EAAE,MAAM,CAAC;KACjB,CAAC;IACF,KAAK,CAAC,EAAE,uBAAuB,EAAE,CAAC;IAClC,gBAAgB,CAAC,EAAE,KAAK,GAAG,QAAQ,GAAG,MAAM,CAAC;CAC9C;AAED,gEAAgE;AAChE,MAAM,WAAW,oBAAqB,SAAQ,oBAAoB;IAChE,MAAM,EAAE,MAAM,CAAC;IACf,QAAQ,EAAE,gBAAgB,EAAE,CAAC;IAC7B,MAAM,CAAC,EAAE,MAAM,CAAC;IAChB,KAAK,CAAC,EAAE,uBAAuB,EAAE,CAAC;IAClC,OAAO,CAAC,EAAE,wBAAwB,CAAC;IACnC,MAAM,CAAC,EAAE,MAAM,EAAE,CAAC;IAClB,YAAY,CAAC,EAAE,MAAM,CAAC;IACtB,gBAAgB,CAAC,EAAE,KAAK,GAAG,QAAQ,GAAG,MAAM,CAAC;IAC7C,UAAU,CAAC,EAAE,MAAM,CAAC;IACpB,iBAAiB,CAAC,EAAE,MAAM,CAAC;CAC5B;AAED,MAAM,WAAW,qBAAqB;IACpC,gBAAgB,EAAE,MAAM,GAAG,IAAI,CAAC;IAChC,KAAK,EAAE,MAAM,GAAG,IAAI,CAAC;CACtB;AAED,MAAM,WAAW,0BAA2B,SAAQ,oBAAoB;IACtE,MAAM,EAAE,YAAY,CAAC;IACrB,MAAM,EAAE,MAAM,CAAC;IACf,OAAO,CAAC,EAAE,wBAAwB,CAAC;IACnC,MAAM,CAAC,EAAE,MAAM,EAAE,CAAC;IAClB,YAAY,CAAC,EAAE,MAAM,CAAC;CACvB;AAED,0CAA0C;AAC1C,MAAM,WAAW,wBAAyB,SAAQ,oBAAoB;IACpE,MAAM,EAAE,UAAU,CAAC;IACnB,MAAM,EAAE,MAAM,GAAG,MAAM,EAAE,CAAC;IAC1B,MAAM,CAAC,EAAE,MAAM,CAAC;IAChB,OAAO,CAAC,EAAE,wBAAwB,CAAC;IACnC,MAAM,CAAC,EAAE,MAAM,EAAE,CAAC;IAClB,YAAY,CAAC,EAAE,MAAM,CAAC;IACtB,UAAU,CAAC,EAAE,MAAM,CAAC;IACpB,iBAAiB,CAAC,EAAE,MAAM,CAAC;CAC5B;AAED,MAAM,WAAW,4BAA6B,SAAQ,oBAAoB;IACxE,MAAM,EAAE,eAAe,CAAC;IACxB,UAAU,EAAE,MAAM,CAAC;IACnB,QAAQ,EAAE,gBAAgB,EAAE,CAAC;IAC7B,eAAe,CAAC,EAAE,MAAM,CAAC;IACzB,cAAc,CAAC,EAAE,MAAM,CAAC;IACxB,cAAc,CAAC,EAAE,MAAM,EAAE,CAAC;IAC1B,aAAa,CAAC,EAAE,MAAM,EAAE,CAAC;IACzB,KAAK,CAAC,EAAE,uBAAuB,EAAE,CAAC;IAClC,gBAAgB,CAAC,EAAE,KAAK,GAAG,QAAQ,GAAG,MAAM,CAAC;CAC9C;AAED,MAAM,WAAW,2BAA2B;IAC1C,UAAU,EAAE,MAAM,CAAC;IACnB,WAAW,CAAC,EAAE,MAAM,CAAC;IACrB,cAAc,CAAC,EAAE,MAAM,EAAE,CAAC;IAC1B,aAAa,CAAC,EAAE,MAAM,EAAE,CAAC;CAC1B;AAED,MAAM,MAAM,gBAAgB,GACxB,4BAA4B,GAC5B,0BAA0B,GAC1B,sBAAsB,GACtB,wBAAwB,GACxB,oBAAoB,GACpB,0BAA0B,GAC1B,wBAAwB,GACxB,4BAA4B,CAAC;AAEjC;;kCAEkC;AAClC,MAAM,MAAM,mBAAmB,GAC3B,YAAY,GACZ,UAAU,GACV,gBAAgB,GAChB,SAAS,GACT,SAAS,GACT,SAAS,GACT,OAAO,GACP,aAAa,GACb,YAAY,GACZ,QAAQ,GACR,SAAS,CAAC;AAEd,MAAM,WAAW,cAAc;IAC7B,gBAAgB,CAAC,EAAE,MAAM,CAAC;IAC1B,UAAU,CAAC,EAAE,MAAM,CAAC;IACpB,QAAQ,CAAC,EAAE,MAAM,CAAC;IAClB,cAAc,CAAC,EAAE,MAAM,CAAC;IACxB,YAAY,CAAC,EAAE,MAAM,CAAC;CACvB;AAED,MAAM,WAAW,gBAAgB;IAC/B,eAAe,EAAE,MAAM,EAAE,CAAC;IAC1B,OAAO,CAAC,EAAE,MAAM,CAAC;IACjB,WAAW,EAAE,MAAM,CAAC,MAAM,EAAE,OAAO,CAAC,CAAC;IACrC,gBAAgB,CAAC,EAAE,cAAc,CAAC;CACnC;AAED;;GAEG;AACH,MAAM,WAAW,qBAAqB;IACpC,OAAO,EAAE,MAAM,EAAE,CAAC;IAClB,cAAc,EAAE,MAAM,CAAC,MAAM,EAAE,YAAY,GAAG,gBAAgB,CAAC,CAAC;IAChE,QAAQ,EAAE;QACR,mBAAmB,EAAE,OAAO,CAAC;QAC7B,UAAU,CAAC,EAAE,MAAM,CAAC;QACpB,gBAAgB,CAAC,EAAE,MAAM,CAAC;QAC1B,aAAa,CAAC,EAAE,gBAAgB,CAAC;KAClC,CAAC;IACF,iBAAiB,CAAC,EAAE;QAClB,sBAAsB,CAAC,EAAE,OAAO,CAAC;QACjC,mBAAmB,CAAC,EAAE,MAAM,CAAC;QAC7B,iBAAiB,CAAC,EAAE,OAAO,CAAC;QAC5B,kBAAkB,CAAC,EAAE,OAAO,CAAC;QAC7B,oBAAoB,CAAC,EAAE,OAAO,CAAC;KAChC,CAAC;IACF,UAAU,CAAC,EAAE,IAAI,GAAG,KAAK,CAAC;CAC3B;AAED,MAAM,WAAW,yBAAyB;IACxC,gBAAgB,EAAE,MAAM,GAAG,IAAI,CAAC;IAChC,gBAAgB,EAAE,OAAO,CAAC;IAC1B,yBAAyB,EAAE,gBAAgB,EAAE,GAAG,IAAI,CAAC;IACrD,KAAK,EAAE,MAAM,GAAG,IAAI,CAAC;CACtB;AAED,MAAM,WAAW,uBAAuB;IACtC,SAAS,EAAE,MAAM,EAAE,GAAG,IAAI,CAAC;IAC3B,WAAW,EAAE,MAAM,CAAC;IACpB,KAAK,EAAE,MAAM,GAAG,IAAI,CAAC;CACtB"}
|