@modular-prompt/driver 0.16.0 → 0.17.0
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 +93 -10
- package/dist/cache-controller.d.ts +4 -0
- package/dist/cache-controller.d.ts.map +1 -1
- package/dist/driver-registry/config-based-factory.d.ts.map +1 -1
- package/dist/driver-registry/config-based-factory.js +6 -0
- package/dist/driver-registry/config-based-factory.js.map +1 -1
- package/dist/driver-registry/factory-helper.d.ts.map +1 -1
- package/dist/driver-registry/factory-helper.js +9 -2
- package/dist/driver-registry/factory-helper.js.map +1 -1
- package/dist/driver-registry/index.d.ts +1 -1
- package/dist/driver-registry/index.d.ts.map +1 -1
- package/dist/driver-registry/types.d.ts +15 -1
- package/dist/driver-registry/types.d.ts.map +1 -1
- package/dist/formatter/converter.d.ts.map +1 -1
- package/dist/formatter/converter.js +31 -2
- package/dist/formatter/converter.js.map +1 -1
- package/dist/index.d.ts +5 -3
- package/dist/index.d.ts.map +1 -1
- package/dist/index.js +5 -3
- package/dist/index.js.map +1 -1
- package/dist/local-inference/adapters.d.ts +6 -0
- package/dist/local-inference/adapters.d.ts.map +1 -1
- package/dist/local-inference/driver.d.ts.map +1 -1
- package/dist/local-inference/driver.js +45 -24
- package/dist/local-inference/driver.js.map +1 -1
- package/dist/local-inference/process-client.d.ts +4 -2
- package/dist/local-inference/process-client.d.ts.map +1 -1
- package/dist/local-inference/process-client.js +24 -8
- package/dist/local-inference/process-client.js.map +1 -1
- package/dist/local-inference/process-communication.d.ts +9 -2
- package/dist/local-inference/process-communication.d.ts.map +1 -1
- package/dist/local-inference/process-communication.js +37 -5
- package/dist/local-inference/process-communication.js.map +1 -1
- package/dist/local-inference/protocol.d.ts +4 -0
- package/dist/local-inference/protocol.d.ts.map +1 -1
- package/dist/local-inference/request-queue.d.ts +1 -1
- package/dist/local-inference/request-queue.d.ts.map +1 -1
- package/dist/local-inference/request-queue.js +26 -7
- package/dist/local-inference/request-queue.js.map +1 -1
- package/dist/local-inference/stream-utils.d.ts +6 -0
- package/dist/local-inference/stream-utils.d.ts.map +1 -1
- package/dist/local-inference/stream-utils.js.map +1 -1
- package/dist/mlx-ml/mlx-cache-controller.d.ts +9 -0
- package/dist/mlx-ml/mlx-cache-controller.d.ts.map +1 -1
- package/dist/mlx-ml/mlx-cache-controller.js +158 -33
- package/dist/mlx-ml/mlx-cache-controller.js.map +1 -1
- package/dist/mlx-ml/mlx-cache-support.d.ts +2 -2
- package/dist/mlx-ml/mlx-cache-support.d.ts.map +1 -1
- package/dist/mlx-ml/mlx-cache-support.js +8 -3
- package/dist/mlx-ml/mlx-cache-support.js.map +1 -1
- package/dist/mlx-ml/mlx-driver.d.ts +0 -1
- package/dist/mlx-ml/mlx-driver.d.ts.map +1 -1
- package/dist/mlx-ml/mlx-driver.js +1 -8
- package/dist/mlx-ml/mlx-driver.js.map +1 -1
- package/dist/mlx-ml/process/index.d.ts +1 -1
- package/dist/mlx-ml/process/index.d.ts.map +1 -1
- package/dist/mlx-ml/process/index.js +2 -2
- package/dist/mlx-ml/process/index.js.map +1 -1
- package/dist/models-config/index.d.ts +1 -1
- package/dist/models-config/index.d.ts.map +1 -1
- package/dist/models-config/index.js +1 -1
- package/dist/models-config/index.js.map +1 -1
- package/dist/models-config/resolve.d.ts +9 -1
- package/dist/models-config/resolve.d.ts.map +1 -1
- package/dist/models-config/resolve.js +94 -2
- package/dist/models-config/resolve.js.map +1 -1
- package/dist/models-config/types.d.ts +3 -1
- package/dist/models-config/types.d.ts.map +1 -1
- package/dist/pytorch/process/index.d.ts +4 -2
- package/dist/pytorch/process/index.d.ts.map +1 -1
- package/dist/pytorch/process/index.js +24 -7
- package/dist/pytorch/process/index.js.map +1 -1
- package/dist/pytorch/pytorch-cache-controller.d.ts +84 -0
- package/dist/pytorch/pytorch-cache-controller.d.ts.map +1 -0
- package/dist/pytorch/pytorch-cache-controller.js +742 -0
- package/dist/pytorch/pytorch-cache-controller.js.map +1 -0
- package/dist/pytorch/pytorch-cache-support.d.ts +23 -0
- package/dist/pytorch/pytorch-cache-support.d.ts.map +1 -0
- package/dist/pytorch/pytorch-cache-support.js +47 -0
- package/dist/pytorch/pytorch-cache-support.js.map +1 -0
- package/dist/pytorch/pytorch-driver.d.ts +8 -1
- package/dist/pytorch/pytorch-driver.d.ts.map +1 -1
- package/dist/pytorch/pytorch-driver.js +40 -0
- package/dist/pytorch/pytorch-driver.js.map +1 -1
- package/dist/runtime/check.d.ts.map +1 -1
- package/dist/runtime/check.js +8 -6
- package/dist/runtime/check.js.map +1 -1
- package/dist/runtime/index.d.ts +2 -2
- package/dist/runtime/index.d.ts.map +1 -1
- package/dist/runtime/index.js +2 -2
- package/dist/runtime/index.js.map +1 -1
- package/dist/runtime/manifest-core.d.mts +1 -0
- package/dist/runtime/manifest-core.mjs +1 -0
- package/dist/runtime/manifest-core.mjs.map +1 -1
- package/dist/runtime/manifest.d.ts +2 -0
- package/dist/runtime/manifest.d.ts.map +1 -1
- package/dist/runtime/manifest.js.map +1 -1
- package/dist/runtime/paths-core.d.mts +15 -1
- package/dist/runtime/paths-core.d.mts.map +1 -1
- package/dist/runtime/paths-core.mjs +50 -5
- package/dist/runtime/paths-core.mjs.map +1 -1
- package/dist/runtime/paths.d.ts +2 -2
- package/dist/runtime/paths.d.ts.map +1 -1
- package/dist/runtime/paths.js +2 -2
- package/dist/runtime/paths.js.map +1 -1
- package/dist/runtime/pytorch-template-core.d.mts +11 -0
- package/dist/runtime/pytorch-template-core.d.mts.map +1 -0
- package/dist/runtime/pytorch-template-core.mjs +54 -0
- package/dist/runtime/pytorch-template-core.mjs.map +1 -0
- package/dist/runtime/setup-commands-core.d.mts +3 -0
- package/dist/runtime/setup-commands-core.d.mts.map +1 -1
- package/dist/runtime/setup-commands-core.mjs +4 -0
- package/dist/runtime/setup-commands-core.mjs.map +1 -1
- package/dist/runtime/setup-commands.d.ts +1 -1
- package/dist/runtime/setup-commands.d.ts.map +1 -1
- package/dist/runtime/setup-commands.js +1 -1
- package/dist/runtime/setup-commands.js.map +1 -1
- package/docs/DRIVER_API.md +455 -0
- package/docs/LOCAL_MODEL_SETUP.md +765 -0
- package/docs/mlx-api-selection.md +301 -0
- package/package.json +9 -5
- package/scripts/runtime-cli.bin.test.ts +142 -0
- package/scripts/runtime-cli.js +305 -35
- package/scripts/runtime-cli.test.ts +163 -0
- package/src/mlx-ml/python/__main__.py +1 -1
- package/src/mlx-ml/python/backends/base.py +88 -18
- package/src/mlx-ml/python/backends/mlx_lm.py +28 -3
- package/src/mlx-ml/python/backends/mlx_vlm.py +679 -2
- package/src/mlx-ml/python/handlers/cache.py +4 -0
- package/src/mlx-ml/python/handlers/generate.py +33 -10
- package/src/mlx-ml/python/handlers/tokenize.py +1 -4
- package/src/mlx-ml/python/pyproject.toml +1 -1
- package/src/mlx-ml/python/server.py +2 -0
- package/src/mlx-ml/python/uv.lock +8 -8
- package/src/pytorch/templates/cpu-minimal/backends/base.py +139 -0
- package/src/pytorch/templates/cpu-minimal/backends/transformers_lm.py +1167 -0
- package/src/pytorch/{python → templates/cpu-minimal}/handlers/__init__.py +1 -0
- package/src/pytorch/templates/cpu-minimal/handlers/cache.py +88 -0
- package/src/pytorch/templates/cpu-minimal/handlers/generate.py +157 -0
- package/src/pytorch/{python → templates/cpu-minimal}/pyproject.toml +2 -2
- package/src/pytorch/{python → templates/cpu-minimal}/server.py +20 -2
- package/src/pytorch/templates/cpu-minimal/tests/test_cache_handler.py +284 -0
- package/src/pytorch/templates/cpu-minimal/tests/test_capabilities.py +14 -0
- package/src/pytorch/templates/cpu-minimal/tests/test_server.py +141 -0
- package/src/pytorch/templates/cpu-minimal/tests/test_transformers_errors.py +89 -0
- package/src/pytorch/templates/cpu-minimal/tests/test_transformers_lm_cache.py +554 -0
- package/src/pytorch/templates/cpu-minimal/utils/__init__.py +0 -0
- package/src/pytorch/{python → templates/cpu-minimal}/utils/token_utils.py +2 -2
- package/src/pytorch/templates/cpu-minimal/utils/transformers_errors.py +54 -0
- package/src/pytorch/{python → templates/cpu-minimal}/uv.lock +149 -109
- package/src/pytorch/templates/cuda/__main__.py +19 -0
- package/src/pytorch/templates/cuda/backends/__init__.py +3 -0
- package/src/pytorch/{python → templates/cuda}/backends/base.py +54 -6
- package/src/pytorch/templates/cuda/backends/transformers_lm.py +379 -0
- package/src/pytorch/templates/cuda/handlers/__init__.py +7 -0
- package/src/pytorch/templates/cuda/handlers/cache.py +93 -0
- package/src/pytorch/templates/cuda/handlers/cancel.py +53 -0
- package/src/pytorch/templates/cuda/handlers/capabilities.py +6 -0
- package/src/pytorch/templates/cuda/handlers/completion.py +15 -0
- package/src/pytorch/templates/cuda/handlers/format_test.py +70 -0
- package/src/pytorch/templates/cuda/handlers/generate.py +152 -0
- package/src/pytorch/templates/cuda/handlers/render.py +40 -0
- package/src/pytorch/templates/cuda/handlers/tokenize.py +63 -0
- package/src/pytorch/templates/cuda/pyproject.toml +37 -0
- package/src/pytorch/templates/cuda/server.py +158 -0
- package/src/pytorch/templates/cuda/tests/__init__.py +0 -0
- package/src/pytorch/templates/cuda/tests/test_cache_handler.py +207 -0
- package/src/pytorch/templates/cuda/tests/test_capabilities.py +14 -0
- package/src/pytorch/templates/cuda/tests/test_server.py +145 -0
- package/src/pytorch/templates/cuda/tests/test_transformers_errors.py +89 -0
- package/src/pytorch/templates/cuda/tests/test_transformers_lm_cache.py +288 -0
- package/src/pytorch/templates/cuda/utils/__init__.py +0 -0
- package/src/pytorch/templates/cuda/utils/chat_template_constraints.py +164 -0
- package/src/pytorch/templates/cuda/utils/prompt_builder.py +54 -0
- package/src/pytorch/templates/cuda/utils/template_render.py +80 -0
- package/src/pytorch/templates/cuda/utils/token_utils.py +376 -0
- package/src/pytorch/templates/cuda/utils/transformers_errors.py +54 -0
- package/src/pytorch/templates/cuda/uv.lock +734 -0
- package/src/pytorch/python/backends/transformers_lm.py +0 -127
- package/src/pytorch/python/handlers/generate.py +0 -68
- /package/src/pytorch/{python → templates/cpu-minimal}/__main__.py +0 -0
- /package/src/pytorch/{python → templates/cpu-minimal}/backends/__init__.py +0 -0
- /package/src/pytorch/{python → templates/cpu-minimal}/handlers/cancel.py +0 -0
- /package/src/pytorch/{python → templates/cpu-minimal}/handlers/capabilities.py +0 -0
- /package/src/pytorch/{python → templates/cpu-minimal}/handlers/completion.py +0 -0
- /package/src/pytorch/{python → templates/cpu-minimal}/handlers/format_test.py +0 -0
- /package/src/pytorch/{python → templates/cpu-minimal}/handlers/render.py +0 -0
- /package/src/pytorch/{python → templates/cpu-minimal}/handlers/tokenize.py +0 -0
- /package/src/pytorch/{python/utils → templates/cpu-minimal/tests}/__init__.py +0 -0
- /package/src/pytorch/{python → templates/cpu-minimal}/utils/chat_template_constraints.py +0 -0
- /package/src/pytorch/{python → templates/cpu-minimal}/utils/prompt_builder.py +0 -0
- /package/src/pytorch/{python → templates/cpu-minimal}/utils/template_render.py +0 -0
|
@@ -2,13 +2,14 @@
|
|
|
2
2
|
* PyTorch Driver 外部インターフェース
|
|
3
3
|
*/
|
|
4
4
|
import { Readable } from 'stream';
|
|
5
|
-
import type { InferenceCapabilities, InferenceFormatTestResult, InferenceMessage, InferenceRenderResult, InferenceTokenizeResult, InferenceToolDefinition } from '../../local-inference/protocol.js';
|
|
5
|
+
import type { InferenceCapabilities, InferenceCachePrefillResult, InferenceFormatTestResult, InferenceMessage, InferenceRenderResult, InferenceTokenizeResult, InferenceToolDefinition } from '../../local-inference/protocol.js';
|
|
6
6
|
export interface PyTorchProcessOptions {
|
|
7
7
|
/** デフォルトの ~/.modular-prompt/runtimes/pytorch/.venv を上書き */
|
|
8
8
|
venvPath?: string;
|
|
9
9
|
/** Python 子プロセスへ渡す PYTORCH_DEVICE(例: cpu, cuda) */
|
|
10
10
|
device?: string;
|
|
11
11
|
}
|
|
12
|
+
export declare const PYTORCH_CUDA_UNAVAILABLE_ERROR: string;
|
|
12
13
|
export declare class PyTorchProcess {
|
|
13
14
|
readonly modelName: string;
|
|
14
15
|
private readonly client;
|
|
@@ -22,7 +23,8 @@ export declare class PyTorchProcess {
|
|
|
22
23
|
primer?: string;
|
|
23
24
|
}, tools?: InferenceToolDefinition[], reasoningEffort?: 'low' | 'medium' | 'high'): Promise<InferenceRenderResult>;
|
|
24
25
|
tokenize(messages: InferenceMessage[], tools?: InferenceToolDefinition[], reasoningEffort?: 'low' | 'medium' | 'high'): Promise<InferenceTokenizeResult>;
|
|
25
|
-
|
|
26
|
+
cachePrefill(cachePath: string, messages: InferenceMessage[], baseCachePath?: string, trimToTokens?: number, prefixOffsets?: number[], prefixHashes?: string[], tools?: InferenceToolDefinition[], reasoningEffort?: 'low' | 'medium' | 'high', images?: string[], maxImageSize?: number): Promise<InferenceCachePrefillResult>;
|
|
27
|
+
generate(prompt: string | number[], options?: Record<string, unknown>, images?: string[], maxImageSize?: number, cachePath?: string, cacheTrimTokens?: number, primer?: string): Promise<Readable>;
|
|
26
28
|
exit(): Promise<void>;
|
|
27
29
|
cancelActiveRequest(): void;
|
|
28
30
|
getStatus(): {
|
|
@@ -1 +1 @@
|
|
|
1
|
-
{"version":3,"file":"index.d.ts","sourceRoot":"","sources":["../../../src/pytorch/process/index.ts"],"names":[],"mappings":"AAAA;;GAEG;AAEH,OAAO,EAAE,QAAQ,EAAE,MAAM,QAAQ,CAAC;
|
|
1
|
+
{"version":3,"file":"index.d.ts","sourceRoot":"","sources":["../../../src/pytorch/process/index.ts"],"names":[],"mappings":"AAAA;;GAEG;AAEH,OAAO,EAAE,QAAQ,EAAE,MAAM,QAAQ,CAAC;AAUlC,OAAO,KAAK,EACV,qBAAqB,EACrB,2BAA2B,EAC3B,yBAAyB,EACzB,gBAAgB,EAChB,qBAAqB,EACrB,uBAAuB,EACvB,uBAAuB,EACxB,MAAM,mCAAmC,CAAC;AAM3C,MAAM,WAAW,qBAAqB;IACpC,2DAA2D;IAC3D,QAAQ,CAAC,EAAE,MAAM,CAAC;IAClB,mDAAmD;IACnD,MAAM,CAAC,EAAE,MAAM,CAAC;CACjB;AAED,eAAO,MAAM,8BAA8B,QAEyB,CAAC;AA4BrE,qBAAa,cAAc;IACzB,QAAQ,CAAC,SAAS,EAAE,MAAM,CAAC;IAC3B,OAAO,CAAC,QAAQ,CAAC,MAAM,CAAyB;gBAEpC,SAAS,EAAE,MAAM,EAAE,OAAO,CAAC,EAAE,qBAAqB;IAqCxD,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,MAAM,CAAC,MAAM,EAAE,OAAO,CAAC,GAAG;QAAE,MAAM,CAAC,EAAE,MAAM,CAAA;KAAE,EACvD,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,EAC3C,MAAM,CAAC,EAAE,MAAM,EAAE,EACjB,YAAY,CAAC,EAAE,MAAM,GACpB,OAAO,CAAC,2BAA2B,CAAC;IAejC,QAAQ,CACZ,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;IAYd,IAAI,IAAI,OAAO,CAAC,IAAI,CAAC;IAI3B,mBAAmB,IAAI,IAAI;IAI3B,SAAS;;;;;;CAGV"}
|
|
@@ -2,11 +2,20 @@
|
|
|
2
2
|
* PyTorch Driver 外部インターフェース
|
|
3
3
|
*/
|
|
4
4
|
import { existsSync } from 'fs';
|
|
5
|
-
import {
|
|
5
|
+
import { join } from 'path';
|
|
6
|
+
import { getPytorchRuntimePythonDir, getVenvPath, SETUP_PYTORCH_CLI, SETUP_PYTORCH_MONOREPO, } from '../../runtime/index.js';
|
|
6
7
|
import { InferenceProcessClient } from '../../local-inference/process-client.js';
|
|
7
8
|
import { mapOptionsToPython } from '../../mlx-ml/process/parameter-mapper.js';
|
|
8
|
-
const
|
|
9
|
-
const
|
|
9
|
+
const pytorchPythonDir = getPytorchRuntimePythonDir();
|
|
10
|
+
export const PYTORCH_CUDA_UNAVAILABLE_ERROR = 'CUDA device requested, but CUDA is not available in this PyTorch runtime. ' +
|
|
11
|
+
'Install a CUDA-enabled torch wheel and verify the NVIDIA driver.';
|
|
12
|
+
const PYTORCH_CUDA_ERROR_MARKER = 'CUDA device requested, but CUDA is not available';
|
|
13
|
+
function formatPytorchProcessExitError(code, signal, stderr) {
|
|
14
|
+
if (stderr?.includes(PYTORCH_CUDA_ERROR_MARKER)) {
|
|
15
|
+
return PYTORCH_CUDA_UNAVAILABLE_ERROR;
|
|
16
|
+
}
|
|
17
|
+
return `PyTorch process exited unexpectedly (code=${code}, signal=${signal})`;
|
|
18
|
+
}
|
|
10
19
|
function resolveVenvPath(options) {
|
|
11
20
|
return (options?.venvPath ??
|
|
12
21
|
process.env.MODULAR_PROMPT_PYTORCH_VENV ??
|
|
@@ -20,10 +29,15 @@ export class PyTorchProcess {
|
|
|
20
29
|
client;
|
|
21
30
|
constructor(modelName, options) {
|
|
22
31
|
this.modelName = modelName;
|
|
32
|
+
if (!existsSync(join(pytorchPythonDir, 'pyproject.toml')) ||
|
|
33
|
+
!existsSync(join(pytorchPythonDir, '__main__.py'))) {
|
|
34
|
+
throw new Error(`PyTorch runtime Python project not found at ${pytorchPythonDir}. ` +
|
|
35
|
+
`Run: ${SETUP_PYTORCH_MONOREPO} (monorepo) or ${SETUP_PYTORCH_CLI} (npm).`);
|
|
36
|
+
}
|
|
23
37
|
const venvPath = resolveVenvPath(options);
|
|
24
38
|
if (!existsSync(venvPath)) {
|
|
25
39
|
throw new Error(`PyTorch venv not found at ${venvPath}. ` +
|
|
26
|
-
`Run: ${SETUP_PYTORCH_MONOREPO}
|
|
40
|
+
`Run: ${SETUP_PYTORCH_MONOREPO} (monorepo) or ${SETUP_PYTORCH_CLI} (npm).`);
|
|
27
41
|
}
|
|
28
42
|
const device = options?.device ?? process.env.PYTORCH_DEVICE;
|
|
29
43
|
const extraEnv = device ? { PYTORCH_DEVICE: device } : undefined;
|
|
@@ -35,7 +49,7 @@ export class PyTorchProcess {
|
|
|
35
49
|
extraEnv,
|
|
36
50
|
loggerPrefix: 'PyTorch',
|
|
37
51
|
mapSamplingOptions: (opts) => mapOptionsToPython(opts, false),
|
|
38
|
-
processExitErrorMessage: (code, signal) =>
|
|
52
|
+
processExitErrorMessage: (code, signal, stderr) => formatPytorchProcessExitError(code, signal, stderr),
|
|
39
53
|
});
|
|
40
54
|
}
|
|
41
55
|
async ensureInitialized() {
|
|
@@ -53,8 +67,11 @@ export class PyTorchProcess {
|
|
|
53
67
|
async tokenize(messages, tools, reasoningEffort) {
|
|
54
68
|
return this.client.tokenize(messages, tools, reasoningEffort);
|
|
55
69
|
}
|
|
56
|
-
async
|
|
57
|
-
return this.client.
|
|
70
|
+
async cachePrefill(cachePath, messages, baseCachePath, trimToTokens, prefixOffsets, prefixHashes, tools, reasoningEffort, images, maxImageSize) {
|
|
71
|
+
return this.client.cachePrefill(cachePath, messages, baseCachePath, trimToTokens, prefixOffsets, prefixHashes, tools, reasoningEffort, images, maxImageSize);
|
|
72
|
+
}
|
|
73
|
+
async generate(prompt, options, images, maxImageSize, cachePath, cacheTrimTokens, primer) {
|
|
74
|
+
return this.client.generate(prompt, options, images, maxImageSize, cachePath, cacheTrimTokens, primer);
|
|
58
75
|
}
|
|
59
76
|
async exit() {
|
|
60
77
|
return this.client.exit();
|
|
@@ -1 +1 @@
|
|
|
1
|
-
{"version":3,"file":"index.js","sourceRoot":"","sources":["../../../src/pytorch/process/index.ts"],"names":[],"mappings":"AAAA;;GAEG;AAGH,OAAO,EAAE,UAAU,EAAE,MAAM,IAAI,CAAC;AAChC,OAAO,EACL,
|
|
1
|
+
{"version":3,"file":"index.js","sourceRoot":"","sources":["../../../src/pytorch/process/index.ts"],"names":[],"mappings":"AAAA;;GAEG;AAGH,OAAO,EAAE,UAAU,EAAE,MAAM,IAAI,CAAC;AAChC,OAAO,EAAE,IAAI,EAAE,MAAM,MAAM,CAAC;AAC5B,OAAO,EACL,0BAA0B,EAC1B,WAAW,EACX,iBAAiB,EACjB,sBAAsB,GACvB,MAAM,wBAAwB,CAAC;AAChC,OAAO,EAAE,sBAAsB,EAAE,MAAM,yCAAyC,CAAC;AAUjF,OAAO,EAAE,kBAAkB,EAAE,MAAM,0CAA0C,CAAC;AAG9E,MAAM,gBAAgB,GAAG,0BAA0B,EAAE,CAAC;AAStD,MAAM,CAAC,MAAM,8BAA8B,GACzC,4EAA4E;IAC5E,kEAAkE,CAAC;AAErE,MAAM,yBAAyB,GAC7B,kDAAkD,CAAC;AAErD,SAAS,6BAA6B,CACpC,IAAmB,EACnB,MAAqB,EACrB,MAAe;IAEf,IAAI,MAAM,EAAE,QAAQ,CAAC,yBAAyB,CAAC,EAAE,CAAC;QAChD,OAAO,8BAA8B,CAAC;IACxC,CAAC;IACD,OAAO,6CAA6C,IAAI,YAAY,MAAM,GAAG,CAAC;AAChF,CAAC;AAED,SAAS,eAAe,CAAC,OAA+B;IACtD,OAAO,CACL,OAAO,EAAE,QAAQ;QACjB,OAAO,CAAC,GAAG,CAAC,2BAA2B;QACvC,WAAW,CAAC,SAAS,CAAC,CACvB,CAAC;AACJ,CAAC;AAED,SAAS,sBAAsB,CAAC,QAAgB;IAC9C,OAAO,QAAQ,KAAK,WAAW,CAAC,SAAS,CAAC,CAAC;AAC7C,CAAC;AAED,MAAM,OAAO,cAAc;IAChB,SAAS,CAAS;IACV,MAAM,CAAyB;IAEhD,YAAY,SAAiB,EAAE,OAA+B;QAC5D,IAAI,CAAC,SAAS,GAAG,SAAS,CAAC;QAE3B,IACE,CAAC,UAAU,CAAC,IAAI,CAAC,gBAAgB,EAAE,gBAAgB,CAAC,CAAC;YACrD,CAAC,UAAU,CAAC,IAAI,CAAC,gBAAgB,EAAE,aAAa,CAAC,CAAC,EAClD,CAAC;YACD,MAAM,IAAI,KAAK,CACb,+CAA+C,gBAAgB,IAAI;gBACjE,QAAQ,sBAAsB,kBAAkB,iBAAiB,SAAS,CAC7E,CAAC;QACJ,CAAC;QAED,MAAM,QAAQ,GAAG,eAAe,CAAC,OAAO,CAAC,CAAC;QAC1C,IAAI,CAAC,UAAU,CAAC,QAAQ,CAAC,EAAE,CAAC;YAC1B,MAAM,IAAI,KAAK,CACb,6BAA6B,QAAQ,IAAI;gBACvC,QAAQ,sBAAsB,kBAAkB,iBAAiB,SAAS,CAC7E,CAAC;QACJ,CAAC;QAED,MAAM,MAAM,GAAG,OAAO,EAAE,MAAM,IAAI,OAAO,CAAC,GAAG,CAAC,cAAc,CAAC;QAC7D,MAAM,QAAQ,GAAG,MAAM,CAAC,CAAC,CAAC,EAAE,cAAc,EAAE,MAAM,EAAE,CAAC,CAAC,CAAC,SAAS,CAAC;QAEjE,IAAI,CAAC,MAAM,GAAG,IAAI,sBAAsB,CAAC;YACvC,SAAS;YACT,gBAAgB,EAAE,gBAAgB;YAClC,QAAQ;YACR,cAAc,EAAE,sBAAsB,CAAC,QAAQ,CAAC,CAAC,CAAC,CAAC,SAAS,CAAC,CAAC,CAAC,SAAS;YACxE,QAAQ;YACR,YAAY,EAAE,SAAS;YACvB,kBAAkB,EAAE,CAAC,IAAI,EAAE,EAAE,CAAC,kBAAkB,CAAC,IAAuC,EAAE,KAAK,CAAC;YAChG,uBAAuB,EAAE,CAAC,IAAI,EAAE,MAAM,EAAE,MAAM,EAAE,EAAE,CAChD,6BAA6B,CAAC,IAAI,EAAE,MAAM,EAAE,MAAM,CAAC;SACtD,CAAC,CAAC;IACL,CAAC;IAED,KAAK,CAAC,iBAAiB;QACrB,OAAO,IAAI,CAAC,MAAM,CAAC,iBAAiB,EAAE,CAAC;IACzC,CAAC;IAED,KAAK,CAAC,eAAe;QACnB,OAAO,IAAI,CAAC,MAAM,CAAC,eAAe,EAAE,CAAC;IACvC,CAAC;IAED,KAAK,CAAC,UAAU,CACd,QAA4B,EAC5B,OAA6B;QAE7B,OAAO,IAAI,CAAC,MAAM,CAAC,UAAU,CAAC,QAAQ,EAAE,OAAO,CAAC,CAAC;IACnD,CAAC;IAED,KAAK,CAAC,MAAM,CACV,QAA4B,EAC5B,OAAuD,EACvD,KAAiC,EACjC,eAA2C;QAE3C,OAAO,IAAI,CAAC,MAAM,CAAC,MAAM,CAAC,QAAQ,EAAE,OAAO,EAAE,KAAK,EAAE,eAAe,CAAC,CAAC;IACvE,CAAC;IAED,KAAK,CAAC,QAAQ,CACZ,QAA4B,EAC5B,KAAiC,EACjC,eAA2C;QAE3C,OAAO,IAAI,CAAC,MAAM,CAAC,QAAQ,CAAC,QAAQ,EAAE,KAAK,EAAE,eAAe,CAAC,CAAC;IAChE,CAAC;IAED,KAAK,CAAC,YAAY,CAChB,SAAiB,EACjB,QAA4B,EAC5B,aAAsB,EACtB,YAAqB,EACrB,aAAwB,EACxB,YAAuB,EACvB,KAAiC,EACjC,eAA2C,EAC3C,MAAiB,EACjB,YAAqB;QAErB,OAAO,IAAI,CAAC,MAAM,CAAC,YAAY,CAC7B,SAAS,EACT,QAAQ,EACR,aAAa,EACb,YAAY,EACZ,aAAa,EACb,YAAY,EACZ,KAAK,EACL,eAAe,EACf,MAAM,EACN,YAAY,CACb,CAAC;IACJ,CAAC;IAED,KAAK,CAAC,QAAQ,CACZ,MAAyB,EACzB,OAAiC,EACjC,MAAiB,EACjB,YAAqB,EACrB,SAAkB,EAClB,eAAwB,EACxB,MAAe;QAEf,OAAO,IAAI,CAAC,MAAM,CAAC,QAAQ,CACzB,MAAM,EACN,OAAO,EACP,MAAM,EACN,YAAY,EACZ,SAAS,EACT,eAAe,EACf,MAAM,CACP,CAAC;IACJ,CAAC;IAED,KAAK,CAAC,IAAI;QACR,OAAO,IAAI,CAAC,MAAM,CAAC,IAAI,EAAE,CAAC;IAC5B,CAAC;IAED,mBAAmB;QACjB,IAAI,CAAC,MAAM,CAAC,mBAAmB,EAAE,CAAC;IACpC,CAAC;IAED,SAAS;QACP,OAAO,IAAI,CAAC,MAAM,CAAC,SAAS,EAAE,CAAC;IACjC,CAAC;CACF"}
|
|
@@ -0,0 +1,84 @@
|
|
|
1
|
+
import type { CacheHandle, CachePrepareParams, PromptCacheController } from '../cache-controller.js';
|
|
2
|
+
import type { FormatterOptions } from '../formatter/types.js';
|
|
3
|
+
import type { InferenceMessage } from '../local-inference/protocol.js';
|
|
4
|
+
import type { PyTorchProcess } from './process/index.js';
|
|
5
|
+
export interface PyTorchCacheControllerOptions {
|
|
6
|
+
/** 固定キャッシュディレクトリ。指定時はauto-cleanupが無効になる */
|
|
7
|
+
cacheDir?: string;
|
|
8
|
+
}
|
|
9
|
+
/**
|
|
10
|
+
* PyTorch Transformers backend 用の PromptCacheController。
|
|
11
|
+
*
|
|
12
|
+
* キャッシュ本体の形式・検証・寿命は Python backend が所有する。
|
|
13
|
+
* このクラスは、プロンプト要素から安定した ref を生成し、LIP の
|
|
14
|
+
* cache_prefill / generate と PromptCacheController の usage 契約を橋渡しする。
|
|
15
|
+
*/
|
|
16
|
+
export declare class PyTorchCacheController implements PromptCacheController {
|
|
17
|
+
private cacheByHash;
|
|
18
|
+
private inflightRequests;
|
|
19
|
+
private process?;
|
|
20
|
+
private cacheDir;
|
|
21
|
+
private managedDir;
|
|
22
|
+
private cacheDirReady;
|
|
23
|
+
private closed;
|
|
24
|
+
private bound;
|
|
25
|
+
private modelKind;
|
|
26
|
+
private cleanupHandler?;
|
|
27
|
+
private messageProcessor?;
|
|
28
|
+
private formatterOptions;
|
|
29
|
+
private lastHandle?;
|
|
30
|
+
private lastElementHashes?;
|
|
31
|
+
private lastHandleModel?;
|
|
32
|
+
private lastHandleFormatterOptionsHash?;
|
|
33
|
+
private lastHandleToolsHash?;
|
|
34
|
+
private lastHandleReasoningEffort?;
|
|
35
|
+
private cacheIndex;
|
|
36
|
+
/** Token counts returned by the backend for process-local refs. */
|
|
37
|
+
private cacheTokenCounts;
|
|
38
|
+
private stats;
|
|
39
|
+
private static readonly EMPTY_HANDLE;
|
|
40
|
+
constructor(options?: PyTorchCacheControllerOptions);
|
|
41
|
+
/** Select backend-local storage before the controller is bound. */
|
|
42
|
+
setModelKind(modelKind?: 'lm' | 'vlm'): void;
|
|
43
|
+
bind(process: PyTorchProcess, formatterOptions: FormatterOptions, messageProcessor?: (messages: InferenceMessage[]) => InferenceMessage[]): Promise<void>;
|
|
44
|
+
private ensureCacheDir;
|
|
45
|
+
private get indexPath();
|
|
46
|
+
private readMetaTokenCount;
|
|
47
|
+
private readPrefixMeta;
|
|
48
|
+
private loadIndex;
|
|
49
|
+
private saveIndex;
|
|
50
|
+
private computeFormatterOptionsHash;
|
|
51
|
+
private computeToolsHash;
|
|
52
|
+
private computeElementHashes;
|
|
53
|
+
private computeCacheKey;
|
|
54
|
+
private generateCachePath;
|
|
55
|
+
private generateFreshCachePath;
|
|
56
|
+
private getIndexedCachePath;
|
|
57
|
+
private getEntryCachePath;
|
|
58
|
+
private toIndexCachePath;
|
|
59
|
+
private updateLastCache;
|
|
60
|
+
private clearLastCache;
|
|
61
|
+
private computeTokenPrefixHash;
|
|
62
|
+
private computePrefixInfo;
|
|
63
|
+
private findBestBase;
|
|
64
|
+
private addToIndex;
|
|
65
|
+
recordQuery(): void;
|
|
66
|
+
recordPromptTokens(newPromptTokens: number, cacheTokensUsed: number): void;
|
|
67
|
+
readCacheTokenCount(cachePath: string): number;
|
|
68
|
+
getStats(): {
|
|
69
|
+
totalQueries: number;
|
|
70
|
+
memoryHit: number;
|
|
71
|
+
diskHit: number;
|
|
72
|
+
incremental: number;
|
|
73
|
+
fresh: number;
|
|
74
|
+
totalPromptTokens: number;
|
|
75
|
+
prefillReusedTokens: number;
|
|
76
|
+
cacheGrowthTokens: number;
|
|
77
|
+
};
|
|
78
|
+
prepare(params: CachePrepareParams): Promise<CacheHandle>;
|
|
79
|
+
private isUnsupportedIncrementalError;
|
|
80
|
+
private createCache;
|
|
81
|
+
release(ref: string): void;
|
|
82
|
+
close(): Promise<void>;
|
|
83
|
+
}
|
|
84
|
+
//# sourceMappingURL=pytorch-cache-controller.d.ts.map
|
|
@@ -0,0 +1 @@
|
|
|
1
|
+
{"version":3,"file":"pytorch-cache-controller.d.ts","sourceRoot":"","sources":["../../src/pytorch/pytorch-cache-controller.ts"],"names":[],"mappings":"AAQA,OAAO,KAAK,EACV,WAAW,EACX,kBAAkB,EAClB,qBAAqB,EACtB,MAAM,wBAAwB,CAAC;AAEhC,OAAO,KAAK,EAAE,gBAAgB,EAAE,MAAM,uBAAuB,CAAC;AAE9D,OAAO,KAAK,EACV,gBAAgB,EAEjB,MAAM,gCAAgC,CAAC;AACxC,OAAO,KAAK,EAAE,cAAc,EAAE,MAAM,oBAAoB,CAAC;AA2BzD,MAAM,WAAW,6BAA6B;IAC5C,2CAA2C;IAC3C,QAAQ,CAAC,EAAE,MAAM,CAAC;CACnB;AAWD;;;;;;GAMG;AACH,qBAAa,sBAAuB,YAAW,qBAAqB;IAClE,OAAO,CAAC,WAAW,CAAkC;IACrD,OAAO,CAAC,gBAAgB,CAA2C;IACnE,OAAO,CAAC,OAAO,CAAC,CAAiB;IACjC,OAAO,CAAC,QAAQ,CAAS;IACzB,OAAO,CAAC,UAAU,CAAU;IAC5B,OAAO,CAAC,aAAa,CAAS;IAC9B,OAAO,CAAC,MAAM,CAAS;IACvB,OAAO,CAAC,KAAK,CAAS;IACtB,OAAO,CAAC,SAAS,CAAsB;IACvC,OAAO,CAAC,cAAc,CAAC,CAAa;IACpC,OAAO,CAAC,gBAAgB,CAAC,CAAuD;IAChF,OAAO,CAAC,gBAAgB,CAAwB;IAChD,OAAO,CAAC,UAAU,CAAC,CAAc;IACjC,OAAO,CAAC,iBAAiB,CAAC,CAAW;IACrC,OAAO,CAAC,eAAe,CAAC,CAAS;IACjC,OAAO,CAAC,8BAA8B,CAAC,CAAS;IAChD,OAAO,CAAC,mBAAmB,CAAC,CAAS;IACrC,OAAO,CAAC,yBAAyB,CAAC,CAAS;IAC3C,OAAO,CAAC,UAAU,CAA2C;IAC7D,mEAAmE;IACnE,OAAO,CAAC,gBAAgB,CAA6B;IACrD,OAAO,CAAC,KAAK,CAUX;IAEF,OAAO,CAAC,MAAM,CAAC,QAAQ,CAAC,YAAY,CAGlC;gBAEU,OAAO,CAAC,EAAE,6BAA6B;IAYnD,mEAAmE;IACnE,YAAY,CAAC,SAAS,CAAC,EAAE,IAAI,GAAG,KAAK,GAAG,IAAI;IAOtC,IAAI,CACR,OAAO,EAAE,cAAc,EACvB,gBAAgB,EAAE,gBAAgB,EAClC,gBAAgB,CAAC,EAAE,CAAC,QAAQ,EAAE,gBAAgB,EAAE,KAAK,gBAAgB,EAAE,GACtE,OAAO,CAAC,IAAI,CAAC;YA4BF,cAAc;IAM5B,OAAO,KAAK,SAAS,GAEpB;IAED,OAAO,CAAC,kBAAkB;IAY1B,OAAO,CAAC,cAAc;YAmBR,SAAS;YAkBT,SAAS;IAqBvB,OAAO,CAAC,2BAA2B;IAKnC,OAAO,CAAC,gBAAgB;IAMxB,OAAO,CAAC,oBAAoB;IAW5B,OAAO,CAAC,eAAe;IAoBvB,OAAO,CAAC,iBAAiB;IAIzB,OAAO,CAAC,sBAAsB;IAiB9B,OAAO,CAAC,mBAAmB;IAY3B,OAAO,CAAC,iBAAiB;IAQzB,OAAO,CAAC,gBAAgB;IAUxB,OAAO,CAAC,eAAe;IAavB,OAAO,CAAC,cAAc;IAStB,OAAO,CAAC,sBAAsB;YAQhB,iBAAiB;YA2EjB,YAAY;IAsI1B,OAAO,CAAC,UAAU;IA2BlB,WAAW,IAAI,IAAI;IAInB,kBAAkB,CAAC,eAAe,EAAE,MAAM,EAAE,eAAe,EAAE,MAAM,GAAG,IAAI;IAK1E,mBAAmB,CAAC,SAAS,EAAE,MAAM,GAAG,MAAM;IAI9C,QAAQ;;;;;;;;;;IAcF,OAAO,CAAC,MAAM,EAAE,kBAAkB,GAAG,OAAO,CAAC,WAAW,CAAC;IAuC/D,OAAO,CAAC,6BAA6B;YAKvB,WAAW;IA6LzB,OAAO,CAAC,GAAG,EAAE,MAAM,GAAG,IAAI;IAapB,KAAK,IAAI,OAAO,CAAC,IAAI,CAAC;CAqC7B"}
|