repo_name stringlengths 1 62 | dataset stringclasses 1 value | lang stringclasses 11 values | pr_id int64 1 20.1k | owner stringlengths 2 34 | reviewer stringlengths 2 39 | diff_hunk stringlengths 15 262k | code_review_comment stringlengths 1 99.6k |
|---|---|---|---|---|---|---|---|
openrouter-runner | github_2023 | python | 68 | OpenRouterTeam | alexanderatallah | @@ -70,70 +82,56 @@ def __init__(self, params: VllmParams):
async def generate(self, payload: CompletionPayload, params):
assert self.engine is not None, "Engine not initialized"
+ # Track usage as a running total
+ # NOTE: This does NOT yet include cold-start GPU time | how hard is this to add @sambarnes ? |
openrouter-runner | github_2023 | python | 68 | OpenRouterTeam | louisgv | @@ -95,45 +100,29 @@ class Usage(BaseModel):
prompt_tokens: int
completion_tokens: int
- # TODO: add these in after deprecating the old fields
- # duration: float
- # gpu_type: GPUType
- # gpu_count: int
+ duration: float
+ gpu_type: GPUType
+ gpu_count: int
class ResponseBody(BaseModel):
text: str
- prompt_tokens: int # TODO: deprecate
- completion_tokens: int # TODO: deprecate
usage: Usage
done: bool
def create_response_text(
text: str,
- prompt_tokens: int = 0,
- completion_tokens: int = 0,
+ usage: Usage,
done: bool = False,
) -> str:
- return ResponseBody(
- text=text,
- prompt_tokens=prompt_tokens,
- completion_tokens=completion_tokens,
- usage=Usage(
- prompt_tokens=prompt_tokens,
- completion_tokens=completion_tokens,
- ),
- done=done,
- ).json(ensure_ascii=False)
+ r = ResponseBody(text=text, usage=usage, done=done)
+ return r.json(ensure_ascii=False) | The main reason we used this was due to the pydantic version shipped with vllm - it's deprecated in latest (due to modal installation), but not if we have the vllm's pydantic typing xd.....
I wanted to do this way back in September when I saw the red deprecation strike! - Let's try it out on dev and see if it work! |
openrouter-runner | github_2023 | python | 68 | OpenRouterTeam | louisgv | @@ -70,84 +84,82 @@ def __init__(self, params: VllmParams):
async def generate(self, payload: CompletionPayload, params):
assert self.engine is not None, "Engine not initialized"
- t_start_inference = time.perf_counter()
+ # Track usage as a running total. For the first request to the
+ # container, cold-start time is included in the usage duration.
+ t_start_inference = time.time()
+ t_start_usage_duration = t_start_inference
+ if self.is_first_request:
+ self.is_first_request = False
+ t_start_usage_duration = self.t_cold_start
+
+ resp = ResponseBody(
+ text="",
+ usage=Usage(
+ prompt_tokens=0,
+ completion_tokens=0,
+ duration=0.0,
+ gpu_type=self.gpu_type,
+ gpu_count=self.gpu_count,
+ ),
+ )
try:
results_generator = self.engine.generate(
payload.prompt, params, payload.id
)
- final_output = None
+ output = ""
+ index = 0
finish_reason = None
- if payload.stream:
- index = 0
-
- async for request_output in results_generator:
- # Skipping invalid UTF8 tokens:
- if (
- request_output.outputs[0].text
- and request_output.outputs[0].text[-1] == "\ufffd"
- ):
- continue
-
- final_output = request_output
- token = final_output.outputs[0].text[index:]
- index = len(final_output.outputs[0].text)
- finish_reason = final_output.outputs[0].finish_reason
- yield create_sse_data(
- token,
- prompt_tokens=len(final_output.prompt_token_ids),
- completion_tokens=len(
- final_output.outputs[0].token_ids
- ),
- done=False,
- finish_reason=finish_reason,
- )
-
- output = ""
- else:
- async for request_output in results_generator:
- final_output = request_output
- yield " "
-
- output = final_output.outputs[0].text
- finish_reason = final_output.outputs[0].finish_reason
-
- prompt_tokens = len(final_output.prompt_token_ids)
- completion_tokens = len(final_output.outputs[0].token_ids)
- if payload.stream:
- yield create_sse_data(
- "",
- prompt_tokens,
- completion_tokens,
- done=True,
- finish_reason=finish_reason,
- )
- else:
- yield create_response_text(
- output,
- prompt_tokens,
- completion_tokens,
- done=True,
- finish_reason=finish_reason,
- )
-
- elapsed = time.perf_counter() - t_start_inference
+ async for current in results_generator:
+ output = current.outputs[0].text
+ finish_reason = current.outputs[0].finish_reason
+ resp.usage.prompt_tokens = len(current.prompt_token_ids)
+ resp.usage.completion_tokens = len(current.outputs[0].token_ids)
+ resp.usage.duration = time.time() - t_start_usage_duration
+
+ # Non-streaming requests continue generating w/o yielding intermediate results
+ if not payload.stream:
+ yield " " # QUESTION[sam]: why yield a single space for non-streaming?
+ continue | @sambarnes we need to yield "something" to keep the connection going, based on an experiment I did some time ago with fetch + SSE. Basically a hack to keep the connection going since this is a one-way pull, not a bi-directional connection like websocket where there's a dedicated handshake before transmission.
It's kinda like... UDP? (but on TCP lol) |
openrouter-runner | github_2023 | python | 68 | OpenRouterTeam | louisgv | @@ -70,84 +84,82 @@ def __init__(self, params: VllmParams):
async def generate(self, payload: CompletionPayload, params):
assert self.engine is not None, "Engine not initialized"
- t_start_inference = time.perf_counter()
+ # Track usage as a running total. For the first request to the
+ # container, cold-start time is included in the usage duration.
+ t_start_inference = time.time()
+ t_start_usage_duration = t_start_inference
+ if self.is_first_request:
+ self.is_first_request = False
+ t_start_usage_duration = self.t_cold_start
+
+ resp = ResponseBody(
+ text="",
+ usage=Usage(
+ prompt_tokens=0,
+ completion_tokens=0,
+ duration=0.0,
+ gpu_type=self.gpu_type,
+ gpu_count=self.gpu_count,
+ ),
+ )
try:
results_generator = self.engine.generate(
payload.prompt, params, payload.id
)
- final_output = None
+ output = ""
+ index = 0
finish_reason = None
- if payload.stream:
- index = 0
-
- async for request_output in results_generator:
- # Skipping invalid UTF8 tokens:
- if (
- request_output.outputs[0].text
- and request_output.outputs[0].text[-1] == "\ufffd"
- ):
- continue
-
- final_output = request_output
- token = final_output.outputs[0].text[index:]
- index = len(final_output.outputs[0].text)
- finish_reason = final_output.outputs[0].finish_reason
- yield create_sse_data(
- token,
- prompt_tokens=len(final_output.prompt_token_ids),
- completion_tokens=len(
- final_output.outputs[0].token_ids
- ),
- done=False,
- finish_reason=finish_reason,
- )
-
- output = ""
- else:
- async for request_output in results_generator:
- final_output = request_output
- yield " "
-
- output = final_output.outputs[0].text
- finish_reason = final_output.outputs[0].finish_reason
-
- prompt_tokens = len(final_output.prompt_token_ids)
- completion_tokens = len(final_output.outputs[0].token_ids)
- if payload.stream:
- yield create_sse_data(
- "",
- prompt_tokens,
- completion_tokens,
- done=True,
- finish_reason=finish_reason,
- )
- else:
- yield create_response_text(
- output,
- prompt_tokens,
- completion_tokens,
- done=True,
- finish_reason=finish_reason,
- )
-
- elapsed = time.perf_counter() - t_start_inference
+ async for current in results_generator:
+ output = current.outputs[0].text
+ finish_reason = current.outputs[0].finish_reason
+ resp.usage.prompt_tokens = len(current.prompt_token_ids)
+ resp.usage.completion_tokens = len(current.outputs[0].token_ids)
+ resp.usage.duration = time.time() - t_start_usage_duration
+
+ # Non-streaming requests continue generating w/o yielding intermediate results
+ if not payload.stream:
+ yield " " # HACK: Keep the connection alive while generating
+ continue
+
+ # Skipping invalid UTF8 tokens:
+ if output and output[-1] == "\ufffd":
+ continue
+
+ # Streaming requests send SSE messages with each new generated part
+ token = output[index:]
+ index = len(output)
+ resp.text = token
+ resp.finish_reason = finish_reason
+ yield sse(resp.json(ensure_ascii=False))
+
+ resp.text = "" if payload.stream else output
+ resp.done = True
+ resp.finish_reason = finish_reason
+ data = resp.json(ensure_ascii=False)
+ yield sse(data) if payload.stream else data
+
logger.info(
"Completed generation",
extra={
"model": self.engine_args.model,
- "tokens": completion_tokens,
- "tps": completion_tokens / elapsed,
- "duration": elapsed,
+ "tokens": resp.usage.completion_tokens,
+ "tps": resp.usage.completion_tokens / t_start_inference,
+ "duration": resp.usage.duration,
},
)
except Exception as err:
e = create_error_text(err)
logger.exception(
"Failed generation", extra={"model": self.engine_args.model}
)
- if payload.stream:
- yield create_sse_data(e)
- else:
- yield e
+ # TODO: Existing code prior to refactor was yielding error text
+ # in a ResponseBody model, with empty usage data. Non-streaming
+ # requests yielded only error text as a string. This behavior
+ # is preserved here, but should maybe be revisited.
+ resp.text = e
+ resp.usage.prompt_tokens = 0
+ resp.usage.completion_tokens = 0 | Feel like we don't have to reset it here. The exception could have been thrown mid generation :-? |
openrouter-runner | github_2023 | python | 39 | OpenRouterTeam | louisgv | @@ -29,3 +29,12 @@ async def test_completion_streaming():
output = await completion(prompt, stream=True)
assert output is not None
+
+
+@pytest.mark.integration
+@pytest.mark.asyncio
+async def test_large_prompt():
+ """Should accept & complete a large prompt"""
+ prompt = get_words_from_file(n=2500)
+ output = await completion(prompt, max_tokens=5000)
+ assert output is not None | Ooh interesting - how else and what other aspect of an LLM endpoint would you test from an e2e? |
openrouter-runner | github_2023 | python | 64 | OpenRouterTeam | alexanderatallah | @@ -85,10 +85,28 @@ class CompletionPayload(BaseModel):
runner: RunnerConfiguration | None = None
-class ResponseBody(BaseModel):
- text: str
+class Usage(BaseModel):
+ """
+ An OpenAI-style usage struct, containing the expected
+ token counts as well as additional data about the GPU
+ usage of the request.
+ """
+
prompt_tokens: int
completion_tokens: int
+ total_tokens: int
+
+ # TODO: add these in after deprecating the old fields
+ # duration: float
+ # gpu_type: GPUType
+ # gpu_count: int | yeah it seems redundant to me. this repo doesn't follow their spec anyway, so I think we could omit it |
openrouter-runner | github_2023 | python | 64 | OpenRouterTeam | alexanderatallah | @@ -87,22 +88,29 @@ async def generate(self, payload: CompletionPayload, params):
and request_output.outputs[0].text[-1] == "\ufffd"
):
continue
- token = request_output.outputs[0].text[index:]
- index = len(request_output.outputs[0].text)
- yield create_sse_data(token)
+
+ final_output = request_output
+ token = final_output.outputs[0].text[index:]
+ index = len(final_output.outputs[0].text)
+ yield create_sse_data(
+ token,
+ prompt_tokens=len(final_output.prompt_token_ids),
+ completion_tokens=len(
+ final_output.outputs[0].token_ids
+ ),
+ done=False,
+ ) | OpenAI doesn't do any accounting for streaming :(
We should, by sending the running token counts along with the running total time |
openrouter-runner | github_2023 | python | 61 | OpenRouterTeam | sambarnes | @@ -15,11 +15,20 @@
def completion(
+ request: Request,
payload: CompletionPayload,
):
model_path = get_model_path(payload.model)
logger.info(
- "Received completion request", extra={"model": str(model_path)}
+ "Received completion request",
+ extra={ | IIRC at opensea we did something that set this context globally for a request so all logs in the call stack used it 🤔 but that's just something in the back of my mind, we can do it later if necessary |
openrouter-runner | github_2023 | python | 61 | OpenRouterTeam | sambarnes | @@ -15,11 +15,20 @@
def completion(
+ request: Request,
payload: CompletionPayload,
):
model_path = get_model_path(payload.model)
logger.info(
- "Received completion request", extra={"model": str(model_path)}
+ "Received completion request",
+ extra={
+ "model": str(model_path),
+ "user-agent": request.headers.get("user-agent"), | I forget if this key needs to be snake_case for datadog to pick it up |
openrouter-runner | github_2023 | python | 60 | OpenRouterTeam | alexanderatallah | @@ -1,9 +1,12 @@
from enum import Enum
-from typing import List, Optional, Union
+from typing import Final, List, Optional, Union
from fastapi.responses import JSONResponse, PlainTextResponse
from pydantic import BaseModel
+A100_40G: Final[float] = 0.001036
+A100_80G: Final[float] = 0.001553 |
can we name this variable based on the units? ($/sec?) |
openrouter-runner | github_2023 | python | 53 | OpenRouterTeam | alexanderatallah | @@ -77,11 +77,9 @@ async def generate(self, payload: CompletionPayload, params):
try:
import time
- tags = {"model": self.engine_args.model}
- with timer("engine.generate()", tags=tags): | we're removing this bc it was instant right? |
openrouter-runner | github_2023 | python | 49 | OpenRouterTeam | montasaurus | @@ -31,6 +33,17 @@ def add_observability(image: Image):
)
+@contextmanager
+def timer(action: str, tags: dict[str, str | int] = None) -> None:
+ """A simple timer context manager with structured logging for its output."""
+ start = time.perf_counter()
+ yield
+ elapsed = time.perf_counter() - start
+
+ extra = tags or {} | {"duration": elapsed} | ```suggestion
extra = (tags or {}) | {"duration": elapsed}
```
to fix the precedence |
openrouter-runner | github_2023 | others | 51 | OpenRouterTeam | louisgv | @@ -0,0 +1,39 @@
+name: Release (prod)
+
+on:
+ push:
+ branches: [main]
+ paths: ['modal/**'] | We should add the poetry lock file as well |
openrouter-runner | github_2023 | others | 51 | OpenRouterTeam | louisgv | @@ -0,0 +1,39 @@
+name: Release (prod)
+
+on:
+ push:
+ branches: [main]
+ paths: ['modal/**']
+
+jobs:
+ release:
+ name: Release (prod)
+ runs-on: ubuntu-latest
+ env:
+ MODAL_TOKEN_ID: ${{ secrets.MODAL_TOKEN_ID }}
+ MODAL_TOKEN_SECRET: ${{ secrets.MODAL_TOKEN_SECRET }} | Done! cc @alexanderatallah to config @sambarnes 's perm |
openrouter-runner | github_2023 | python | 48 | OpenRouterTeam | sambarnes | @@ -53,7 +58,7 @@ def download_model(model_name: str):
ignore_patterns.append("*.bin")
# Clean doesn't remove the cache, so using `local_files_only` here returns the cache even when the local dir is empty.
- print(f"Checking for {model_name}")
+ logger.info(f"Checking for {model_name}") | in the future we might wanna move more towards structured logging with some of these
that way we can e.g. group logs by model in DD
`logger.info("Checking for model", extra=dict(model=model_name)"`
def not a right now thing, but worth considering going forward
|
openrouter-runner | github_2023 | python | 48 | OpenRouterTeam | sambarnes | @@ -0,0 +1,107 @@
+import json
+import logging
+import os
+import sys
+
+from datadog_api_client import ApiClient, Configuration
+from datadog_api_client.v2.api.logs_api import LogsApi
+from datadog_api_client.v2.model.content_encoding import ContentEncoding
+from datadog_api_client.v2.model.http_log import HTTPLog
+from datadog_api_client.v2.model.http_log_item import HTTPLogItem
+
+
+def get_logger(name: str):
+ return logging.getLogger(name)
+
+
+# Define a custom logging handler that sends logs to Datadog
+class DatadogHandler(logging.Handler):
+ def __init__(self, api_client: ApiClient):
+ self.api_client = api_client
+
+ super().__init__()
+
+ def emit(self, record):
+ # Ignore debug messages
+ if record.levelno == logging.DEBUG:
+ return
+
+ environment_name = os.getenv("DD_ENV") or "development"
+
+ log_payload = {
+ "python-logging": {
+ "py-env": "development",
+ "py-message": record.getMessage(),
+ "py-status": record.levelname.lower(),
+ "py-logger": record.name,
+ "py-stacktrace": record.exc_info,
+ "py-exception": record.exc_text,
+ "py-line": record.lineno,
+ "py-file": record.filename,
+ "py-function": record.funcName,
+ "py-level": record.levelno,
+ "py-path": record.pathname,
+ "py-thread": record.thread,
+ "py-threadName": record.threadName,
+ "py-process": record.process,
+ "py-processName": record.processName,
+ "py-args": record.args,
+ "py-msecs": record.msecs,
+ "py-relativeCreated": record.relativeCreated,
+ "py-created": record.created,
+ "py-module": record.module,
+ },
+ "message": record.getMessage(),
+ "environment": environment_name,
+ "source": record.name,
+ "model": record.__dict__.get("model"),
+ "modal": {
+ "cloudProvider": os.getenv("MODAL_CLOUD_PROVIDER") or "",
+ "environment": os.getenv("MODAL_ENVIRONMENT") or "",
+ "imageId": os.getenv("MODAL_IMAGE_ID") or "",
+ "region": os.getenv("MODAL_REGION") or "",
+ "taskId": os.getenv("MODAL_TASK_ID") or "",
+ },
+ }
+
+ # Send the log to Datadog using the Logs API
+ try:
+ body = HTTPLog(
+ [
+ HTTPLogItem(
+ ddsource="Python",
+ ddtags=f"env:{environment_name}",
+ hostname=os.getenv("MODAL_TASK_ID") or "",
+ message=json.dumps(log_payload),
+ service="openrouter-runner",
+ status=record.levelname.lower(),
+ ),
+ ]
+ )
+
+ logs = LogsApi(api_client)
+
+ logs.submit_log(
+ content_encoding=ContentEncoding.DEFLATE,
+ body=body, # type: ignore
+ )
+
+ except Exception as e:
+ print(f"Error sending log to Datadog: {e}")
+
+
+if "DD_API_KEY" in os.environ: | given the new docs say:
> or leave it blank to disable Datadog (e.g. `DD_API_KEY=`
do you know if `DD_API_KEY` will still be on the container as empty string if ppl set it like that?
if so, would probs need to be
```suggestion
if os.environ.get("DD_API_KEY") is not None:
``` |
openrouter-runner | github_2023 | python | 48 | OpenRouterTeam | sambarnes | @@ -0,0 +1,107 @@
+import json
+import logging
+import os
+import sys
+
+from datadog_api_client import ApiClient, Configuration
+from datadog_api_client.v2.api.logs_api import LogsApi
+from datadog_api_client.v2.model.content_encoding import ContentEncoding
+from datadog_api_client.v2.model.http_log import HTTPLog
+from datadog_api_client.v2.model.http_log_item import HTTPLogItem
+
+
+def get_logger(name: str):
+ return logging.getLogger(name)
+
+
+# Define a custom logging handler that sends logs to Datadog
+class DatadogHandler(logging.Handler):
+ def __init__(self, api_client: ApiClient):
+ self.api_client = api_client
+
+ super().__init__()
+
+ def emit(self, record):
+ # Ignore debug messages
+ if record.levelno == logging.DEBUG:
+ return
+
+ environment_name = os.getenv("DD_ENV") or "development"
+
+ log_payload = {
+ "python-logging": {
+ "py-env": "development",
+ "py-message": record.getMessage(),
+ "py-status": record.levelname.lower(),
+ "py-logger": record.name,
+ "py-stacktrace": record.exc_info,
+ "py-exception": record.exc_text,
+ "py-line": record.lineno,
+ "py-file": record.filename,
+ "py-function": record.funcName,
+ "py-level": record.levelno,
+ "py-path": record.pathname,
+ "py-thread": record.thread,
+ "py-threadName": record.threadName,
+ "py-process": record.process,
+ "py-processName": record.processName,
+ "py-args": record.args,
+ "py-msecs": record.msecs,
+ "py-relativeCreated": record.relativeCreated,
+ "py-created": record.created,
+ "py-module": record.module,
+ },
+ "message": record.getMessage(),
+ "environment": environment_name,
+ "source": record.name,
+ "model": record.__dict__.get("model"),
+ "modal": {
+ "cloudProvider": os.getenv("MODAL_CLOUD_PROVIDER") or "",
+ "environment": os.getenv("MODAL_ENVIRONMENT") or "",
+ "imageId": os.getenv("MODAL_IMAGE_ID") or "",
+ "region": os.getenv("MODAL_REGION") or "",
+ "taskId": os.getenv("MODAL_TASK_ID") or "",
+ },
+ }
+
+ # Send the log to Datadog using the Logs API
+ try:
+ body = HTTPLog(
+ [
+ HTTPLogItem(
+ ddsource="Python",
+ ddtags=f"env:{environment_name}",
+ hostname=os.getenv("MODAL_TASK_ID") or "",
+ message=json.dumps(log_payload),
+ service="openrouter-runner",
+ status=record.levelname.lower(),
+ ),
+ ]
+ )
+
+ logs = LogsApi(api_client)
+
+ logs.submit_log(
+ content_encoding=ContentEncoding.DEFLATE,
+ body=body, # type: ignore
+ )
+
+ except Exception as e:
+ print(f"Error sending log to Datadog: {e}") | think we should still do a bare `logging.exception(` in order for the sentry hook to fire & notify us there?
```suggestion
logging.exception("Error sending log to Datadog")
``` |
openrouter-runner | github_2023 | python | 48 | OpenRouterTeam | sambarnes | @@ -0,0 +1,8 @@
+from modal import Image
+
+
+def add_datadog(image: Image):
+ return image.pip_install("datadog-api-client==2.21.0") | mind expanding one why this helper would have api & not ddtrace in it?
(just mentioning cause i see ddtrace manually on the download.py image) |
openrouter-runner | github_2023 | python | 48 | OpenRouterTeam | sambarnes | @@ -0,0 +1,65 @@
+from fastapi import Depends, FastAPI, Request
+from fastapi.responses import JSONResponse
+from fastapi.security import HTTPAuthorizationCredentials
+from pydantic import BaseModel
+
+from runner.endpoints.completion import completion as completion_endpoint
+from runner.shared.common import config
+from runner.shared.download import download_model
+from shared.logging import get_logger
+from shared.protocol import CompletionPayload
+
+api_app = FastAPI() | i like this reorganization :+1: |
openrouter-runner | github_2023 | python | 48 | OpenRouterTeam | louisgv | @@ -11,13 +13,16 @@
models_volume,
)
+logger = get_logger(__name__) | Does this work at the top-level scope? Does the stub's env injection take effect here? |
openrouter-runner | github_2023 | python | 48 | OpenRouterTeam | louisgv | @@ -11,13 +13,16 @@
models_volume,
)
+logger = get_logger(__name__)
+
cache_path = get_model_path("__cache__")
downloader_image = (
- Image.debian_slim()
+ BASE_IMAGE
# Use the barebones hf-transfer package for maximum download speeds. No progress bar, but expect 700MB/s.
.pip_install("huggingface_hub==0.19.4")
.pip_install("hf-transfer==0.1.4")
+ .pip_install("ddtrace==2.4.0")
.env({"HF_HUB_ENABLE_HF_TRANSFER": "1"}) | This would install on top of the `BASE_IMAGE` instance which is not what we want I think. We should make BASE_IMAGE into a function that return a new Image() instance instead. |
openrouter-runner | github_2023 | python | 48 | OpenRouterTeam | louisgv | @@ -0,0 +1,107 @@
+import json
+import logging
+import os
+import sys
+
+from datadog_api_client import ApiClient, Configuration
+from datadog_api_client.v2.api.logs_api import LogsApi
+from datadog_api_client.v2.model.content_encoding import ContentEncoding
+from datadog_api_client.v2.model.http_log import HTTPLog
+from datadog_api_client.v2.model.http_log_item import HTTPLogItem
+
+
+def get_logger(name: str):
+ return logging.getLogger(name)
+
+
+# Define a custom logging handler that sends logs to Datadog
+class DatadogHandler(logging.Handler):
+ def __init__(self, api_client: ApiClient):
+ self.api_client = api_client
+
+ super().__init__()
+
+ def emit(self, record):
+ # Ignore debug messages
+ if record.levelno == logging.DEBUG:
+ return
+
+ environment_name = os.getenv("DD_ENV") or "development"
+
+ log_payload = {
+ "python-logging": {
+ "py-env": "development", | Is this the case even for prod instance? |
openrouter-runner | github_2023 | python | 48 | OpenRouterTeam | alexanderatallah | @@ -0,0 +1,80 @@
+import modal
+from fastapi import Depends, FastAPI, Request
+from fastapi.responses import JSONResponse
+from fastapi.security import HTTPAuthorizationCredentials
+from pydantic import BaseModel
+
+from runner.endpoints.completion import completion as completion_endpoint
+from runner.shared.common import config
+from runner.shared.download import download_model
+from shared.logging import get_logger
+from shared.protocol import CompletionPayload
+
+api_app = FastAPI()
+logger = get_logger(__name__)
+
+
+@api_app.middleware("http")
+async def log_errors(request: Request, call_next): | when is this called? |
openrouter-runner | github_2023 | python | 47 | OpenRouterTeam | montasaurus | @@ -116,7 +117,9 @@ async def generate(self, payload: CompletionPayload, params):
print(f"Request completed: {throughput:.4f} tokens/s")
except Exception as err:
e = create_error_text(err)
- print(e)
+ logging.exception(
+ "Failed generation", extra={"model": self.engine_args.model}
+ ) | I hit the same thing, `logging.basicConfig(level=logging.INFO)` fixed it |
openrouter-runner | github_2023 | python | 46 | OpenRouterTeam | montasaurus | @@ -70,7 +72,13 @@ async def get_job(
@stub.function(
- secret=Secret.from_name("ext-api-key"),
+ image=Image.debian_slim(python_version="3.10").pip_install(
+ "sentry-sdk[fastapi]==1.39.1"
+ ), | Got exactly that in a PR I'm working on! |
openrouter-runner | github_2023 | python | 43 | OpenRouterTeam | louisgv | @@ -10,7 +10,9 @@ def get_model_path(model_name: str):
return models_path / model_name.lower()
-def does_model_exist(model_path: Path):
+# gets passed a str in some circumstances despite being typed as Path
+def does_model_exist(model_path: Path | str):
+ model_path = Path(model_path) | Might need to revert this (and the model_path passing down to our container) to just string. When we instantiate a GPU container, the props passed down will be serialized by the Modal abstraction, using pickle, and afaik it will not be deserialized on our code properly.
3 options:
1. Ping Modal about better type-casting, to which I think their best attempt would be a pydantic pipeline, since from their POV this param has arbitrary type.
2. Make a serialize/deserialize pipeline, which might not work with modal parameterized req
3. Stick to passing down primitive types (revert to string) |
openrouter-runner | github_2023 | typescript | 41 | OpenRouterTeam | louisgv | @@ -0,0 +1,35 @@
+import {
+ awaitJob,
+ enqueueAddModel,
+ getApiUrl,
+ getAuthHeaders,
+ runIfCalledAsScript
+} from 'scripts/shared';
+
+async function main(model?: string) {
+ const modelName = model || process.env.MODEL!;
+ console.log(`Test adding model ${modelName}`);
+ const body = await enqueueAddModel(modelName);
+ console.log('Successfully queued model to add', body);
+
+ // fetch job success or failure for up to the timeout of 1 hour
+ const timeoutMs = 60 * 60 * 1000;
+ await awaitJob(body.job_id, timeoutMs); | nit: Can we think of an alternative name that doesn't repeat the `await`? |
openrouter-runner | github_2023 | python | 41 | OpenRouterTeam | louisgv | @@ -24,7 +25,7 @@ def _make_container(
class _VllmContainer(VllmEngine):
def __init__(
self,
- model_path: str,
+ model_path: Path, | Interesting - so it's possible to use `Path` as a param such that only 1 instance will be spawned to serve a certain path until the max concurrent input is reached? |
openrouter-runner | github_2023 | typescript | 41 | OpenRouterTeam | alexanderatallah | @@ -0,0 +1,34 @@
+import {
+ completion,
+ enqueueAddModel,
+ pollForJobCompletion,
+ runIfCalledAsScript
+} from 'scripts/shared';
+
+async function main() {
+ const modelName = 'microsoft/Orca-2-13b'; | can we do phi-2? (if this test is still helpful for small models)
it has a better license |
openrouter-runner | github_2023 | python | 41 | OpenRouterTeam | alexanderatallah | @@ -24,12 +37,17 @@ class Params(BaseModel):
skip_special_tokens: bool = True
-class Payload(BaseModel):
+class RunnerConfiguration(BaseModel):
+ container: ContainerType
+
+
+class CompletionPayload(BaseModel):
id: str
prompt: str
stream: bool = False
params: Params
model: str
+ runner: RunnerConfiguration | None = None | woah, new union notation |
openrouter-runner | github_2023 | typescript | 41 | OpenRouterTeam | louisgv | @@ -6,9 +6,21 @@ const envFile = `.env.dev`;
config({ path: envFile });
-const url = process.env.API_URL;
-const key = process.env.RUNNER_API_KEY;
+const url = process.env.API_URL!;
+const key = process.env.RUNNER_API_KEY!; | We should assert these instead of using `!` since new devs might not have them. We lean toward strict typing with assertion as much as possible. |
openrouter-runner | github_2023 | typescript | 41 | OpenRouterTeam | louisgv | @@ -0,0 +1,35 @@
+import {
+ enqueueAddModel,
+ getApiUrl,
+ getAuthHeaders,
+ pollForJobCompletion,
+ runIfCalledAsScript
+} from 'scripts/shared';
+
+async function main(model?: string) {
+ const modelName = model || process.env.MODEL!;
+ console.log(`Test adding model ${modelName}`);
+ const body = await enqueueAddModel(modelName); | nit: we should prob reuse the variable from shared |
openrouter-runner | github_2023 | python | 33 | OpenRouterTeam | louisgv | @@ -22,6 +22,13 @@ def download_models(all_models: List[str]):
cache_path = get_model_path("__cache__")
for model_name in all_models:
+ parts = model_name.split(":")
+
+ model_revision = None | This only downloads the model - can we also added it somehow into the protocol's input params? |
openrouter-runner | github_2023 | python | 36 | OpenRouterTeam | alexanderatallah | @@ -28,27 +27,39 @@ def __init__(
model_path: str,
max_model_len: Optional[int] = None,
):
- if num_gpus > 1:
- # Patch issue from https://github.com/vllm-project/vllm/issues/1116
- import ray
-
- ray.shutdown()
- ray.init(num_gpus=num_gpus, ignore_reinit_error=True)
-
- super().__init__(
- VllmParams(
- model=model_path,
- tensor_parallel_size=num_gpus,
- max_model_len=max_model_len,
- )
+ import sentry_sdk
+
+ sentry_sdk.init( | we can't init sentry higher up, for other Modal processes too? |
openrouter-runner | github_2023 | others | 32 | OpenRouterTeam | louisgv | @@ -48,3 +48,11 @@ ignore = [
[tool.ruff.per-file-ignores]
"gcp/shap-e/app.py" = ["E402"] # module level import not at top of file -- expected for sys.path.insert() usage
+[tool.ruff.lint.isort]
+known-third-party = ["modal"]
+known-first-party = [
+ "punctuator",
+ "runner",
+ "shared",
+ "tuner",
+] | I wonder if there's a better way to structure our project? Can ruff target or isort has sub projects? |
openrouter-runner | github_2023 | python | 31 | OpenRouterTeam | louisgv | @@ -0,0 +1,29 @@
+import os
+from typing import Annotated
+
+from fastapi import Depends, FastAPI, testclient
+from fastapi.security import HTTPAuthorizationCredentials
+
+from shared.config import Config
+
+
+def test_auth():
+ """The API auth dependency should prevent unauthorized requests.""" | @sambarnes I think the CI PR targeted the wrong remote (or wrong github org) |
openrouter-runner | github_2023 | python | 31 | OpenRouterTeam | louisgv | @@ -75,16 +72,12 @@ def transform(self, input_str: str):
import threading
import time
- output = [
- None
- ] # Use a list to hold the output to bypass Python's scoping limitations | Ooh - I also forgot about the line-length. Kinda want to keep it at 80 since I split my screen a lot :d |
openrouter-runner | github_2023 | others | 31 | OpenRouterTeam | louisgv | @@ -16,11 +16,13 @@ accelerate = "^0.23.0"
datasets = "^2.14.5"
scipy = "^1.11.3"
wandb = "^0.15.12"
+httpx = "^0.26.0" | seems like we don't need this anymore? Based on your comment, I think this was needed for the e2e test.
For e2e, I wonder if we can do a `modal deply runner` to a `e2e` workspace, tail the modal URL, then call it with our JS scripts?
Also we should prob add a js script to test auth as well :-? |
openrouter-runner | github_2023 | typescript | 31 | OpenRouterTeam | louisgv | @@ -15,6 +15,17 @@ async function main(model?: string) {
max_tokens: 1024,
stop: ['</s>'],
});
+
+ // Unauthorized requests should fail with a 401
+ let gotExpectedError = false;
+ try {
+ await completion(prompt, {model, apiKey: "BADKEY"});
+ } catch (e: any) {
+ gotExpectedError = e.message == "Status: 401";
+ }
+ if (!gotExpectedError) {
+ throw new Error("Unauthorized request returned unexpected response")
+ } | nit: we can just re-throw the error in the try if not 401, no need to do it separately IMO. Also, we should either use `startsWith` or strict equal. |
openrouter-runner | github_2023 | others | 28 | OpenRouterTeam | louisgv | @@ -1,46 +1,260 @@
-# PREREQ:
+# OpenRouter Runner - Setup Guide
-1. Create a modal secret group
- `HUGGINGFACE_TOKEN = <your huggingface token>`
- with name "huggingface"
-2. Create a modal secret group
- `RUNNER_API_KEY = <generate a random key>`
- with name "ext-api-key"
-3. Make sure your current directory, `pwd`, is `/modal`
-4. `modal config set-environment dev # or main`
+Welcome to the setup guide for the OpenRouter Runner. In this guide, you'll achieve the following:
-## Adding new container
+- **Configure Your Environment**: Set up necessary tools and accounts.
+- **Access Tokens**: Create and manage access tokens for secure interactions.
+- **Prepare Containers and Engines**: Get your system ready to run a variety of models.
+- **Deploy Models**: Learn how to add and deploy new models to the Runner.
+- **Test and Monitor**: Understand how to test your deployments and monitor performance.
-1. Copy one of the container file in `runner/containers`
-2. Adapt the class name, image, machine type, engine, etc...
-3. Add the container to [](./containers/__init__.py)
-4. Add models to be run with the containers in a dedicated list
-5. Run: `modal run runner::download`
+By the end of this guide, you'll have a comprehensive setup that allows you to manage and interact with advanced AI engines and access your models via API endpoints.
-## Adding new model
+## Prerequisites
-1. Add the model HF ID to [](./containers/__init__.py) under a list
-2. Run: `modal run runner::download`
+Before you begin, ensure you have the necessary accounts and tools:
-## Configuration
+1. **Modal Account**: Set up your environment on [Modal](https://modal.com/) as this will be your primary deployment platform.
+2. **Hugging Face Account**: Obtain a token from [Hugging Face](https://huggingface.co/) for accessing models and libraries.
-- For stub config, see: [](./shared/common.py)
-- For containers config, see: [](./containers/__init__.py)
-- For endpoint configs, see: [](./__init__.py)
+### Environment and Secrets Setup
-## Testing a model
+Before you start working with the OpenRouter Runner, it's crucial to set up your working environment and secure access tokens. Follow these steps to ensure everything is properly configured:
-1. `cd ..`
-2. Load your environment, e.g. `source .env.dev`
-3. Run: `pnpm x scripts/test-simple.ts YourModel/Identifier`
+1. **Set Current Directory**:
+ First, make sure your current working directory is the `/modal` directory where your project is located. Adjust the path as needed for your setup.
+ ```shell
+ cd path/to/modal
+ ```
-Other tests are available in `scripts`:
+2. **Create Development Environment**:
+ If you haven't already set up a development environment in Modal, now is the time to do so. This provides a separate space where you can develop and test without affecting production.
-```shell
-pnpm x scripts/test-${testname}.ts Gryphe/MythoMax-L2-13b # For dev environment
-pnpm x scripts/test-${testname}.ts Gryphe/MythoMax-L2-13b prod # For prod environment
-```
+ ```shell
+ modal environment create dev
+ ```
+
+ Now if you go to your dashboard on [Modal](https://modal.com/) you should be able to see the dev environment available in the environment drop down box. For easier development navigate to that environment so that you can track deployments and logging.
+
+3. **Configure Modal Environment**:
+ Switch to your development environment using the Modal CLI. This ensures all subsequent commands and deployments are done in the correct context.
+ ```shell
+ modal config set-environment dev # or main
+ ```
+ After running this command, you should see the 'dev' environment available in the environment dropdown box on your Modal dashboard.
+
+ 
+4. **Create Modal Secrets**:
+ Securely store your API keys as secrets in Modal, which will be used by your application to access necessary resources and services.
+ - **HuggingFace Token**:
+ Create a modal secret group with your Hugging Face token. Replace `<your huggingface token>` with the actual token.
+ ```shell
+ modal secret create huggingface HUGGINGFACE_TOKEN=<your huggingface token>
+ ```
+ - **Runner API Key**:
+ Create another secret group for the runner API key. Replace `<generate a random key>` with a strong, random key you've generated. Be sure to save this key somewhere as we'll need it for later!
+ ```shell
+ modal secret create ext-api-key RUNNER_API_KEY=<generate a random key>
+ ```
+
+ Now if you go to your dashboard on [Modal](https://modal.com/) and click on the secrets tab you should be able to see your keys deployed there.
+
+ 
+
+## Getting Started with Deployment
+
+With your environment now fully configured, you're ready to dive into deploying the OpenRouter Runner. This section guides you through deploying the Runner, adding new models or containers, and initiating tests to ensure everything is functioning as expected.
+
+### Adding a New Model with Existing Containers
+
+Adding new models to OpenRouter Runner is straightforward, especially when using models from Hugging Face that are compatible with existing containers. Here's how to do it:
+
+1. **Find and Copy the Model ID**: Browse [Hugging Face](https://huggingface.co/models) for the model you wish to deploy. For example, let's use `"mistralai/Mistral-7B-Instruct-v0.2"`.
+
+2. **Update Model List**: Open the `runner/containers/__init__.py` file. Locate the list of model IDs for the container you're using (e.g., `vllm_7b_model_ids` for a VLLM 7B model container). Add your new model ID to the list: | nit: the official casing is `vLLM` |
openrouter-runner | github_2023 | python | 29 | OpenRouterTeam | louisgv | @@ -89,7 +90,7 @@ async def generate(self, payload: Payload, params):
async for request_output in results_generator:
final_output = request_output
- output = request_output.outputs[0].text | Good catch - will run the e2e test without stream on this path, but this diff makes sense to me |
openrouter-runner | github_2023 | python | 17 | OpenRouterTeam | alexanderatallah | @@ -0,0 +1,54 @@
+from runner.engines.vllm import VllmEngine, VllmParams
+
+# Each container comprises of:
+# 1. An image
+# 2. A stub class wrapping an engine
+
+from modal import gpu, Image
+
+from shared.volumes import models_path
+from runner.shared.common import stub
+
+_gpu = gpu.A100(count=1, memory=80)
+
+_vllm_image = (
+ Image.from_registry(
+ "nvcr.io/nvidia/pytorch:23.09-py3"
+ # "nvcr.io/nvidia/pytorch:22.12-py3"
+ )
+ # Use latest torch
+ .pip_install(
+ "torch",
+ "torchvision",
+ "torchaudio",
+ )
+ .pip_install(
+ "transformers @ git+https://github.com/huggingface/transformers@main"
+ )
+ .pip_install("vllm @ git+https://github.com/vllm-project/vllm.git@main")
+)
+
+
+@stub.cls(
+ volumes={str(models_path): stub.models_volume},
+ image=_vllm_image,
+ gpu=_gpu,
+ allow_concurrent_inputs=16,
+ container_idle_timeout=10 * 60, # 5 minutes,
+ # keep_warm=True,
+)
+class Vllm_A100_32K_Container(VllmEngine):
+ def __init__(
+ self,
+ model_path: str,
+ ):
+ super().__init__(
+ VllmParams(
+ model=model_path,
+ tensor_parallel_size=_gpu.count,
+ trust_remote_code=True,
+ gpu_memory_utilization=0.95,
+ max_model_len=32_000,
+ max_num_batched_tokens=32_000, | do we need this? seems like vllm merge da pr sept 27 to remove the need: https://github.com/vllm-project/vllm/issues/1189#issuecomment-1738255066 |
openrouter-runner | github_2023 | python | 12 | OpenRouterTeam | alexanderatallah | @@ -8,6 +8,7 @@ def _to_lower_list(l: List[str]):
vllm_13b_model_ids = [
+ "mistralai/Mistral-7B-Instruct-v0.1", | do we need the same hardware for this model? (or if we stick with 2 gpus, will they scale efficiently with a 7b?) |
openrouter-runner | github_2023 | others | 8 | OpenRouterTeam | alexanderatallah | @@ -0,0 +1,32 @@
+# Add new model
+
+1. Copy one of the model file in `aux/models` and rename it to your model name.
+2. Change the `model_id` to the HF model ID.
+3. Adapt the Model class name accordingly.
+4. Import the model into [](./models/__init__.py) and add it to the get_model function.
+
+# Configuration
+
+See [](./shared/common.py)
+
+# Deployment
+
+1. Create a modal secret group
+ HUGGINGFACE_TOKEN = <your huggingface token>
+ with name "huggingface"
+2. Create a modal secret group
+ AUX_API_KEY = <generate a random key>
+ with name "ext-api-key"
+3. modal deploy aux/main.py
+4. Make sure `pwd` is `/modal`
+
+5. Run:
+
+```bash
+poetry shell
+
+modal deploy auxilary/main.py --env=dev | you mean `aux/main.py`? |
openrouter-runner | github_2023 | python | 8 | OpenRouterTeam | alexanderatallah | @@ -1,4 +1,18 @@
-from modal.config import Config
+from modal.config import Config as ModalConfig
+from pydantic import BaseModel
-modal_config = Config()
+modal_config = ModalConfig()
keep_warm = None if modal_config.get(key="environment") == "dev" else 1 | shouldn't we always cold-start now? |
openrouter-runner | github_2023 | python | 8 | OpenRouterTeam | alexanderatallah | @@ -0,0 +1,14 @@
+from modal import Stub
+from vllm_runner.shared.config import Config
+
+config = Config(
+ name="aux",
+ api_key_id="AUX_API_KEY",
+ download_dir="/model",
+ num_gpu=1,
+ max_batched_tokens=4096,
+ idle_timeout=5 * 60, # 5 minutes | Small comment: this is slightly confusing because `Config` is coming out of `vllm_runner`, which makes me think it's vllm-specific, but it's also being used as the `container_idle_timeout` on Modal containers.
Should we rename vllm_runner to runner, perhaps? |
openrouter-runner | github_2023 | python | 8 | OpenRouterTeam | alexanderatallah | @@ -0,0 +1,147 @@
+from vllm_runner.shared.protocol import (
+ Payload,
+ CompletionResponse,
+ ErrorPayload,
+ ErrorResponse,
+)
+from vllm_runner.shared.sampling_params import SamplingParams
+
+from typing import List
+from modal import Secret, method, gpu, Image
+
+from ..shared.common import stub, config
+
+
+model_id = "meta-llama/Llama-2-13b-chat-hf"
+
+
+def download_model():
+ import os
+ from huggingface_hub import snapshot_download
+ from pathlib import Path
+
+ # make MODEL_DIR if not existed
+ Path(config.download_dir).mkdir(parents=True, exist_ok=True)
+
+ snapshot_download(
+ model_id,
+ local_dir=config.download_dir,
+ token=os.environ["HUGGINGFACE_TOKEN"],
+ )
+
+
+image = (
+ Image.from_registry("nvcr.io/nvidia/pytorch:22.12-py3")
+ .pip_install(
+ "vllm == 0.1.7",
+ # Pinned to Sep/11/2023
+ # "vllm @ git+https://github.com/vllm-project/vllm.git@b9cecc26359794af863b3484a3464108b7d5ee5f",
+ # Pinned to 08/15/2023
+ # "vllm @ git+https://github.com/vllm-project/vllm.git@805de738f618f8b47ab0d450423d23db1e636fa2",
+ "typing-extensions==4.5.0", # >=4.6 causes typing issues
+ )
+ # Use the barebones hf-transfer package for maximum download speeds. No progress bar, but expect 700MB/s.
+ .pip_install("hf-transfer~=0.1")
+ .env({"HF_HUB_ENABLE_HF_TRANSFER": "1"})
+ .run_function(
+ download_model,
+ secret=Secret.from_name("huggingface"),
+ timeout=60 * 60,
+ )
+)
+
+
+@stub.cls(
+ image=image,
+ gpu=gpu.A100(count=config.num_gpu),
+ secret=Secret.from_name("huggingface"),
+ allow_concurrent_inputs=config.concurrency,
+ container_idle_timeout=config.idle_timeout,
+)
+class Llama2Chat13BModel:
+ async def __aenter__(self):
+ from vllm.engine.arg_utils import AsyncEngineArgs | does it work to call helper functions from inside each of these class methods? |
openrouter-runner | github_2023 | python | 8 | OpenRouterTeam | alexanderatallah | @@ -0,0 +1,147 @@
+from vllm_runner.shared.protocol import (
+ Payload,
+ CompletionResponse,
+ ErrorPayload,
+ ErrorResponse,
+)
+from vllm_runner.shared.sampling_params import SamplingParams
+
+from typing import List
+from modal import Secret, method, gpu, Image
+
+from ..shared.common import stub, config
+
+
+model_id = "meta-llama/Llama-2-13b-chat-hf"
+
+
+def download_model(): | why not use mk_gpu_image? |
openrouter-runner | github_2023 | python | 8 | OpenRouterTeam | alexanderatallah | @@ -0,0 +1,201 @@
+# Copied from: https://github.com/vllm-project/vllm/blob/bc0644574ca12d754a031596bdcfe8e1f0e6ab39/vllm/sampling_params.py | can we put something like `# BEGIN OPENROUTER MODIFICATION` / `# END OPENROUTER MODIFICATION` in the code below so we know why we forked it? |
openrouter-runner | github_2023 | python | 8 | OpenRouterTeam | alexanderatallah | @@ -1,5 +1,7 @@
# Copied from: https://github.com/vllm-project/vllm/blob/bc0644574ca12d754a031596bdcfe8e1f0e6ab39/vllm/sampling_params.py
+# The main reason we copy it here is to keep our CPU endpoint container lean - since all we need from vllm is this params validator and not the whole package | thanks 👍 |
openrouter-runner | github_2023 | python | 10 | OpenRouterTeam | louisgv | @@ -124,7 +121,7 @@ def download_model_to_folder():
@stub.cls(
- gpu=gpu.A100(count=NUM_GPU, memory=20),
+ gpu=gpu.A10G(count=NUM_GPU),
secret=Secret.from_name("huggingface"),
allow_concurrent_inputs=12,
container_idle_timeout=600, | Maybe we can reduce the idle timeout as well? |
openrouter-runner | github_2023 | python | 5 | OpenRouterTeam | alexanderatallah | @@ -88,63 +112,114 @@ def download_model_to_folder():
)
)
-stub = Stub("mythalion", image=image)
+stub = Stub(NAME, image=image)
@stub.cls(
- gpu=gpu.A100(),
+ gpu=gpu.A100(count=NUM_GPU, memory=20),
secret=Secret.from_name("huggingface"),
allow_concurrent_inputs=12,
container_idle_timeout=600,
keep_warm=1, | should we make keep_warm 0 when in the dev environment? |
openrouter-runner | github_2023 | typescript | 5 | OpenRouterTeam | alexanderatallah | @@ -2,7 +2,10 @@ import { once } from 'events';
import fs from 'fs/promises';
import { createInterface } from 'readline';
-export async function getWords(n: number, fileNo = 0): Promise<string> {
+export async function getWords( | nit: from the function naming, it was unclear if this fn generates random words or if it gets words out of a file. (had to read it to see that fileNo is actually important and it's doing the latter) |
openrouter-runner | github_2023 | python | 5 | OpenRouterTeam | alexanderatallah | @@ -158,37 +233,46 @@ def completion(
payload: Payload,
token: HTTPAuthorizationCredentials = Depends(auth_scheme),
):
- if token.credentials != os.environ["MYTHALION_API_KEY"]:
+ if token.credentials != os.environ[API_KEY_ID]:
raise HTTPException(
status_code=status.HTTP_401_UNAUTHORIZED,
detail="Incorrect bearer token",
headers={"WWW-Authenticate": "Bearer"},
)
- from fastapi.responses import StreamingResponse
from vllm.sampling_params import SamplingParams
- sampling_params = SamplingParams(
- # early_stopping=payload.params.early_stopping,
- # length_penalty=payload.params.length_penalty,
- best_of=payload.params.best_of,
- frequency_penalty=payload.params.frequency_penalty,
- ignore_eos=payload.params.ignore_eos,
- logprobs=payload.params.logprobs,
- max_tokens=payload.params.max_tokens,
- n=payload.params.n,
- presence_penalty=payload.params.presence_penalty,
- stop=payload.params.stop,
- temperature=payload.params.temperature,
- top_k=payload.params.top_k,
- top_p=payload.params.top_p,
- use_beam_search=payload.params.use_beam_search,
- )
+ try:
+ sampling_params = SamplingParams(
+ # early_stopping=payload.params.early_stopping,
+ # length_penalty=payload.params.length_penalty,
+ best_of=payload.params.best_of,
+ frequency_penalty=payload.params.frequency_penalty,
+ ignore_eos=payload.params.ignore_eos,
+ logprobs=payload.params.logprobs,
+ max_tokens=payload.params.max_tokens,
+ n=payload.params.n,
+ presence_penalty=payload.params.presence_penalty,
+ stop=payload.params.stop,
+ temperature=payload.params.temperature,
+ top_k=payload.params.top_k,
+ top_p=payload.params.top_p,
+ use_beam_search=payload.params.use_beam_search,
+ )
+ print(sampling_params)
+ except ValueError as e:
+ return create_error_response(status.HTTP_400_BAD_REQUEST, str(e))
+
+ model = Model()
+
+ input_ids = model.tokenize_prompt.remote(payload) | why do these two things remotely? and if each can take more than 50ms, is it faster to parallelize? |
openrouter-runner | github_2023 | python | 5 | OpenRouterTeam | alexanderatallah | @@ -88,63 +112,114 @@ def download_model_to_folder():
)
)
-stub = Stub("mythalion", image=image)
+stub = Stub(NAME, image=image)
@stub.cls(
- gpu=gpu.A100(),
+ gpu=gpu.A100(count=NUM_GPU, memory=20),
secret=Secret.from_name("huggingface"),
allow_concurrent_inputs=12,
container_idle_timeout=600,
keep_warm=1,
)
class Model:
- def __enter__(self):
+ async def __aenter__(self):
from vllm.engine.arg_utils import AsyncEngineArgs
from vllm.engine.async_llm_engine import AsyncLLMEngine
+ from vllm.transformers_utils.tokenizer import get_tokenizer
engine_args = AsyncEngineArgs(
model=MODEL_DIR,
- tensor_parallel_size=1,
+ tensor_parallel_size=NUM_GPU,
# using 95% of GPU memory by default
gpu_memory_utilization=0.95,
disable_log_requests=True,
+ max_num_batched_tokens=4096,
)
self.engine = AsyncLLMEngine.from_engine_args(engine_args)
- @method()
- async def generate(self, payload: Payload, params):
- import time
+ self.engine_model_config = await self.engine.get_model_config()
+ self.max_model_len = self.engine_model_config.get_max_model_len()
- results_generator = self.engine.generate(
- payload.prompt, params, payload.id
+ # A separate tokenizer to map token IDs to strings.
+ self.tokenizer = get_tokenizer(
+ engine_args.tokenizer,
+ tokenizer_mode=engine_args.tokenizer_mode,
+ trust_remote_code=engine_args.trust_remote_code,
)
- t0 = time.time()
- index, tokens = 0, 0
- async for request_output in results_generator:
- if (
- request_output.outputs[0].text
- and "\ufffd" == request_output.outputs[0].text[-1]
- ):
- continue
- token = request_output.outputs[0].text[index:]
- if payload.stream:
- choice = CompletionResponse(text=token).json(ensure_ascii=False)
- yield f"data: {choice}\n\n"
- else:
- yield token
- index = len(request_output.outputs[0].text)
- # Token accounting
- tokens = len(request_output.outputs[0].token_ids)
+ @method()
+ async def tokenize_prompt(self, payload: Payload) -> List[int]:
+ return self.tokenizer(payload.prompt).input_ids
- throughput = tokens / (time.time() - t0)
- print(f"Tokens count: {tokens} tokens")
- print(f"Request completed: {throughput:.4f} tokens/s")
+ @method()
+ async def validate_input(self, params, input_ids) -> Optional[JSONResponse]:
+ token_num = len(input_ids)
+ is_too_high = token_num + params.max_tokens > self.max_model_len
+
+ if is_too_high:
+ return create_error_response(
+ status.HTTP_400_BAD_REQUEST,
+ f"This model's maximum context length is {self.max_model_len} tokens. "
+ f"However, you requested {params.max_tokens + token_num} tokens "
+ f"({token_num} in the messages, "
+ f"{params.max_tokens} in the completion). "
+ f"Please reduce the length of the messages or completion.",
+ )
+ else:
+ return None
- # yield "[DONE]"
- # print(request_output.outputs[0].text)
+ @method()
+ async def generate(self, payload: Payload, params, input_ids):
+ try:
+ import time
+
+ results_generator = self.engine.generate(
+ payload.prompt, params, payload.id, input_ids
+ )
+
+ t0 = time.time()
+ index, tokens = 0, 0
+ output = ""
+ async for request_output in results_generator:
+ if (
+ request_output.outputs[0].text
+ and "\ufffd" == request_output.outputs[0].text[-1]
+ ):
+ continue | nit: would be nice to add a comment here to explain why we are skipping this |
openrouter-runner | github_2023 | python | 2 | OpenRouterTeam | alexanderatallah | @@ -108,6 +108,7 @@ def __enter__(self):
tensor_parallel_size=1,
# using 95% of GPU memory by default
gpu_memory_utilization=0.95,
+ log_requests=False, | can we undeploy this? see https://modal.com/logs/ap-oeR9RaoFC1krDiG94H2iUq?functionId=fu-W8IPkamMB6habWTtkQRiKJ&taskId=ta-3GWMAUS1dPKhkbTYAZVnh0&inputId=in-pw3u7DxpqrC3ynnn9dstHW |
openrouter-runner | github_2023 | python | 1 | OpenRouterTeam | alexanderatallah | @@ -48,13 +48,14 @@ def download_model_to_folder():
)
)
-stub = Stub("vllm-serving", image=image)
+stub = Stub("mythomax", image=image)
+
@stub.cls(
- gpu=gpu.A100(),
- secret=Secret.from_name("huggingface"),
- allow_concurrent_inputs=12,
- container_idle_timeout=300
+ gpu=gpu.A100(),
+ secret=Secret.from_name("huggingface"),
+ allow_concurrent_inputs=12, | nit: we estimated around 20 concurrent inputs earlier. what are the cons of increasing this number? |
openrouter-runner | github_2023 | python | 1 | OpenRouterTeam | alexanderatallah | @@ -48,13 +48,14 @@ def download_model_to_folder():
)
)
-stub = Stub("vllm-serving", image=image)
+stub = Stub("mythomax", image=image)
+
@stub.cls(
- gpu=gpu.A100(),
- secret=Secret.from_name("huggingface"),
- allow_concurrent_inputs=12,
- container_idle_timeout=300
+ gpu=gpu.A100(),
+ secret=Secret.from_name("huggingface"),
+ allow_concurrent_inputs=12,
+ container_idle_timeout=300, | what does this mean? is it just for the cpu container? how fast is it to cold-start? Not sure we should shut it down after 5 min if it takes 30 sec to cold-start |
openrouter-runner | github_2023 | python | 1 | OpenRouterTeam | alexanderatallah | @@ -112,10 +109,8 @@ async def generate(self, question: str):
print(f"Request completed: {throughput:.4f} tokens/s")
# print(request_output.outputs[0].text)
-@stub.function(
- timeout=60 * 10,
- allow_concurrent_inputs=12
-)
+
+@stub.function(timeout=60 * 10, allow_concurrent_inputs=12)
@web_endpoint()
def get(question: str): | can we try using https://vllm.readthedocs.io/en/latest/getting_started/quickstart.html#openai-compatible-server ? |
openrouter-runner | github_2023 | python | 1 | OpenRouterTeam | alexanderatallah | @@ -0,0 +1,168 @@
+import os
+
+from typing import List, Optional, Union
+
+from fastapi import Depends, HTTPException, status, Request
+
+from modal import Image, Secret, Stub, method, gpu, web_endpoint
+from pydantic import BaseModel
+from fastapi.security import HTTPBearer, HTTPAuthorizationCredentials
+
+
+MODEL_DIR = "/model"
+
+MODEL = "PygmalionAI/mythalion-13b"
+
+
+# https://github.com/vllm-project/vllm/blob/320a622ec4d098f2da5d097930f4031517e7327b/vllm/sampling_params.py#L7-L52
+
+
+class Params(BaseModel): | is this where we should specify default settings, or should we do that in the adapter? seems like a good thing for the adapter to own. (for that, Pygmalion has suggested defaults here: https://pygmalionai.github.io/blog/posts/introducing_pygmalion_2/#sillytavern) |
openrouter-runner | github_2023 | python | 1 | OpenRouterTeam | alexanderatallah | @@ -0,0 +1,168 @@
+import os
+
+from typing import List, Optional, Union
+
+from fastapi import Depends, HTTPException, status, Request
+
+from modal import Image, Secret, Stub, method, gpu, web_endpoint
+from pydantic import BaseModel
+from fastapi.security import HTTPBearer, HTTPAuthorizationCredentials
+
+
+MODEL_DIR = "/model"
+
+MODEL = "PygmalionAI/mythalion-13b"
+
+
+# https://github.com/vllm-project/vllm/blob/320a622ec4d098f2da5d097930f4031517e7327b/vllm/sampling_params.py#L7-L52
+
+
+class Params(BaseModel):
+ n: int = (1,)
+ best_of: Optional[int] = (None,)
+ presence_penalty: float = (0.0,)
+ frequency_penalty: float = (0.0,)
+ temperature: float = (1.0,)
+ top_p: float = (1.0,)
+ top_k: int = (-1,)
+ use_beam_search: bool = (False,)
+ length_penalty: float = (1.0,)
+ early_stopping: Union[bool, str] = (False,)
+ stop: Union[None, str, List[str]] = (None,)
+ ignore_eos: bool = (False,)
+ max_tokens: int = (16,)
+ logprobs: Optional[int] = (None,)
+
+
+class Payload(BaseModel):
+ id: str
+ prompt: str
+ params: Params
+
+
+auth_scheme = HTTPBearer()
+
+
+def download_model_to_folder():
+ from huggingface_hub import snapshot_download
+
+ snapshot_download(
+ MODEL,
+ local_dir=MODEL_DIR,
+ token=os.environ["HUGGINGFACE_TOKEN"],
+ )
+
+
+image = (
+ Image.from_registry("nvcr.io/nvidia/pytorch:22.12-py3")
+ .pip_install("torch==2.0.1", index_url="https://download.pytorch.org/whl/cu118")
+ # Pinned to 08/15/2023
+ .pip_install(
+ "vllm @ git+https://github.com/vllm-project/vllm.git@805de738f618f8b47ab0d450423d23db1e636fa2",
+ "typing-extensions==4.5.0", # >=4.6 causes typing issues
+ )
+ # Use the barebones hf-transfer package for maximum download speeds. No progress bar, but expect 700MB/s.
+ .pip_install("hf-transfer~=0.1")
+ .env({"HF_HUB_ENABLE_HF_TRANSFER": "1"})
+ .run_function(
+ download_model_to_folder,
+ secret=Secret.from_name("huggingface"),
+ timeout=60 * 20,
+ )
+)
+
+stub = Stub("mythalion", image=image)
+
+
+@stub.cls(
+ gpu=gpu.A100(),
+ secret=Secret.from_name("huggingface"),
+ allow_concurrent_inputs=12,
+ container_idle_timeout=600,
+)
+class Model:
+ def __enter__(self):
+ from vllm.engine.arg_utils import AsyncEngineArgs
+ from vllm.engine.async_llm_engine import AsyncLLMEngine
+
+ engine_args = AsyncEngineArgs(
+ model=MODEL_DIR,
+ tensor_parallel_size=1,
+ # Only uses 90% of GPU memory by default
+ gpu_memory_utilization=0.95,
+ )
+
+ self.engine = AsyncLLMEngine.from_engine_args(engine_args)
+ self.template = TEMPLATE | template doesn't seem defined (but prob should be managed by the adapter anyway) |
openrouter-runner | github_2023 | python | 1 | OpenRouterTeam | alexanderatallah | @@ -0,0 +1,168 @@
+import os
+
+from typing import List, Optional, Union
+
+from fastapi import Depends, HTTPException, status, Request
+
+from modal import Image, Secret, Stub, method, gpu, web_endpoint
+from pydantic import BaseModel
+from fastapi.security import HTTPBearer, HTTPAuthorizationCredentials
+
+
+MODEL_DIR = "/model"
+
+MODEL = "PygmalionAI/mythalion-13b"
+
+
+# https://github.com/vllm-project/vllm/blob/320a622ec4d098f2da5d097930f4031517e7327b/vllm/sampling_params.py#L7-L52
+
+
+class Params(BaseModel):
+ n: int = (1,)
+ best_of: Optional[int] = (None,)
+ presence_penalty: float = (0.0,)
+ frequency_penalty: float = (0.0,)
+ temperature: float = (1.0,)
+ top_p: float = (1.0,)
+ top_k: int = (-1,)
+ use_beam_search: bool = (False,)
+ length_penalty: float = (1.0,)
+ early_stopping: Union[bool, str] = (False,)
+ stop: Union[None, str, List[str]] = (None,)
+ ignore_eos: bool = (False,)
+ max_tokens: int = (16,)
+ logprobs: Optional[int] = (None,)
+
+
+class Payload(BaseModel):
+ id: str
+ prompt: str
+ params: Params
+
+
+auth_scheme = HTTPBearer()
+
+
+def download_model_to_folder():
+ from huggingface_hub import snapshot_download
+
+ snapshot_download(
+ MODEL,
+ local_dir=MODEL_DIR,
+ token=os.environ["HUGGINGFACE_TOKEN"],
+ )
+
+
+image = (
+ Image.from_registry("nvcr.io/nvidia/pytorch:22.12-py3")
+ .pip_install("torch==2.0.1", index_url="https://download.pytorch.org/whl/cu118")
+ # Pinned to 08/15/2023
+ .pip_install(
+ "vllm @ git+https://github.com/vllm-project/vllm.git@805de738f618f8b47ab0d450423d23db1e636fa2",
+ "typing-extensions==4.5.0", # >=4.6 causes typing issues
+ )
+ # Use the barebones hf-transfer package for maximum download speeds. No progress bar, but expect 700MB/s.
+ .pip_install("hf-transfer~=0.1")
+ .env({"HF_HUB_ENABLE_HF_TRANSFER": "1"})
+ .run_function(
+ download_model_to_folder,
+ secret=Secret.from_name("huggingface"),
+ timeout=60 * 20,
+ )
+)
+
+stub = Stub("mythalion", image=image)
+
+
+@stub.cls(
+ gpu=gpu.A100(),
+ secret=Secret.from_name("huggingface"),
+ allow_concurrent_inputs=12,
+ container_idle_timeout=600,
+)
+class Model:
+ def __enter__(self):
+ from vllm.engine.arg_utils import AsyncEngineArgs
+ from vllm.engine.async_llm_engine import AsyncLLMEngine
+
+ engine_args = AsyncEngineArgs(
+ model=MODEL_DIR,
+ tensor_parallel_size=1,
+ # Only uses 90% of GPU memory by default
+ gpu_memory_utilization=0.95,
+ )
+
+ self.engine = AsyncLLMEngine.from_engine_args(engine_args)
+ self.template = TEMPLATE
+
+ @method()
+ async def generate(self, payload: Payload): | skipping this for review, assuming WIP |
dotnet-cloud-native-build-2023 | github_2023 | csharp | 5 | bradygaster | davidfowl | @@ -30,74 +30,80 @@ protected override async Task ExecuteAsync(CancellationToken stoppingToken)
var orders = await _ordersClient.GetOrders();
activity?.AddTag("order-count", orders?.Count ?? 0);
- var orderTasks = new List<Task>();
foreach (Order? order in orders)
{
- orderTasks.Add(Task.Run(async () =>
- {
- _logger.LogInformation($"Checking inventory for Order {order.OrderId}");
-
- using (var orderActivity = _activitySource.StartActivity("order-processor.process-order"))
- {
- orderActivity?.AddTag("order-id", order.OrderId);
- orderActivity?.AddTag("product-count", order.Cart.Length);
-
- bool canWeFulfillOrder = true;
- var itemTasks = new List<Task>();
-
- foreach (var cartItem in order.Cart)
- {
- itemTasks.Add(Task.Run(async () =>
- {
- _logger.LogInformation($"Checking inventory for product id {cartItem.ProductId} in order {order.OrderId}");
-
- var inStock = await _productsClient.CanInventoryFulfill(cartItem.ProductId, cartItem.Quantity);
-
- if (inStock)
- {
- _logger.LogInformation($"Inventory OK for product id {cartItem.ProductId} in order {order.OrderId}");
- }
- else
- {
- _logger.LogInformation($"Not enough inventory for product id {cartItem.ProductId} in order {order.OrderId}");
-
- canWeFulfillOrder = canWeFulfillOrder && inStock;
- }
- }));
- }
- await Task.WhenAll(itemTasks);
- orderActivity?.SetTag("can-fulfill-order", canWeFulfillOrder);
-
- if (canWeFulfillOrder)
- {
- var invTasks = new List<Task>();
- foreach (var cartItem in order.Cart)
- {
- invTasks.Add(Task.Run(async () =>
- {
- _logger.LogInformation($"Removing {cartItem.Quantity} of product id {cartItem.ProductId} from inventory");
-
- await _productsClient.SubtractInventory(cartItem.ProductId, cartItem.Quantity);
-
- _logger.LogInformation($"Removed {cartItem.Quantity} of product id {cartItem.ProductId} from inventory");
- }));
- }
-
- _logger.LogInformation($"Marking order {order.OrderId} as ready for shipment");
-
- invTasks.Add(_ordersClient.MarkOrderReadyForShipment(order));
-
- _logger.LogInformation($"Marked order {order.OrderId} as ready for shipment");
-
- await Task.WhenAll(invTasks);
- }
- }
- }));
- await Task.WhenAll(orderTasks);
+ // Processing orders sequentially as it looks better in the traces and
+ // solves concurrency issues with multiple orders consuming the same inventory
+ await ProcessOrder(order);
}
}
await Task.Delay(15000, stoppingToken);
}
}
+
+ private async Task ProcessOrder(Order order)
+ {
+ using (var orderActivity = _activitySource.StartActivity("order-processor.process-order"))
+ {
+ _logger.LogInformation($"Checking inventory for Order {order.OrderId}");
+
+ orderActivity?.AddTag("order-id", order.OrderId);
+ orderActivity?.AddTag("product-count", order.Cart.Length);
+
+ var itemTasks = new List<Task<bool>>();
+ foreach (var cartItem in order.Cart)
+ {
+ itemTasks.Add(validateInventory(cartItem, order));
+ }
+ await Task.WhenAll(itemTasks);
+
+ orderActivity?.AddEvent(new ActivityEvent("inventory-check-complete"));
+
+ bool canWeFulfillOrder = true;
+ itemTasks.ForEach(itemTask => canWeFulfillOrder = canWeFulfillOrder && itemTask.Result);
+
+ orderActivity?.SetTag("can-fulfill-order", canWeFulfillOrder);
+
+ if (canWeFulfillOrder)
+ {
+ var invTasks = new List<Task>();
+ foreach (var cartItem in order.Cart)
+ {
+ invTasks.Add(updateInventory(cartItem));
+ }
+
+ _logger.LogInformation($"Marking order {order.OrderId} as ready for shipment");
+ invTasks.Add(_ordersClient.MarkOrderReadyForShipment(order));
+ await Task.WhenAll(invTasks);
+ _logger.LogInformation($"Marked order {order.OrderId} as ready for shipment");
+ }
+ }
+ }
+
+ private async Task<bool> validateInventory(CartItem cartItem, Order order) | These lowercase methods are java |
axum-htmx | github_2023 | others | 15 | robertwayne | robertwayne | @@ -0,0 +1,123 @@
+use axum_core::response::{IntoResponseParts, ResponseParts};
+use http::header::VARY;
+
+use crate::{extractors, headers, HxError};
+
+/// The `Vary: HX-Request` header.
+///
+/// You may want to add this header to the response if your handler responds differently based on
+/// the `HX-Request` request header.
+///
+/// For example, if your server renders the full HTML when the `HX-Request` header is missing or
+/// `false`, and it renders a fragment of that HTML when `HX-Request: true`.
+///
+/// You probably need this only for `GET` requests, as other HTTP methods are not cached by default.
+///
+/// See <https://htmx.org/docs/#caching> for more information.
+#[derive(Debug, Clone)]
+pub struct VaryHxRequest;
+
+impl IntoResponseParts for VaryHxRequest {
+ type Error = HxError;
+
+ fn into_response_parts(self, mut res: ResponseParts) -> Result<ResponseParts, Self::Error> {
+ res.headers_mut()
+ .insert(VARY, headers::HX_REQUEST.try_into()?);
+
+ Ok(res)
+ }
+}
+
+impl extractors::HxRequest {
+ /// Convenience method to create the corresponding `Vary` response header
+ pub fn vary_response() -> VaryHxRequest {
+ VaryHxRequest
+ }
+}
+
+/// The `Vary: HX-Target` header.
+///
+/// You may want to add this header to the response if your handler responds differently based on
+/// the `HX-Target` request header.
+///
+/// You probably need this only for `GET` requests, as other HTTP methods are not cached by default.
+///
+/// See <https://htmx.org/docs/#caching> for more information.
+#[derive(Debug, Clone)]
+pub struct VaryHxTarget;
+
+impl IntoResponseParts for VaryHxTarget {
+ type Error = HxError;
+
+ fn into_response_parts(self, mut res: ResponseParts) -> Result<ResponseParts, Self::Error> {
+ res.headers_mut()
+ .insert(VARY, headers::HX_TARGET.try_into()?); | We can use `http::HeaderValue::from_static(HX_TARGET)` instead. This applies to the other instances of `try_into()?` as well.
Originally I was going to say we could make this `type Error = Infallible`, but after thinking some more, `.insert()` can actually panic, so this should probably be addressed in the existing places as well. That's for another PR, though. Just a note for myself. |
claude-to-chatgpt | github_2023 | python | 9 | jtsang4 | jtsang4 | @@ -74,15 +74,9 @@ def claude_to_chatgpt_response_stream(self, claude_response, prev_decoded_respon
"object": "chat.completion.chunk",
"created": int(time.time()),
"model": "gpt-3.5-turbo-0301",
- "usage": {
- "prompt_tokens": 0,
- "completion_tokens": completion_tokens,
- "total_tokens": completion_tokens,
- },
"choices": [
{
"delta": {
- "role": "assistant", | The "role" field appeared in my previouse tests, please check it again. |
online-edit-web | github_2023 | typescript | 123 | xun082 | xun082 | @@ -55,6 +61,46 @@ export default function CodeEditor({ editorId }: CodeEditorProps) {
border: isOver ? '1px #3b82f6 solid' : undefined,
};
+ _editor &&
+ _editor.addCommand(monaco.KeyMod.CtrlCmd | monaco.KeyCode.KeyS, () => {
+ const prettierValue = getPrettierConfig(fileData);
+ formatWithPrettier(_editor, prettierValue);
+ });
+
+ // 格式化代码
+ const formatWithPrettier = async (item: editor.IStandaloneCodeEditor, prettierConfig: object) => { | 这里的ts类型别用object，很容易有报错，可以用Record这种 record<string,any>这样，又或者record<string,string|boolean|number> |
online-edit-web | github_2023 | typescript | 119 | xun082 | xun082 | @@ -210,14 +211,23 @@ const SearchPage: FC = () => {
<IncludeAndExclude />
</div>
<div className="mx-7 text-[12px] select-none text-gray-500">
- {searchInpVal ? (
- searchResult && searchResult.length ? (
- <div>{`${resultCount.fileCount} 文件中 ${resultCount.matchCount} 个结果`}</div>
+ {data ? (
+ searchInpVal ? (
+ isShowResult ? (
+ <div>{`${resultCount.fileCount} 文件中 ${resultCount.matchCount} 个结果`}</div>
+ ) : (
+ <div>未找到结果</div>
+ )
) : (
- <div>未找到结果</div>
+ '' | 这里返回null |
create-neat | github_2023 | typescript | 309 | xun082 | uaenaTzx | @@ -146,9 +146,9 @@ export default async function createAppTest(projectName: string, options: Record
if (gitCheck(rootDirectory)) exec("git init", { cwd: rootDirectory });
// 安装传入的依赖
- if (process.env.NODE_ENV === "PROD") {
- await dependenciesInstall(rootDirectory, packageManager);
- }
+ // if (process.env.NODE_ENV === "PROD") { | 这块为什么要删除
其他插件可能会带进来一些依赖到 pka 里面 |
create-neat | github_2023 | others | 309 | xun082 | uaenaTzx | @@ -0,0 +1,47 @@
+import fs from 'fs';
+import path from 'path';
+import spawn from 'cross-spawn';
+import {fileURLToPath} from 'url'; | 看着没有格式化 |
create-neat | github_2023 | others | 309 | xun082 | uaenaTzx | @@ -0,0 +1,47 @@
+import fs from 'fs';
+import path from 'path';
+import spawn from 'cross-spawn';
+import {fileURLToPath} from 'url';
+
+const __filename = fileURLToPath(import.meta.url);
+const __dirname = path.dirname(__filename);
+
+//插件目录
+const pluginsDir = path.resolve(__dirname, '../packages/@plugin');
+
+function publishPackage(packagePath) {
+ return new Promise((resolve, reject) => {
+ console.log(`Publishing package at ${packagePath}...`); | log 可以带个颜色（chalk） |
create-neat | github_2023 | others | 309 | xun082 | uaenaTzx | @@ -0,0 +1,47 @@
+import fs from 'fs';
+import path from 'path';
+import spawn from 'cross-spawn';
+import {fileURLToPath} from 'url';
+
+const __filename = fileURLToPath(import.meta.url);
+const __dirname = path.dirname(__filename);
+
+//插件目录
+const pluginsDir = path.resolve(__dirname, '../packages/@plugin');
+
+function publishPackage(packagePath) {
+ return new Promise((resolve, reject) => {
+ console.log(`Publishing package at ${packagePath}...`);
+ const proc = spawn('npm',['publish', '--access', 'public'],{cwd:packagePath, stdio:'inherit'}); | 看着都没有格式化 |
create-neat | github_2023 | others | 309 | xun082 | uaenaTzx | @@ -0,0 +1,47 @@
+import fs from 'fs';
+import path from 'path';
+import spawn from 'cross-spawn';
+import {fileURLToPath} from 'url';
+
+const __filename = fileURLToPath(import.meta.url);
+const __dirname = path.dirname(__filename);
+
+//插件目录
+const pluginsDir = path.resolve(__dirname, '../packages/@plugin');
+
+function publishPackage(packagePath) {
+ return new Promise((resolve, reject) => {
+ console.log(`Publishing package at ${packagePath}...`);
+ const proc = spawn('npm',['publish', '--access', 'public'],{cwd:packagePath, stdio:'inherit'});
+ proc.on('close',(code) => {
+ if(code === 0) {
+ console.log(`Published${packagePath}successfully.`);
+ resolve()
+ }else {
+ reject(new Error(`Error publishing ${packagePath}`)) | 输出信息尽量完整，code 也可以输出 |
create-neat | github_2023 | others | 309 | xun082 | uaenaTzx | @@ -0,0 +1,47 @@
+import fs from 'fs'; | 还缺少一些东西，建议修改的点：
1. 发包过程改成交互式的（参考 cn create）
2. 交互式的具体流程可以包括：是否要批量发，检查当前包是否有更新（否则不用发），是否选择发 beta 包 等等，你再设计看看 |
create-neat | github_2023 | others | 308 | xun082 | uaenaTzx | @@ -72,7 +72,7 @@
},
"engines": {
"node": ">=18",
- "pnpm": "9.4.0"
+ "pnpm": "^9.0.0" | 不要轻易改依赖版本，这个是有什么背景吗
|
create-neat | github_2023 | others | 308 | xun082 | uaenaTzx | @@ -1,5 +1,21 @@
// const path = require("path");
module.exports = {
- rules: [],
+ rules: [ | 这块其实还有很多优化空间，可以看到历史的 config 配置里存在 构建工具 x 框架 的情况，我们看看能不能通过协议的方式来解决问题
如果遇到特殊场景不好做的，可以找我讨论 |
create-neat | github_2023 | javascript | 308 | xun082 | uaenaTzx | @@ -2,11 +2,18 @@ module.exports = {
rules: [
{
test: /\.(ts|tsx|js|jsx)$/,
- exclude: /node_modules/,
- use: {
- loader: "swc-loader",
- exclude: /node_modules/,
- },
+ exclude: [/node_modules/, /public/, /(.|_)min\.js$/], | 同 babel |
create-neat | github_2023 | others | 308 | xun082 | uaenaTzx | @@ -1,7 +1,7 @@
-const PluginConfig = require("./config/index.cjs");
+// const PluginConfig = require("./config/index.cjs");
-const pluginBabel = (buildTool, template) => { | 不需要的就删了吧 |
create-neat | github_2023 | typescript | 308 | xun082 | uaenaTzx | @@ -163,11 +163,12 @@ class Generator {
}
// 根据环境变量加载 plugin/template
+ // 返回增加可选的buildTool，编译器🥱插件(babel/swc)需要 | 这个注释改一下 |
create-neat | github_2023 | typescript | 310 | xun082 | uaenaTzx | @@ -3,21 +3,23 @@ import generator from "@babel/generator";
import fs from "fs-extra";
import chalk from "chalk";
import { parse } from "@babel/parser";
-
-import { relativePathToRoot } from "../utils/constants";
-import { createFiles } from "../utils/createFiles";
-import { createConfigByParseAst } from "../utils/ast/parseAst";
-import { Preset } from "../utils/preset";
-import { readTemplateFileContent } from "../utils/fileController";
-import generateBuildToolConfigFromEJS from "../utils/generateBuildToolConfigFromEJS";
-import { buildToolType } from "../types";
-
-import GeneratorAPI from "./GeneratorAPI";
-import ConfigTransform from "./ConfigTransform";
-import TemplateAPI from "./TemplateAPI";
-import FileTree from "./FileTree";
-import BaseAPI from "./BaseAPI";
-
+import { pathToFileURL } from "url";
+
+import { CNRootDirectory, relativePathToRoot } from "../utils/constants.js";
+import { createFiles } from "../utils/createFiles.js";
+import { createConfigByParseAst } from "../utils/ast/parseAst.js";
+import { Preset } from "../utils/preset.js";
+import { readTemplateFileContent } from "../utils/fileController.js";
+import generateBuildToolConfigFromEJS from "../utils/generateBuildToolConfigFromEJS.js";
+import { buildToolType } from "../types/index.js";
+
+import GeneratorAPI from "./GeneratorAPI.js";
+import ConfigTransform from "./ConfigTransform.js";
+import TemplateAPI from "./TemplateAPI.js";
+import FileTree from "./FileTree.js";
+import BaseAPI from "./BaseAPI.js"; | 应该是 ts 后缀 |
create-neat | github_2023 | typescript | 310 | xun082 | uaenaTzx | @@ -231,10 +234,9 @@ class Generator {
// 单独处理一个框架相关依赖，主要是将框架相关的依赖包插入到pkg内，以及将需要的构建工具配置合并到构建工具模板中
async templateGenerate() {
- const templateGenerator = await this.loadBase(
- `packages/core/template/template-${this.templateName}/generator/index.cjs`,
- "",
- );
+ const templatePath = `packages/core/dist/template/template-${this.templateName}/generator/index.js`;
+ console.log("templatePath: ", templatePath); | log 删了吧 |
create-neat | github_2023 | typescript | 307 | xun082 | uaenaTzx | @@ -18,6 +17,8 @@ import dependenciesInstall from "./dependenciesInstall";
import { createReadmeString } from "./createFiles";
import { buildToolConfigDevDependencies, buildToolScripts } from "./constants";
+const { resolveApp } = require("@laconic/utils"); | 这个文件并没有引用这个呀 |
create-neat | github_2023 | typescript | 306 | xun082 | uaenaTzx | @@ -1,28 +1,35 @@
+/** 通用协议常量，定义了跨不同工具和框架交互的协议类型。*/
const globalProtocol = {
- ENTRY_FILE: "ENTRY_FILE", // 入口文件配置，如：入口文件引入全局 less、scss
+ /** 入口文件配置，例如引入全局的less、scss文件。 */
+ ENTRY_FILE: "ENTRY_FILE",
+ /** 更改文件导出的内容。 */
UPDATE_EXPORT_CONTENT_PROTOCOL: "UPDATE_EXPORT_CONTENT_PROTOCOL",
+ /** 向文件插入import语句。 */
INSERT_IMPORT_PROTOCOL: "INSERT_IMPORT_PROTOCOL",
+ /** 替换文件中对应slot的内容。 */
SLOT_CONTENT_PROTOCOL: "SLOT_CONTENT_PROTOCOL",
-};
+} as const;
-// 插件对框架的协议
+/** 插件对框架的协议，扩展了通用协议，并添加处理样式的协议。 */ | 扩展了通用协议，并添加处理样式的协议
这个没必要注释 |
create-neat | github_2023 | typescript | 306 | xun082 | uaenaTzx | @@ -1,28 +1,35 @@
+/** 通用协议常量，定义了跨不同工具和框架交互的协议类型。*/
const globalProtocol = {
- ENTRY_FILE: "ENTRY_FILE", // 入口文件配置，如：入口文件引入全局 less、scss
+ /** 入口文件配置，例如引入全局的less、scss文件。 */
+ ENTRY_FILE: "ENTRY_FILE",
+ /** 更改文件导出的内容。 */
UPDATE_EXPORT_CONTENT_PROTOCOL: "UPDATE_EXPORT_CONTENT_PROTOCOL",
+ /** 向文件插入import语句。 */
INSERT_IMPORT_PROTOCOL: "INSERT_IMPORT_PROTOCOL",
+ /** 替换文件中对应slot的内容。 */
SLOT_CONTENT_PROTOCOL: "SLOT_CONTENT_PROTOCOL",
-};
+} as const;
-// 插件对框架的协议
+/** 插件对框架的协议，扩展了通用协议，并添加处理样式的协议。 */
const pluginToTemplateProtocol = {
...globalProtocol,
+ /** 处理样式协议，如：less、scss。 */
PROCESS_STYLE_PLUGIN: "PROCESS_STYLE_PLUGIN",
-};
+} as const;
-// 插件对构建工具的协议
+/** 插件对构建工具的协议 */
const pluginToBuildToolProtocol = {
...globalProtocol,
-};
+} as const;
-// 框架对构建工具的协议
+/** 框架对构建工具的协议 */
const templateToBuildToolProtocol = {
...globalProtocol,
- ADD_CONFIG: "ADD_CONFIG", //根据框架，不同的打包工具需要不同的插件，有些是都需要用的，有些是框架独有的
-};
+ /** 根据框架，不同的打包工具需要不同的插件，有些是都需要用的，有些是框架独有的 */ | 根据框架，不同的打包工具需要不同的插件，有些是都需要用的，有些是框架独有的
解释有些奇怪，add config 不一定是添加插件配置，描述有些局限 |
create-neat | github_2023 | typescript | 306 | xun082 | uaenaTzx | @@ -1,28 +1,35 @@
+/** 通用协议常量，定义了跨不同工具和框架交互的协议类型。*/
const globalProtocol = {
- ENTRY_FILE: "ENTRY_FILE", // 入口文件配置，如：入口文件引入全局 less、scss
+ /** 入口文件配置，例如引入全局的less、scss文件。 */
+ ENTRY_FILE: "ENTRY_FILE",
+ /** 更改文件导出的内容。 */
UPDATE_EXPORT_CONTENT_PROTOCOL: "UPDATE_EXPORT_CONTENT_PROTOCOL",
+ /** 向文件插入import语句。 */
INSERT_IMPORT_PROTOCOL: "INSERT_IMPORT_PROTOCOL",
+ /** 替换文件中对应slot的内容。 */ | slot 是什么也得说明清楚，比如：对目标文件的指定插槽进行内容注入 |
create-neat | github_2023 | typescript | 306 | xun082 | uaenaTzx | @@ -1,28 +1,35 @@
+/** 通用协议常量，定义了跨不同工具和框架交互的协议类型。*/
const globalProtocol = {
- ENTRY_FILE: "ENTRY_FILE", // 入口文件配置，如：入口文件引入全局 less、scss
+ /** 入口文件配置，例如引入全局的less、scss文件。 */ | 英文字母和中文留空格，比如 xx 这样 |
create-neat | github_2023 | typescript | 298 | xun082 | uaenaTzx | @@ -66,6 +66,11 @@ export const buildToolScripts = {
},
};
+export const templateFileExtension = { | 注释明确 |
create-neat | github_2023 | typescript | 298 | xun082 | uaenaTzx | @@ -174,13 +197,18 @@ class FileTree {
* @param url 添加文件的原始的真实路径
* @param parentDir 父文件夹路径
*/
- addToTreeByTemplateDirPathAndEjs(url: string, parentDir: string, options: any) {
+ addToTreeByTemplateDirPathAndEjs(url: string, parentDir: string, options: any, isTs: any) { | any 不对，而且这里是可以读取到 process 的值的，不太需要传参 |
create-neat | github_2023 | typescript | 298 | xun082 | uaenaTzx | @@ -107,13 +108,25 @@ class FileTree {
return file;
}
+ private transformFileExtension(fileContent: string, fileName: string, fileExtension: string) { | 注释尽量补充完善 |
create-neat | github_2023 | others | 302 | xun082 | uaenaTzx | @@ -1,16 +1,104 @@
-module.exports = (generatorAPI) => {
+const path = require("path");
+const fs = require("fs");
+
+function fileRender(files) {
+ try {
+ const outputDir = path.join(__dirname, "template");
+ Object.entries(files).forEach(([filePath, content]) => {
+ const fullPath = path.join(outputDir, filePath);
+ fs.ensureFileSync(fullPath);
+ fs.writeFileSync(fullPath, content, "utf-8");
+ });
+ } catch (error) {
+ console.log("文件渲染失败: ", error);
+ }
+}
+
+const basicConfig = {
+ scripts: {
+ postinstall: "husky install",
+ "commit-msg": "commitlint --edit $1",
+ },
+ "lint-staged": {
+ "*.{ts,tsx,js,jsx}": ["pnpm format:ci", "pnpm lint:ci"],
+ },
+ devDependencies: {
+ husky: "^9.0.11",
+ "lint-staged": "^15.2.0",
+ "@commitlint/cli": "^18.4.3",
+ "@commitlint/config-conventional": "^18.4.3",
+ },
+};
+
+const generateBasicConfig = (generatorAPI) => {
+ generatorAPI.extendPackage(basicConfig);
+ generatorAPI.protocolGenerate({
+ [pluginToTemplateProtocol.RENDER_FILE]: { | 这个不走 template 渲染吗 |
create-neat | github_2023 | others | 302 | xun082 | uaenaTzx | @@ -1,16 +1,104 @@
-module.exports = (generatorAPI) => {
+const path = require("path");
+const fs = require("fs");
+
+function fileRender(files) {
+ try {
+ const outputDir = path.join(__dirname, "template");
+ Object.entries(files).forEach(([filePath, content]) => {
+ const fullPath = path.join(outputDir, filePath);
+ fs.ensureFileSync(fullPath);
+ fs.writeFileSync(fullPath, content, "utf-8");
+ });
+ } catch (error) {
+ console.log("文件渲染失败: ", error);
+ }
+}
+
+const basicConfig = {
+ scripts: {
+ postinstall: "husky install",
+ "commit-msg": "commitlint --edit $1",
+ },
+ "lint-staged": {
+ "*.{ts,tsx,js,jsx}": ["pnpm format:ci", "pnpm lint:ci"],
+ },
+ devDependencies: {
+ husky: "^9.0.11",
+ "lint-staged": "^15.2.0",
+ "@commitlint/cli": "^18.4.3",
+ "@commitlint/config-conventional": "^18.4.3",
+ },
+};
+
+const generateBasicConfig = (generatorAPI) => {
+ generatorAPI.extendPackage(basicConfig);
+ generatorAPI.protocolGenerate({
+ [pluginToTemplateProtocol.RENDER_FILE]: {
+ params: {
+ files: {
+ ".commitlintrc.js": `module.exports = {
+ extends: ["@commitlint/config-conventional"]
+ }`,
+ },
+ content: {
+ fileRender,
+ },
+ },
+ },
+ });
+};
+
+const generateStrictConfig = (generatorAPI) => {
generatorAPI.extendPackage({
+ ...basicConfig,
scripts: {
- postinstall: "husky install",
+ ...basicConfig.scripts,
+ "pre-commit": "lint-staged && npm run test",
},
"lint-staged": {
- "*.{ts,tsx,js,jsx}": ["pnpm format:ci", "pnpm lint:ci"],
+ ...basicConfig["lint-staged"],
+ "*.{css,scss,less}": ["pnpm stylelint"], | 包管理器不是写死的 |
create-neat | github_2023 | others | 302 | xun082 | uaenaTzx | @@ -1,16 +1,104 @@
-module.exports = (generatorAPI) => {
+const path = require("path");
+const fs = require("fs");
+
+function fileRender(files) {
+ try {
+ const outputDir = path.join(__dirname, "template");
+ Object.entries(files).forEach(([filePath, content]) => {
+ const fullPath = path.join(outputDir, filePath);
+ fs.ensureFileSync(fullPath);
+ fs.writeFileSync(fullPath, content, "utf-8");
+ });
+ } catch (error) {
+ console.log("文件渲染失败: ", error);
+ }
+}
+
+const basicConfig = {
+ scripts: {
+ postinstall: "husky install",
+ "commit-msg": "commitlint --edit $1",
+ },
+ "lint-staged": {
+ "*.{ts,tsx,js,jsx}": ["pnpm format:ci", "pnpm lint:ci"],
+ },
+ devDependencies: {
+ husky: "^9.0.11",
+ "lint-staged": "^15.2.0",
+ "@commitlint/cli": "^18.4.3",
+ "@commitlint/config-conventional": "^18.4.3",
+ },
+};
+
+const generateBasicConfig = (generatorAPI) => {
+ generatorAPI.extendPackage(basicConfig);
+ generatorAPI.protocolGenerate({
+ [pluginToTemplateProtocol.RENDER_FILE]: {
+ params: {
+ files: {
+ ".commitlintrc.js": `module.exports = {
+ extends: ["@commitlint/config-conventional"]
+ }`,
+ },
+ content: {
+ fileRender,
+ },
+ },
+ },
+ });
+};
+
+const generateStrictConfig = (generatorAPI) => {
generatorAPI.extendPackage({
+ ...basicConfig,
scripts: {
- postinstall: "husky install",
+ ...basicConfig.scripts,
+ "pre-commit": "lint-staged && npm run test",
},
"lint-staged": {
- "*.{ts,tsx,js,jsx}": ["pnpm format:ci", "pnpm lint:ci"],
+ ...basicConfig["lint-staged"],
+ "*.{css,scss,less}": ["pnpm stylelint"],
},
devDependencies: {
- husky: "^9.0.11",
- "lint-staged": "^15.2.0",
- "@commitlint/cli": "^18.4.3",
- "@commitlint/config-conventional": "^18.4.3",
+ ...basicConfig.devDependencies,
+ stylelint: "^15.10.0",
},
});
+ generatorAPI.protocolGenerate({
+ [pluginToTemplateProtocol.RENDER_FILE]: {
+ params: {
+ files: {
+ ".commitlintrc.js": `module.exports ={ | +1 |
create-neat | github_2023 | typescript | 302 | xun082 | uaenaTzx | @@ -8,6 +8,8 @@ const globalProtocol = {
const pluginToTemplateProtocol = {
...globalProtocol,
PROCESS_STYLE_PLUGIN: "PROCESS_STYLE_PLUGIN",
+ INSERT_IMPORT: "INSERT_IMPORT",
+ RENDER_FILE: "RENDER_FILE", | 没看到相关的处理器，不过这个协议确实意义不大 |
create-neat | github_2023 | others | 299 | xun082 | uaenaTzx | @@ -35,11 +35,22 @@ module.exports = (generatorAPI) => {
[pluginToTemplateProtocol.UPDATE_EXPORT_CONTENT_PROTOCOL]: {
params: {
url: 'src/App',
- exportContent: 'Observer',
+ exportContent: 'observer', | 这个改动是有什么原因 |
create-neat | github_2023 | others | 299 | xun082 | uaenaTzx | @@ -35,11 +35,22 @@ module.exports = (generatorAPI) => {
[pluginToTemplateProtocol.UPDATE_EXPORT_CONTENT_PROTOCOL]: {
params: {
url: 'src/App',
- exportContent: 'Observer',
+ exportContent: 'observer',
astOptions: {
parserOptions: { sourceType: "module", plugins: ["typescript", "jsx"] }
}
},
},
+ [pluginToTemplateProtocol.SLOT_CONTENT_PROTOCOL]: {
+ params: {
+ slotConfig: [
+ {
+ url: 'src/App',
+ slotName: 'store-slot',
+ slotContent: "console.log('当前store的number值为: ', store.number);", | 内容修正 |
create-neat | github_2023 | javascript | 299 | xun082 | uaenaTzx | @@ -2,7 +2,9 @@ import React from "react";
import "./index.css";
function App() {
+ // <!--slot: store-slot --> | 不用加 // 容易出现歧义 |
create-neat | github_2023 | typescript | 299 | xun082 | uaenaTzx | @@ -0,0 +1,24 @@
+import prettier from "prettier";
+import prettierPluginVue from "prettier-plugin-vue"; | 只有 vue 的吗，这个不能搞成通用的吗 |
create-neat | github_2023 | others | 295 | xun082 | uaenaTzx | @@ -10,9 +10,31 @@ module.exports = (generatorAPI) => {
});
generatorAPI.protocolGenerate({
+ [pluginToTemplateProtocol.INSERT_IMPORT_PROTOCOL]: {
+ params: {
+ imports: [
+ {
+ dir: "src/App",
+ module: [ | modules吧 |
create-neat | github_2023 | others | 295 | xun082 | uaenaTzx | @@ -10,9 +10,31 @@ module.exports = (generatorAPI) => {
});
generatorAPI.protocolGenerate({
+ [pluginToTemplateProtocol.INSERT_IMPORT_PROTOCOL]: {
+ params: {
+ imports: [
+ {
+ dir: "src/App",
+ module: [
+ {
+ name: "mobx",
+ from: "mobx-react-lite",
+ },
+ {
+ name: "store",
+ from: "./store",
+ }
+ ]
+ }
+ ],
+ astOptions: { | 这块的传参确实有一定学习成本 |
create-neat | github_2023 | typescript | 294 | xun082 | uaenaTzx | @@ -0,0 +1,35 @@
+import path from "path";
+
+import { FileData } from "../models/FileTree";
+
+/**
+ * 根据给定的文件路径，从嵌套的文件结构中检索目标文件的数据。
+ * | jsdoc 不需要这一行 |
create-neat | github_2023 | typescript | 294 | xun082 | uaenaTzx | @@ -52,3 +52,14 @@ export const createCallExpression = (
*/
export const createObjectProperty = (name: string, value: Parameters<typeof objectProperty>[1]) =>
objectProperty(identifier(name), value);
+
+/**
+ * 封装遍历ast中ExportDefaultDeclaration的通用函数
+ * @param {object} path
+ * @param {object} t
+ * @param {string} content 要包裹导出的字符串
+ */
+export const exportDefaultDeclarationUtils = (path: any, t: any, content: string) => { | any 尽量避免吧 |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.