-
Notifications
You must be signed in to change notification settings - Fork 1.6k
feat: Add support of scheduling attention dp request #6246
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
base: main
Are you sure you want to change the base?
Conversation
📝 WalkthroughWalkthroughThis change introduces a new Changes
Sequence Diagram(s)sequenceDiagram
participant Client
participant LLMAPI
participant Executor
participant Worker
participant RequestQueue
Client->>LLMAPI: generate / generate_async(..., scheduling_params)
LLMAPI->>Executor: generate_async(..., scheduling_params)
Executor->>RequestQueue: enqueue request (with scheduling_params)
Worker->>RequestQueue: process request (attention DP logic)
RequestQueue->>RequestQueue: filter & schedule requests by attention DP rank/capacity
RequestQueue-->>Worker: scheduled requests
Worker-->>Executor: scheduled requests
Executor-->>LLMAPI: results
LLMAPI-->>Client: results
Estimated code review effort🎯 4 (Complex) | ⏱️ ~40 minutes Possibly related PRs
Suggested reviewers
Note ⚡️ Unit Test Generation is now available in beta!Learn more here, or try it out under "Finishing Touches" below. ✨ Finishing Touches
🧪 Generate unit tests
🪧 TipsChatThere are 3 ways to chat with CodeRabbit:
SupportNeed help? Create a ticket on our support page for assistance with any issues or questions. Note: Be mindful of the bot's finite context window. It's strongly recommended to break down tasks such as reading entire modules into smaller chunks. For a focused discussion, use review comments to chat about specific files and their changes, instead of using the PR comments. CodeRabbit Commands (Invoked using PR comments)
Other keywords and placeholders
Documentation and Community
|
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Actionable comments posted: 0
🧹 Nitpick comments (1)
tensorrt_llm/schedule_params.py (1)
10-11
: Fix line length to comply with coding standards.The docstring line exceeds the 120-character limit. Consider reformatting for better readability.
- attention_dp_rank (int): The rank of target attention dp - attention_dp_relax (bool): Whether to allow the request to be scheduled to other attention dp for better throughput + attention_dp_rank (int): The rank of target attention dp + attention_dp_relax (bool): Whether to allow the request to be scheduled to other attention dp + for better throughput
📜 Review details
Configuration used: .coderabbit.yaml
Review profile: CHILL
Plan: Pro
📒 Files selected for processing (6)
tensorrt_llm/_torch/pyexecutor/executor_request_queue.py
(1 hunks)tensorrt_llm/executor/executor.py
(3 hunks)tensorrt_llm/executor/request.py
(3 hunks)tensorrt_llm/executor/worker.py
(1 hunks)tensorrt_llm/llmapi/llm.py
(5 hunks)tensorrt_llm/schedule_params.py
(1 hunks)
🧠 Learnings (1)
tensorrt_llm/executor/worker.py (1)
Learnt from: amitz-nv
PR: #5616
File: tensorrt_llm/executor/worker.py:375-384
Timestamp: 2025-07-17T09:01:27.374Z
Learning: In tensorrt_llm/executor/worker.py, the LoRA adapter cache optimization logic that checks is_adapter_in_cpu_cache()
and conditionally passes None for weights/config has a known race condition issue that cannot be solved with simple error handling or verification checks. This is a known limitation that requires a more comprehensive solution.
🪛 Ruff (0.12.2)
tensorrt_llm/schedule_params.py
11-11: Line too long (123 > 120)
(E501)
🧰 Additional context used
🧠 Learnings (1)
tensorrt_llm/executor/worker.py (1)
Learnt from: amitz-nv
PR: #5616
File: tensorrt_llm/executor/worker.py:375-384
Timestamp: 2025-07-17T09:01:27.374Z
Learning: In tensorrt_llm/executor/worker.py, the LoRA adapter cache optimization logic that checks is_adapter_in_cpu_cache()
and conditionally passes None for weights/config has a known race condition issue that cannot be solved with simple error handling or verification checks. This is a known limitation that requires a more comprehensive solution.
🪛 Ruff (0.12.2)
tensorrt_llm/schedule_params.py
11-11: Line too long (123 > 120)
(E501)
🔇 Additional comments (12)
tensorrt_llm/executor/request.py (1)
13-13
: Clean integration of ScheduleParams parameter.The addition of
ScheduleParams
import and integration into theGenerationRequest
constructor follows the established pattern for other optional parameters. The parameter is properly typed and stored as an instance attribute.Also applies to: 90-90, 115-115
tensorrt_llm/executor/worker.py (1)
511-513
: Proper conditional attachment of schedule parameters.The implementation correctly follows the existing pattern for PyTorch backend-specific parameters (like
py_multimodal_data
andpy_logits_post_processors
). The conditional checks ensure that schedule parameters are only attached when using the PyTorch backend and when they are actually provided.tensorrt_llm/schedule_params.py (1)
5-15
: Well-designed dataclass with optimization features.The implementation uses
slots=True
for memory efficiency andkw_only=True
to enforce keyword-only arguments, which is excellent for API clarity. The optional typing and clear docstring make the parameters self-documenting.tensorrt_llm/llmapi/llm.py (3)
33-33
: Proper import of ScheduleParams.The import is correctly placed with other parameter-related imports.
239-240
: Consistent parameter handling in synchronous generate method.The
schedule_params
parameter is properly integrated into the method signature and correctly handled in the batching logic using the_item_at
helper function, following the established pattern for other optional parameters.Also applies to: 288-288
314-314
: Clean integration in asynchronous generate method.The parameter is properly typed and forwarded through the call chain to the executor's
generate_async
method, maintaining the parameter propagation pattern throughout the system.Also applies to: 425-425
tensorrt_llm/executor/executor.py (3)
32-32
: LGTM: Clean import additionThe import of
ScheduleParams
is properly placed with other related imports.
124-124
: LGTM: Consistent parameter additionThe
schedule_params
parameter follows the established pattern of other optional parameters in the method signature.
147-148
: LGTM: Proper parameter propagationThe
schedule_params
is correctly passed to theGenerationRequest
constructor, maintaining the parameter flow through the system.tensorrt_llm/_torch/pyexecutor/executor_request_queue.py (3)
247-249
: LGTM: Clean method call delegationThe replacement of direct balancing logic with a dedicated scheduling method improves code organization and readability.
257-257
: LGTM: Consistent counter updateThe counter update properly tracks the total number of new requests processed.
264-344
: Ignore the logic-error suggestion for attention_dp_rankThe comparison
req_item.request.schedule_params.attention_dp_rank == self.dist.tp_rankis intentional—“dp” here refers to the dimension-parallel (attention) split, which is driven by the tensor-parallel rank (
tp_rank
). There is no separatedp_rank
in thedist
context, so the existing check is correct.Likely an incorrect or invalid review comment.
ac0d750
to
fb1dd7f
Compare
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Actionable comments posted: 1
🧹 Nitpick comments (2)
tensorrt_llm/_torch/pyexecutor/executor_request_queue.py (2)
312-317
: Optimize communication overhead.The padding approach for allgather creates unnecessary communication overhead as noted in the TODO. Consider using variable-length communication or a more efficient serialization approach.
Would you like me to suggest an alternative implementation that reduces communication overhead?
327-330
: Simplify the condition logic.The nested condition can be simplified for better readability.
- elif req_item.request.schedule_params is None or \ - req_item.request.schedule_params is not None and \ - req_item.request.schedule_params.attention_dp_rank is None: + elif (req_item.request.schedule_params is None or + req_item.request.schedule_params.attention_dp_rank is None):
📜 Review details
Configuration used: .coderabbit.yaml
Review profile: CHILL
Plan: Pro
📒 Files selected for processing (1)
tensorrt_llm/_torch/pyexecutor/executor_request_queue.py
(4 hunks)
⏰ Context from checks skipped due to timeout of 90000ms. You can increase the timeout in your CodeRabbit configuration to a maximum of 15 minutes (900000ms). (1)
- GitHub Check: Pre-commit Check
🔇 Additional comments (3)
tensorrt_llm/_torch/pyexecutor/executor_request_queue.py (3)
247-249
: LGTM! Clean refactoring of scheduling logic.The extraction of scheduling logic into a dedicated
_schedule_attention_dp_requests
method improves code organization and readability. The method call and parameter passing are correct.Also applies to: 257-257
406-406
: LGTM! Clean integration of pre-scheduled requests.The method signature update to accept
new_requests_cur_rank
and the subsequent merging logic correctly integrate pre-scheduled requests with the load balancing algorithm. The variable renaming tonew_scheduled_requests
improves code clarity.Also applies to: 410-410, 424-424, 457-457, 462-462, 464-465
274-275
: DP rank comparison is correct
TheExecutorRequestQueue
uses the tensor-parallel group as the attention-DP group—there is nodist.dp_rank
. Comparingschedule_params.attention_dp_rank
againstself.dist.tp_rank
and coordinating viatp_allgather
is the intended behavior.Likely an incorrect or invalid review comment.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Actionable comments posted: 1
📜 Review details
Configuration used: .coderabbit.yaml
Review profile: CHILL
Plan: Pro
📒 Files selected for processing (1)
tests/unittest/_torch/test_executor_request_queue.py
(1 hunks)
🪛 Ruff (0.12.2)
tests/unittest/_torch/test_executor_request_queue.py
835-835: Local variable req_id
is assigned to but never used
Remove assignment to unused variable req_id
(F841)
🧰 Additional context used
🪛 Ruff (0.12.2)
tests/unittest/_torch/test_executor_request_queue.py
835-835: Local variable req_id
is assigned to but never used
Remove assignment to unused variable req_id
(F841)
⏰ Context from checks skipped due to timeout of 90000ms. You can increase the timeout in your CodeRabbit configuration to a maximum of 15 minutes (900000ms). (1)
- GitHub Check: Pre-commit Check
/bot run --disable-fail-fast |
PR_Github #12549 [ run ] triggered by Bot |
PR_Github #12549 [ run ] completed with state |
8b7483b
to
537de6d
Compare
/bot run --disable-fail-fast |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Actionable comments posted: 0
♻️ Duplicate comments (2)
tensorrt_llm/_torch/pyexecutor/executor_request_queue.py (1)
305-305
: Fix potential request order reversal.Using
extendleft
will reverse the order of waiting requests. If maintaining FIFO order is important, consider usingextend
and reversing the list first, or use a different approach.- self.waiting_queue.extendleft(new_requests_cur_rank_waiting) + # Maintain FIFO order by extending from the right + self.waiting_queue.extend(reversed(new_requests_cur_rank_waiting))tests/unittest/_torch/test_executor_request_queue.py (1)
835-835
: Remove unused variable assignment.The variable
req_id
is assigned but never used in this test.- req_id = schedule_queue.enqueue_request(mock_request) + schedule_queue.enqueue_request(mock_request)
🧹 Nitpick comments (2)
tensorrt_llm/schedule_params.py (1)
11-11
: Fix line length to comply with style guide.Line exceeds the 120 character limit.
- attention_dp_relax (bool): Whether to allow the request to be scheduled to other attention dp for better throughput + attention_dp_relax (bool): Whether to allow the request to be scheduled to other attention dp for better + throughputtensorrt_llm/_torch/pyexecutor/executor_request_queue.py (1)
327-330
: Simplify complex conditional logic.The nested conditions for identifying non-scheduled requests can be simplified for better readability.
- elif req_item.request.py_schedule_params is None or \ - req_item.request.py_schedule_params is not None and \ - req_item.request.py_schedule_params.attention_dp_rank is None: + elif (req_item.request.py_schedule_params is None or + req_item.request.py_schedule_params.attention_dp_rank is None):
📜 Review details
Configuration used: .coderabbit.yaml
Review profile: CHILL
Plan: Pro
📒 Files selected for processing (8)
tensorrt_llm/_torch/pyexecutor/executor_request_queue.py
(4 hunks)tensorrt_llm/executor/executor.py
(3 hunks)tensorrt_llm/executor/request.py
(3 hunks)tensorrt_llm/executor/worker.py
(1 hunks)tensorrt_llm/llmapi/llm.py
(5 hunks)tensorrt_llm/schedule_params.py
(1 hunks)tests/unittest/_torch/test_executor_request_queue.py
(1 hunks)tests/unittest/api_stability/references/llm.yaml
(1 hunks)
🧠 Learnings (1)
tensorrt_llm/_torch/pyexecutor/executor_request_queue.py (1)
Learnt from: yiqingy0
PR: #5198
File: jenkins/mergeWaiveList.py:0-0
Timestamp: 2025-07-22T08:33:49.076Z
Learning: In the TensorRT-LLM waive list merging system, removed lines are always located at the end of the merge waive lists, which is why the mergeWaiveList.py script uses reverse traversal - it's an optimization for this specific domain constraint.
🧬 Code Graph Analysis (1)
tensorrt_llm/_torch/pyexecutor/executor_request_queue.py (1)
tensorrt_llm/_torch/distributed/communicator.py (2)
tp_allgather
(138-139)tp_size
(46-47)
🪛 Ruff (0.12.2)
tests/unittest/_torch/test_executor_request_queue.py
835-835: Local variable req_id
is assigned to but never used
Remove assignment to unused variable req_id
(F841)
tensorrt_llm/schedule_params.py
11-11: Line too long (123 > 120)
(E501)
🚧 Files skipped from review as they are similar to previous changes (4)
- tensorrt_llm/executor/request.py
- tensorrt_llm/executor/worker.py
- tensorrt_llm/executor/executor.py
- tensorrt_llm/llmapi/llm.py
🧰 Additional context used
🧠 Learnings (1)
tensorrt_llm/_torch/pyexecutor/executor_request_queue.py (1)
Learnt from: yiqingy0
PR: #5198
File: jenkins/mergeWaiveList.py:0-0
Timestamp: 2025-07-22T08:33:49.076Z
Learning: In the TensorRT-LLM waive list merging system, removed lines are always located at the end of the merge waive lists, which is why the mergeWaiveList.py script uses reverse traversal - it's an optimization for this specific domain constraint.
🧬 Code Graph Analysis (1)
tensorrt_llm/_torch/pyexecutor/executor_request_queue.py (1)
tensorrt_llm/_torch/distributed/communicator.py (2)
tp_allgather
(138-139)tp_size
(46-47)
🪛 Ruff (0.12.2)
tests/unittest/_torch/test_executor_request_queue.py
835-835: Local variable req_id
is assigned to but never used
Remove assignment to unused variable req_id
(F841)
tensorrt_llm/schedule_params.py
11-11: Line too long (123 > 120)
(E501)
🔇 Additional comments (3)
tests/unittest/api_stability/references/llm.yaml (1)
138-140
: LGTM!The addition of the optional
schedule_params
parameter togenerate_async
maintains backward compatibility while enabling the new scheduling functionality.tensorrt_llm/_torch/pyexecutor/executor_request_queue.py (1)
274-276
: Ignore incorrect renaming suggestion:.py_schedule_params
is intentional
Thepy_schedule_params
attribute is explicitly set on the executor request intensorrt_llm/executor/worker.py
:if self._is_pytorch_backend and request.schedule_params is not None: executor_request.py_schedule_params = request.schedule_paramsAll usages of
req_item.request.py_schedule_params
intensorrt_llm/_torch/pyexecutor/executor_request_queue.py
are therefore correct and should remain unchanged.Likely an incorrect or invalid review comment.
tests/unittest/_torch/test_executor_request_queue.py (1)
459-857
: Excellent test coverage!The comprehensive test suite effectively covers all major scheduling scenarios including rank filtering, capacity limits, relax flag prioritization, edge cases, and integration with the LLM API. The use of mocked distributed operations and helper functions makes the tests clear and maintainable.
PR_Github #12641 [ run ] triggered by Bot |
PR_Github #12641 [ run ] completed with state |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Actionable comments posted: 1
♻️ Duplicate comments (1)
tensorrt_llm/_torch/pyexecutor/executor_request_queue.py (1)
301-301
: Fix potential request order reversal.Using
extendleft
will reverse the order of pending requests, potentially breaking FIFO ordering when these requests are processed in the next scheduling round.- self.waiting_queue.extendleft(pending_requests) + # Maintain FIFO order by extending from the right + self.waiting_queue.extend(reversed(pending_requests))
🧹 Nitpick comments (1)
tensorrt_llm/_torch/pyexecutor/executor_request_queue.py (1)
324-331
: Optimize request removal for better performance.Using
remove()
in a loop over a potentially large list is inefficient (O(n²) complexity). Consider using a more efficient approach.- # Try to put the unscheduled requests to the target dp rank until the max_num_active_requests is reached - for req_item in unscheduled_requests[:]: - if req_item.request.py_schedule_params is not None: - target_dp_rank = req_item.request.py_schedule_params.attention_dp_rank - if self.all_ranks_num_active_requests[ - target_dp_rank] < self.max_num_active_requests: - self.all_ranks_num_active_requests[target_dp_rank] += 1 - # Ensure all ranks have the same unscheduled requests - unscheduled_requests.remove(req_item) - - # If the target dp rank is the current rank, add it to the new_requests_cur_rank - if target_dp_rank == self.dist.tp_rank: - new_requests_cur_rank.append(req_item) + # Try to put the unscheduled requests to the target dp rank until the max_num_active_requests is reached + remaining_unscheduled = [] + for req_item in unscheduled_requests: + if req_item.request.py_schedule_params is not None: + target_dp_rank = req_item.request.py_schedule_params.attention_dp_rank + if self.all_ranks_num_active_requests[ + target_dp_rank] < self.max_num_active_requests: + self.all_ranks_num_active_requests[target_dp_rank] += 1 + + # If the target dp rank is the current rank, add it to the new_requests_cur_rank + if target_dp_rank == self.dist.tp_rank: + new_requests_cur_rank.append(req_item) + else: + remaining_unscheduled.append(req_item) + else: + remaining_unscheduled.append(req_item) + unscheduled_requests[:] = remaining_unscheduled
📜 Review details
Configuration used: .coderabbit.yaml
Review profile: CHILL
Plan: Pro
📒 Files selected for processing (2)
tensorrt_llm/_torch/pyexecutor/executor_request_queue.py
(5 hunks)tensorrt_llm/executor/worker.py
(1 hunks)
✅ Files skipped from review due to trivial changes (1)
- tensorrt_llm/executor/worker.py
🧰 Additional context used
🧠 Learnings (1)
tensorrt_llm/_torch/pyexecutor/executor_request_queue.py (2)
Learnt from: amitz-nv
PR: #5616
File: tensorrt_llm/executor/worker.py:375-384
Timestamp: 2025-07-17T09:01:27.402Z
Learning: In tensorrt_llm/executor/worker.py, the LoRA adapter cache optimization logic that checks is_adapter_in_cpu_cache()
and conditionally passes None for weights/config has a known race condition issue that cannot be solved with simple error handling or verification checks. This is a known limitation that requires a more comprehensive solution.
Learnt from: yiqingy0
PR: #5198
File: jenkins/mergeWaiveList.py:0-0
Timestamp: 2025-07-22T08:33:49.109Z
Learning: In the TensorRT-LLM waive list merging system, removed lines are always located at the end of the merge waive lists, which is why the mergeWaiveList.py script uses reverse traversal - it's an optimization for this specific domain constraint.
🪛 Ruff (0.12.2)
tensorrt_llm/_torch/pyexecutor/executor_request_queue.py
316-316: Line too long (133 > 120)
(E501)
⏰ Context from checks skipped due to timeout of 90000ms. You can increase the timeout in your CodeRabbit configuration to a maximum of 15 minutes (900000ms). (1)
- GitHub Check: Pre-commit Check
🔇 Additional comments (3)
tensorrt_llm/_torch/pyexecutor/executor_request_queue.py (3)
223-225
: LGTM! Improved performance metrics accuracy.Moving the performance metrics update to the specific fetch methods ensures metrics are calculated after all filtering and scheduling steps, providing more accurate measurements.
235-241
: LGTM! Necessary refactoring for scheduling logic.Converting to an instance variable allows the scheduling methods to access and modify active request counts across ranks. The allgather operation correctly synchronizes state across all ranks.
400-459
: LGTM! Proper integration of pre-scheduled requests.The method signature change correctly allows merging of pre-scheduled attention DP requests with load-balanced requests. The heap-based balancing logic is preserved and the final merge ensures all scheduled requests are returned together.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Actionable comments posted: 6
♻️ Duplicate comments (3)
tensorrt_llm/_torch/pyexecutor/executor_request_queue.py (2)
303-303
: Fix potential request order reversal.Using
extendleft
will reverse the order of waiting requests. If maintaining FIFO order is important, consider usingextend
and reversing the list first, or use a different approach.- self.waiting_queue.extendleft(pending_requests) + # Maintain FIFO order by extending from the right + self.waiting_queue.extend(reversed(pending_requests))
315-319
: Strengthen error handling for scheduling assertion.The assertion provides good validation but could fail unexpectedly if there are bugs in the filtering logic. Consider adding more descriptive error handling.
- assert self.all_ranks_num_active_requests[target_dp_rank] <= self.max_num_active_requests, \ - f"The number of active requests on rank {target_dp_rank} is {self.all_ranks_num_active_requests[target_dp_rank]}, " \ - f"which is greater than the max_num_active_requests {self.max_num_active_requests}" + if self.all_ranks_num_active_requests[target_dp_rank] > self.max_num_active_requests: + raise RuntimeError( + f"Scheduling inconsistency: rank {target_dp_rank} has " + f"{self.all_ranks_num_active_requests[target_dp_rank]} active requests, " + f"exceeding limit of {self.max_num_active_requests}. " + f"This indicates a bug in the filtering logic." + )tests/unittest/_torch/test_executor_request_queue.py (1)
698-699
: Remove unused variable assignment.The variable
result
is assigned but never used in this test.- result = attention_dp_queue._schedule_attention_dp_requests( - scheduled_requests, unscheduled_requests) + attention_dp_queue._schedule_attention_dp_requests( + scheduled_requests, unscheduled_requests)
🧹 Nitpick comments (1)
tests/unittest/_torch/test_executor_request_queue.py (1)
726-728
: Improve test documentation.The comment doesn't accurately describe what's being tested. The test is checking that an assertion is raised when trying to schedule a request to a rank that's already over capacity.
- # Should raise assertion error because we're trying to schedule a request when rank is at capacity + # Should raise assertion error because rank 0 already exceeds max capacity (9 > 8)
📜 Review details
Configuration used: .coderabbit.yaml
Review profile: CHILL
Plan: Pro
📒 Files selected for processing (2)
tensorrt_llm/_torch/pyexecutor/executor_request_queue.py
(5 hunks)tests/unittest/_torch/test_executor_request_queue.py
(1 hunks)
🧰 Additional context used
🧠 Learnings (1)
tensorrt_llm/_torch/pyexecutor/executor_request_queue.py (2)
Learnt from: amitz-nv
PR: #5616
File: tensorrt_llm/executor/worker.py:375-384
Timestamp: 2025-07-17T09:01:27.402Z
Learning: In tensorrt_llm/executor/worker.py, the LoRA adapter cache optimization logic that checks is_adapter_in_cpu_cache()
and conditionally passes None for weights/config has a known race condition issue that cannot be solved with simple error handling or verification checks. This is a known limitation that requires a more comprehensive solution.
Learnt from: yiqingy0
PR: #5198
File: jenkins/mergeWaiveList.py:0-0
Timestamp: 2025-07-22T08:33:49.109Z
Learning: In the TensorRT-LLM waive list merging system, removed lines are always located at the end of the merge waive lists, which is why the mergeWaiveList.py script uses reverse traversal - it's an optimization for this specific domain constraint.
🪛 Ruff (0.12.2)
tests/unittest/_torch/test_executor_request_queue.py
677-677: Local variable result
is assigned to but never used
Remove assignment to unused variable result
(F841)
698-698: Local variable result
is assigned to but never used
Remove assignment to unused variable result
(F841)
tensorrt_llm/_torch/pyexecutor/executor_request_queue.py
318-318: Line too long (133 > 120)
(E501)
🔇 Additional comments (1)
tensorrt_llm/_torch/pyexecutor/executor_request_queue.py (1)
460-460
: Consider the order when merging scheduled requests.The order of extending
new_scheduled_requests
withnew_requests_cur_rank
may affect processing order. Ensure this aligns with the expected behavior - typically, pre-scheduled requests should be processed before newly balanced ones.- new_scheduled_requests.extend(new_requests_cur_rank) + # Process pre-scheduled requests first, then newly balanced ones + result = new_requests_cur_rank + new_scheduled_requests + return result
d01a546
to
f783e2c
Compare
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Actionable comments posted: 2
🔭 Outside diff range comments (1)
tensorrt_llm/schedule_params.py (1)
1-16
: Add NVIDIA copyright header.All TensorRT-LLM source files must contain an NVIDIA copyright header that includes the current year.
Add the copyright header at the beginning of the file:
+# SPDX-FileCopyrightText: Copyright (c) 2025 NVIDIA CORPORATION & AFFILIATES. All rights reserved. +# SPDX-License-Identifier: Apache-2.0 +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + from dataclasses import dataclass from typing import Optional
♻️ Duplicate comments (8)
tensorrt_llm/_torch/pyexecutor/executor_request_queue.py (5)
249-254
: Fix duplicate performance metrics update.The performance metrics are updated twice for the same requests, causing inflated latency measurements as noted in past reviews.
Remove the duplicate metric updates:
- # Update performance metrics - # TODO: Check whether we should update the performance metrics for all ranks - if self.enable_iter_perf_stats and self.dist.rank == 0: - self._update_new_active_requests_queue_latency(scheduled_requests) - self._update_new_active_requests_queue_latency(unscheduled_requests) -
280-281
: Fix misleading comment about copying.The comment says "avoid modifying the original" but the code doesn't create a proper copy.
- # Create a copy to avoid modifying the original all_ranks_num_active_requests - all_ranks_num_active_requests = self.all_ranks_num_active_requests.copy() + # Create a copy to avoid modifying the original all_ranks_num_active_requests + all_ranks_num_active_requests = self.all_ranks_num_active_requests.copy()
302-302
: Fix potential request order reversal with extendleft.Using
extendleft
reverses the order of pending requests, breaking FIFO order.- self.waiting_queue.extendleft(pending_requests) + # Maintain FIFO order by extending from the right + self.waiting_queue.extend(reversed(pending_requests))
315-319
: Replace assertion with proper error handling.The assertion could fail unexpectedly and should be replaced with descriptive error handling.
- assert self.all_ranks_num_active_requests[target_dp_rank] <= self.max_num_active_requests, \ - f"The number of active requests on rank {target_dp_rank}" \ - f"is {self.all_ranks_num_active_requests[target_dp_rank]}, " \ - f"which is greater than the max_num_active_requests {self.max_num_active_requests}" + if self.all_ranks_num_active_requests[target_dp_rank] > self.max_num_active_requests: + raise RuntimeError( + f"Scheduling inconsistency: rank {target_dp_rank} has " + f"{self.all_ranks_num_active_requests[target_dp_rank]} active requests, " + f"exceeding limit of {self.max_num_active_requests}. " + f"This indicates a bug in the filtering logic." + )
233-236
: Consider using local variable for active request counts.The instance variable
self.all_ranks_num_active_requests
could lead to side effects if methods are called in different orders.Consider passing the active request counts as a parameter between methods instead of storing as an instance variable to avoid potential side effects and improve thread safety.
tests/unittest/_torch/test_executor_request_queue.py (3)
677-679
: Remove unused variable assignment.The variable
result
is assigned but never used in this test.- result = attention_dp_queue._schedule_attention_dp_requests( - scheduled_requests, unscheduled_requests) + attention_dp_queue._schedule_attention_dp_requests( + scheduled_requests, unscheduled_requests)
764-766
: Add assertion to verify scheduling result.The test schedules requests but doesn't verify the final result.
result = attention_dp_queue._schedule_attention_dp_requests( scheduled, unscheduled) + + # Verify that all requests were scheduled successfully + assert len(unscheduled) == 0 # All should have been scheduled + assert req_no_params not in result # Only requests for current rank
698-700
: Remove another unused variable assignment.The variable
result
is assigned but never used in this test.- result = attention_dp_queue._schedule_attention_dp_requests( - scheduled_requests, unscheduled_requests) + attention_dp_queue._schedule_attention_dp_requests( + scheduled_requests, unscheduled_requests)
🧹 Nitpick comments (4)
tensorrt_llm/schedule_params.py (1)
7-12
: Improve docstring format and clarity.The docstring could benefit from following Google style more closely and providing clearer parameter descriptions.
- """Schedule parameters. - - Args: - attention_dp_rank (int): The rank of target attention dp - attention_dp_relax (bool): Whether to allow the request to be scheduled to other attention dp for better throughput - """ + """Parameters for controlling attention data parallel scheduling behavior. + + Args: + attention_dp_rank: The target rank for attention data parallelism. If None, the request + can be scheduled to any available rank. + attention_dp_relax: Whether to allow relaxed scheduling. If True, the request can be + scheduled to other attention dp ranks for better throughput when the target rank + is at capacity. If None or False, strict scheduling is enforced. + """tensorrt_llm/_torch/pyexecutor/executor_request_queue.py (1)
265-266
: Clarify counter semantics for filtered requests.The counter
num_fetch_requests
includes all new requests but some may be filtered out and put back to waiting queue. Consider tracking scheduled vs unscheduled separately.# Update counters self.num_fetch_requests += len(new_requests) self.num_fetch_requests_cur_rank += len(new_requests_cur_rank) + # TODO: Consider tracking num_scheduled_requests and num_pending_requests separately
tests/unittest/_torch/test_executor_request_queue.py (2)
319-805
: Consider adding edge case tests.The test coverage is comprehensive but could benefit from additional edge cases.
Consider adding tests for:
- Concurrent access scenarios with multiple threads
- Very large request queues to test scalability
- Invalid attention_dp_rank values (negative, >= tp_size)
- Mixed relax=True/False requests for the same rank
- Boundary conditions when all ranks are at capacity
349-351
: Remove hardcoded initialization of all_ranks_num_active_requests.The test fixture hardcodes
all_ranks_num_active_requests
which might not reflect real scenarios where this is populated via allgather.Consider creating a helper method to properly initialize this based on test requirements:
- # Initialize all_ranks_num_active_requests - queue.all_ranks_num_active_requests = [2, 1, 3, 0] # 4 ranks + # Initialize via a helper method that simulates allgather + queue.all_ranks_num_active_requests = [] # Will be set per test
📜 Review details
Configuration used: .coderabbit.yaml
Review profile: CHILL
Plan: Pro
📒 Files selected for processing (7)
tensorrt_llm/_torch/pyexecutor/executor_request_queue.py
(5 hunks)tensorrt_llm/executor/executor.py
(3 hunks)tensorrt_llm/executor/request.py
(3 hunks)tensorrt_llm/executor/worker.py
(1 hunks)tensorrt_llm/llmapi/llm.py
(5 hunks)tensorrt_llm/schedule_params.py
(1 hunks)tests/unittest/_torch/test_executor_request_queue.py
(1 hunks)
🚧 Files skipped from review as they are similar to previous changes (4)
- tensorrt_llm/executor/worker.py
- tensorrt_llm/executor/request.py
- tensorrt_llm/llmapi/llm.py
- tensorrt_llm/executor/executor.py
🧰 Additional context used
📓 Path-based instructions (2)
**/*.py
📄 CodeRabbit Inference Engine (CODING_GUIDELINES.md)
**/*.py
: Python code should conform to Python 3.8+.
Indent Python code with 4 spaces. Do not use tabs.
Always maintain the namespace when importing in Python, even if only one class or function from a module is used.
Python filenames should use snake_case (e.g., some_file.py).
Python classes should use PascalCase (e.g., class SomeClass).
Python functions and methods should use snake_case (e.g., def my_awesome_function():).
Python local variables should use snake_case. Prefix k for variable names that start with a number (e.g., k_99th_percentile).
Python global variables should use upper snake_case and prefix G (e.g., G_MY_GLOBAL).
Python constants should use upper snake_case (e.g., MY_CONSTANT).
Avoid shadowing variables declared in an outer scope in Python.
Initialize all externally visible members of a Python class in the constructor.
For interfaces that may be used outside a Python file, prefer docstrings over comments.
Comments in Python should be reserved for code within a function, or interfaces that are local to a file.
Use Google style docstrings for Python classes and functions, which can be parsed by Sphinx.
Attributes and variables in Python can be documented inline; attribute docstrings will be rendered under the docstring for the class.
Avoid using reflection in Python when functionality can be easily achieved without it.
When using try-except blocks in Python, limit the except to the smallest set of errors possible.
When using try-except blocks to handle multiple possible variable types in Python, keep the body of the try as small as possible, using the else block to implement the logic.
Files:
tensorrt_llm/_torch/pyexecutor/executor_request_queue.py
tests/unittest/_torch/test_executor_request_queue.py
tensorrt_llm/schedule_params.py
**/*.{cpp,h,hpp,cc,cxx,cu,py}
📄 CodeRabbit Inference Engine (CODING_GUIDELINES.md)
All TensorRT-LLM Open Source Software code should contain an NVIDIA copyright header that includes the current year. This includes .cpp, .h, .cu, .py, and any other source files which are compiled or interpreted.
Files:
tensorrt_llm/_torch/pyexecutor/executor_request_queue.py
tests/unittest/_torch/test_executor_request_queue.py
tensorrt_llm/schedule_params.py
🪛 Ruff (0.12.2)
tests/unittest/_torch/test_executor_request_queue.py
677-677: Local variable result
is assigned to but never used
Remove assignment to unused variable result
(F841)
698-698: Local variable result
is assigned to but never used
Remove assignment to unused variable result
(F841)
tensorrt_llm/schedule_params.py
11-11: Line too long (123 > 120)
(E501)
⏰ Context from checks skipped due to timeout of 90000ms. You can increase the timeout in your CodeRabbit configuration to a maximum of 15 minutes (900000ms). (1)
- GitHub Check: Pre-commit Check
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Actionable comments posted: 1
♻️ Duplicate comments (2)
tensorrt_llm/_torch/pyexecutor/executor_request_queue.py (2)
267-270
: Use local variable instead of instance variable.Storing active request counts in an instance variable can lead to unexpected side effects if methods are called in different orders.
# Get active request counts across all ranks - self.all_ranks_num_active_requests = [] + all_ranks_num_active_requests = [] responses_list = self.dist.tp_allgather(num_active_requests) for num_active_requests in responses_list: - self.all_ranks_num_active_requests.append(num_active_requests) + all_ranks_num_active_requests.append(num_active_requests)Then pass
all_ranks_num_active_requests
as a parameter to methods that need it.
303-309
: Fix incorrect sorting for relax mode prioritization.The comment says "Prioritize the requests that are not in relax mode" but
reverse=True
actually puts relax=True requests first.- # Prioritize the requests that are not in relax mode + # Prioritize the requests that are not in relax mode (relax=False first) def get_relax_value(req_item): if req_item.request.py_scheduling_params is None: return True return req_item.request.py_scheduling_params.attention_dp_relax - new_requests = sorted(new_requests, key=get_relax_value, reverse=True) + new_requests = sorted(new_requests, key=get_relax_value)
🧹 Nitpick comments (1)
tensorrt_llm/scheduling_params.py (1)
7-12
: Fix line length and improve docstring format.The docstring line exceeds the 120 character limit and should follow Google style format for better consistency.
class SchedulingParams: - """Schedule parameters. - - Args: - attention_dp_rank (int): The rank of target attention dp - attention_dp_relax (bool): Whether to allow the request to be scheduled to other attention dp for better throughput + """Scheduling parameters for attention data parallelism. + + Attributes: + attention_dp_rank: The rank of target attention dp. + attention_dp_relax: Whether to allow the request to be scheduled to other + attention dp ranks for better throughput. """
📜 Review details
Configuration used: .coderabbit.yaml
Review profile: CHILL
Plan: Pro
📒 Files selected for processing (7)
tensorrt_llm/_torch/pyexecutor/executor_request_queue.py
(8 hunks)tensorrt_llm/executor/executor.py
(3 hunks)tensorrt_llm/executor/request.py
(3 hunks)tensorrt_llm/executor/worker.py
(1 hunks)tensorrt_llm/llmapi/llm.py
(5 hunks)tensorrt_llm/scheduling_params.py
(1 hunks)tests/unittest/api_stability/references/llm.yaml
(2 hunks)
🚧 Files skipped from review as they are similar to previous changes (4)
- tensorrt_llm/executor/request.py
- tensorrt_llm/executor/worker.py
- tensorrt_llm/executor/executor.py
- tensorrt_llm/llmapi/llm.py
🧰 Additional context used
📓 Path-based instructions (2)
**/*.py
📄 CodeRabbit Inference Engine (CODING_GUIDELINES.md)
**/*.py
: Python code should conform to Python 3.8+.
Indent Python code with 4 spaces. Do not use tabs.
Always maintain the namespace when importing in Python, even if only one class or function from a module is used.
Python filenames should use snake_case (e.g., some_file.py).
Python classes should use PascalCase (e.g., class SomeClass).
Python functions and methods should use snake_case (e.g., def my_awesome_function():).
Python local variables should use snake_case, and prefix k for variable names that start with a number (e.g., k_99th_percentile).
Python global variables should use upper snake_case and prefix G (e.g., G_MY_GLOBAL).
Python constants should use upper snake_case (e.g., MY_CONSTANT).
Avoid shadowing variables declared in an outer scope in Python.
Initialize all externally visible members of a Python class in the constructor.
For interfaces that may be used outside a Python file, prefer docstrings over comments.
Comments in Python should be reserved for code within a function, or interfaces that are local to a file.
Use Google style docstrings for Python classes and functions, which can be parsed by Sphinx.
Attributes and variables in Python can be documented inline; attribute docstrings will be rendered under the docstring for the class.
Avoid using reflection in Python when functionality can be easily achieved without it.
When using try-except blocks in Python, limit the except to the smallest set of errors possible.
When using try-except blocks to handle multiple possible variable types in Python, keep the body of the try as small as possible, using the else block to implement the logic.
Files:
tensorrt_llm/_torch/pyexecutor/executor_request_queue.py
tensorrt_llm/scheduling_params.py
**/*.{cpp,h,hpp,cc,cxx,cu,py}
📄 CodeRabbit Inference Engine (CODING_GUIDELINES.md)
All TensorRT-LLM Open Source Software code should contain an NVIDIA copyright header that includes the current year. This includes .cpp, .h, .cu, .py, and any other source files which are compiled or interpreted.
Files:
tensorrt_llm/_torch/pyexecutor/executor_request_queue.py
tensorrt_llm/scheduling_params.py
🧠 Learnings (1)
tensorrt_llm/_torch/pyexecutor/executor_request_queue.py (2)
Learnt from: yiqingy0
PR: #5198
File: jenkins/mergeWaiveList.py:0-0
Timestamp: 2025-07-22T08:33:49.109Z
Learning: In the TensorRT-LLM waive list merging system, removed lines are always located at the end of the merge waive lists, which is why the mergeWaiveList.py script uses reverse traversal - it's an optimization for this specific domain constraint.
Learnt from: amitz-nv
PR: #5616
File: tensorrt_llm/executor/worker.py:375-384
Timestamp: 2025-07-17T09:01:27.402Z
Learning: In tensorrt_llm/executor/worker.py, the LoRA adapter cache optimization logic that checks is_adapter_in_cpu_cache()
and conditionally passes None for weights/config has a known race condition issue that cannot be solved with simple error handling or verification checks. This is a known limitation that requires a more comprehensive solution.
🪛 Ruff (0.12.2)
tensorrt_llm/scheduling_params.py
11-11: Line too long (123 > 120)
(E501)
⏰ Context from checks skipped due to timeout of 90000ms. You can increase the timeout in your CodeRabbit configuration to a maximum of 15 minutes (900000ms). (1)
- GitHub Check: Pre-commit Check
🔇 Additional comments (4)
tests/unittest/api_stability/references/llm.yaml (1)
129-131
: LGTM! API changes maintain backward compatibility.The addition of the optional
scheduling_params
parameter to bothgenerate
andgenerate_async
methods is well-designed with proper type annotations and default values, ensuring existing code continues to work without modifications.Also applies to: 141-143
tensorrt_llm/_torch/pyexecutor/executor_request_queue.py (3)
86-119
: Well-implemented attention DP filtering logic.The method correctly handles attention DP request filtering while maintaining FIFO order with
extendleft(reversed())
. Good use ofcopy()
to avoid side effects on the shared state.
198-236
: Clean integration of attention DP flag.The method correctly propagates the
enable_attention_dp
flag to the waiting queue processing logic.
390-442
: Good update to merge pre-scheduled requests.The method signature change and logic update correctly handle merging requests that were already scheduled to the current rank with newly balanced requests.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Actionable comments posted: 2
🧹 Nitpick comments (2)
tests/unittest/_torch/test_executor_request_queue.py (2)
339-351
: Consider using a more descriptive initialization pattern.The fixture is well-structured, but the hardcoded active request counts could be more self-documenting.
def attention_dp_queue(mock_dist_attention_dp): """Create an ExecutorRequestQueue instance for attention DP testing.""" queue = ExecutorRequestQueue(dist=mock_dist_attention_dp, enable_attention_dp=True, max_batch_size=4, max_beam_width=2, max_num_active_requests=8, enable_iter_perf_stats=True, is_disaggregated=False) - # Initialize all_ranks_num_active_requests - queue.all_ranks_num_active_requests = [2, 1, 3, 0] # 4 ranks + # Initialize all_ranks_num_active_requests with test values + # Rank 0: 2 active, Rank 1: 1 active, Rank 2: 3 active, Rank 3: 0 active + queue.all_ranks_num_active_requests = [2, 1, 3, 0] return queue
667-696
: Fix typo in variable name.attention_dp_queue.all_ranks_num_active_requests = [5, 6, 3, 7] attention_dp_queue.waiting_queue.extend(req_list) - avaiable_active_requests = attention_dp_queue.max_num_active_requests * 4 - sum( + available_active_requests = attention_dp_queue.max_num_active_requests * 4 - sum( attention_dp_queue.all_ranks_num_active_requests) result = attention_dp_queue._get_from_waiting_queue( attention_dp_queue.waiting_queue, - avaiable_active_requests, + available_active_requests, enable_attention_dp=True) - assert len(result) == avaiable_active_requests + assert len(result) == available_active_requests
📜 Review details
Configuration used: .coderabbit.yaml
Review profile: CHILL
Plan: Pro
📒 Files selected for processing (1)
tests/unittest/_torch/test_executor_request_queue.py
(3 hunks)
🧰 Additional context used
📓 Path-based instructions (2)
**/*.py
📄 CodeRabbit Inference Engine (CODING_GUIDELINES.md)
**/*.py
: Python code should conform to Python 3.8+.
Indent Python code with 4 spaces. Do not use tabs.
Always maintain the namespace when importing in Python, even if only one class or function from a module is used.
Python filenames should use snake_case (e.g., some_file.py).
Python classes should use PascalCase (e.g., class SomeClass).
Python functions and methods should use snake_case (e.g., def my_awesome_function():).
Python local variables should use snake_case, and prefix k for variable names that start with a number (e.g., k_99th_percentile).
Python global variables should use upper snake_case and prefix G (e.g., G_MY_GLOBAL).
Python constants should use upper snake_case (e.g., MY_CONSTANT).
Avoid shadowing variables declared in an outer scope in Python.
Initialize all externally visible members of a Python class in the constructor.
For interfaces that may be used outside a Python file, prefer docstrings over comments.
Comments in Python should be reserved for code within a function, or interfaces that are local to a file.
Use Google style docstrings for Python classes and functions, which can be parsed by Sphinx.
Attributes and variables in Python can be documented inline; attribute docstrings will be rendered under the docstring for the class.
Avoid using reflection in Python when functionality can be easily achieved without it.
When using try-except blocks in Python, limit the except to the smallest set of errors possible.
When using try-except blocks to handle multiple possible variable types in Python, keep the body of the try as small as possible, using the else block to implement the logic.
Files:
tests/unittest/_torch/test_executor_request_queue.py
**/*.{cpp,h,hpp,cc,cxx,cu,py}
📄 CodeRabbit Inference Engine (CODING_GUIDELINES.md)
All TensorRT-LLM Open Source Software code should contain an NVIDIA copyright header that includes the current year. This includes .cpp, .h, .cu, .py, and any other source files which are compiled or interpreted.
Files:
tests/unittest/_torch/test_executor_request_queue.py
🪛 Ruff (0.12.2)
tests/unittest/_torch/test_executor_request_queue.py
521-521: Local variable result
is assigned to but never used
Remove assignment to unused variable result
(F841)
639-640: Avoid equality comparisons to True
; use ...:
for truth checks
Replace comparison
(E712)
646-647: Avoid equality comparisons to True
; use ...:
for truth checks
Replace comparison
(E712)
654-655: Avoid equality comparisons to True
; use ...:
for truth checks
Replace comparison
(E712)
663-664: Avoid equality comparisons to False
; use not ...:
for false checks
Replace comparison
(E712)
⏰ Context from checks skipped due to timeout of 90000ms. You can increase the timeout in your CodeRabbit configuration to a maximum of 15 minutes (900000ms). (1)
- GitHub Check: Pre-commit Check
🔇 Additional comments (4)
tests/unittest/_torch/test_executor_request_queue.py (4)
319-336
: LGTM!The fixture properly sets up a mock distributed environment with 4 tensor parallel ranks for testing attention DP functionality.
354-373
: LGTM!The helper function properly creates mock requests with optional scheduling parameters, handling both cases when parameters are provided and when they're None.
377-396
: LGTM!The test correctly verifies that requests scheduled for the current rank are returned and the active request count is properly updated.
398-417
: LGTM!The test properly verifies that requests scheduled for other ranks are not returned but their active counts are updated.
@Superjomn this is touching the LLM API, could you please review? Thanks. |
Signed-off-by: Shunkang <[email protected]>
Signed-off-by: Shunkang <[email protected]>
Signed-off-by: Shunkang <[email protected]>
Signed-off-by: Patrice Castonguay <[email protected]>
Signed-off-by: Patrice Castonguay <[email protected]>
Signed-off-by: Patrice Castonguay <[email protected]>
Signed-off-by: Shunkang <[email protected]>
60589e0
to
27f16fb
Compare
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Actionable comments posted: 2
🔭 Outside diff range comments (1)
tensorrt_llm/_torch/pyexecutor/executor_request_queue.py (1)
276-311
: Fix undefined variable error.The variable
num_new_requests_all_ranks
is used on line 306 but is never defined in this method.Either remove these lines if the counters are not needed:
- # Update counters - self.num_fetch_requests += num_new_requests_all_ranks - self.num_fetch_requests_cur_rank += len(new_requests_cur_rank)Or define the variable before using it:
# Schedule attention dp requests all_ranks_new_requests = self._schedule_attention_dp_requests( new_requests, all_ranks_num_active_requests) new_requests_cur_rank = all_ranks_new_requests[self.dist.tp_rank] + + # Calculate total number of new requests across all ranks + num_new_requests_all_ranks = sum(len(reqs) for reqs in all_ranks_new_requests.values()) # Update performance metrics
♻️ Duplicate comments (5)
tensorrt_llm/_torch/pyexecutor/executor_request_queue.py (4)
136-136
: Fix incorrect attribute name.The attribute should be
py_schedule_params
instead ofpy_scheduling_params
.- scheduling_params = req_item.request.py_scheduling_params + scheduling_params = req_item.request.py_schedule_params
326-328
: Fix incorrect attribute name in scheduling method.The attribute should be
py_schedule_params
instead ofpy_scheduling_params
.- if req_item.request.py_scheduling_params is None: + if req_item.request.py_schedule_params is None: return True - return req_item.request.py_scheduling_params.attention_dp_relax + return req_item.request.py_schedule_params.attention_dp_relax
330-330
: Fix sorting logic to match the comment.The comment says "Prioritize the requests that are not in relax mode" but the current sorting puts relax=True first.
- new_requests = sorted(new_requests, key=get_relax_value, reverse=True) + new_requests = sorted(new_requests, key=get_relax_value)
337-337
: Fix another incorrect attribute name reference.- target_dp_rank = req_item.request.py_scheduling_params.attention_dp_rank + target_dp_rank = req_item.request.py_schedule_params.attention_dp_ranktests/unittest/_torch/test_executor_request_queue.py (1)
672-697
: Simplify boolean comparisons in assertions.Avoid explicit comparisons to
True
andFalse
.- assert attention_dp_queue._can_process_attention_dp_request( - req_no_params, [0, 0, 0, 0]) == True + assert attention_dp_queue._can_process_attention_dp_request( + req_no_params, [0, 0, 0, 0]) req_relax = RequestQueueItem( 2, create_mock_request_with_py_schedule_params(attention_dp_rank=0, attention_dp_relax=True)) - assert attention_dp_queue._can_process_attention_dp_request( - req_relax, [0, 0, 0, 0]) == True + assert attention_dp_queue._can_process_attention_dp_request( + req_relax, [0, 0, 0, 0]) req_target = RequestQueueItem( 3, create_mock_request_with_py_schedule_params(attention_dp_rank=1, attention_dp_relax=False)) all_ranks = [0, 0, 0, 0] - assert attention_dp_queue._can_process_attention_dp_request( - req_target, all_ranks) == True + assert attention_dp_queue._can_process_attention_dp_request( + req_target, all_ranks) assert all_ranks[1] == 1 req_no_capacity = RequestQueueItem( 4, create_mock_request_with_py_schedule_params(attention_dp_rank=0, attention_dp_relax=False)) all_ranks_full = [8, 0, 0, 0] # Rank 0 is at capacity - assert attention_dp_queue._can_process_attention_dp_request( - req_no_capacity, all_ranks_full) == False + assert not attention_dp_queue._can_process_attention_dp_request( + req_no_capacity, all_ranks_full)
🧹 Nitpick comments (1)
tensorrt_llm/scheduling_params.py (1)
11-11
: Fix line length to comply with 120 character limit.The docstring line exceeds the maximum line length of 120 characters.
- attention_dp_relax (bool): Whether to allow the request to be scheduled to other attention dp for better throughput + attention_dp_relax (bool): Whether to allow the request to be scheduled to other attention dp + for better throughput
📜 Review details
Configuration used: .coderabbit.yaml
Review profile: CHILL
Plan: Pro
📒 Files selected for processing (10)
cpp/include/tensorrt_llm/common/tllmException.h
(2 hunks)cpp/tensorrt_llm/common/tllmException.cpp
(1 hunks)tensorrt_llm/_torch/pyexecutor/executor_request_queue.py
(9 hunks)tensorrt_llm/executor/executor.py
(3 hunks)tensorrt_llm/executor/request.py
(3 hunks)tensorrt_llm/executor/worker.py
(1 hunks)tensorrt_llm/llmapi/llm.py
(5 hunks)tensorrt_llm/scheduling_params.py
(1 hunks)tests/unittest/_torch/test_executor_request_queue.py
(3 hunks)tests/unittest/api_stability/references/llm.yaml
(2 hunks)
🚧 Files skipped from review as they are similar to previous changes (5)
- tests/unittest/api_stability/references/llm.yaml
- tensorrt_llm/executor/worker.py
- tensorrt_llm/llmapi/llm.py
- tensorrt_llm/executor/executor.py
- tensorrt_llm/executor/request.py
🧰 Additional context used
📓 Path-based instructions (4)
**/*.{cpp,h,hpp,cc,cxx}
📄 CodeRabbit Inference Engine (CODING_GUIDELINES.md)
**/*.{cpp,h,hpp,cc,cxx}
: Closing braces of namespaces should have a comment saying the namespace it closes (e.g., } // namespace foo).
Prefer const or constexpr variables over #defines whenever possible.
A variable that is not modified after its initialization should be declared as const.
Except 0 (only used in comparison for checking signness/existence/emptiness) and nullptr, true, false, all other literals should only be used for variable initialization.
Use the Allman indentation style for braces.
Put the semicolon for an empty for or while loop in a new line.
The statement forming the body of a switch, while, do .. while or for statement shall be a compound statement (use brace-delimited statements).
If and else should always be followed by brace-delimited statements, even if empty or a single statement.
C++ filenames should use camel case with first letter lowercase (e.g., thisIsAFilename.cpp) and must be case-insensitive unique within a compilation target.
All types (including class names) are camel case with uppercase first letter (e.g., FooBarClass).
Local variables, methods, and namespaces use camel case with first letter lowercase (e.g., localFooBar).
Non-magic-number global variables that are non-static and not defined in anonymous namespace use camel case prefixed by 'g' (e.g., gDontUseGlobalFoos).
Non-magic-number global variables that are static or defined in an anonymous namespace use camel case prefixed by 's' (e.g., sMutableStaticGlobal).
Locally visible static variable uses camel case with lowercase prefix 's' as the first letter (e.g., static std::once_flag sFlag;).
Class member variables use camel case prefixed with 'm' (e.g., mNbFooValues). Public member variables do not require the 'm' prefix but it is encouraged for clarity.
Enumerations, global constants, static constants at class-scope, and function-scope magic-number/literal constants are uppercase snakecase with prefix 'k' (e.g., kDIGIT_NUM).
If you must use macros, follow uppercase snak...
Files:
cpp/tensorrt_llm/common/tllmException.cpp
cpp/include/tensorrt_llm/common/tllmException.h
**/*.{cpp,h,hpp,cc,cxx,cu,py}
📄 CodeRabbit Inference Engine (CODING_GUIDELINES.md)
All code must contain an NVIDIA copyright header that includes the current year. This includes .cpp, .h, .cu, .py, and any other source files which are compiled or interpreted.
Files:
cpp/tensorrt_llm/common/tllmException.cpp
cpp/include/tensorrt_llm/common/tllmException.h
tensorrt_llm/_torch/pyexecutor/executor_request_queue.py
tests/unittest/_torch/test_executor_request_queue.py
tensorrt_llm/scheduling_params.py
**/*.{h,hpp}
📄 CodeRabbit Inference Engine (CODING_GUIDELINES.md)
Use a preprocessor guard in header files. The guard name must have prefix TRTLLM_ followed by the filename, all in caps, and no trailing underscore.
Files:
cpp/include/tensorrt_llm/common/tllmException.h
**/*.py
📄 CodeRabbit Inference Engine (CODING_GUIDELINES.md)
**/*.py
: The code developed for TensorRT-LLM should conform to Python 3.8+.
Indent Python code with 4 spaces. Do not use tabs.
Always maintain the namespace when importing in Python, even if only one class or function from a module is used.
Python filenames should use snake_case (e.g., some_file.py).
Python classes should use PascalCase (e.g., class SomeClass).
Python functions and methods should use snake_case (e.g., def my_awesome_function():).
Python local variables should use snake_case. Prefix k for variable names that start with a number (e.g., k_99th_percentile).
Python global variables should use upper snake_case and prefix G (e.g., G_MY_GLOBAL).
Python constants should use upper snake_case (e.g., MY_CONSTANT).
Avoid shadowing variables declared in an outer scope in Python.
Initialize all externally visible members of a class in the constructor in Python.
For interfaces that may be used outside a file, prefer docstrings over comments in Python.
Comments in Python should be reserved for code within a function, or interfaces that are local to a file.
Use Google style docstrings for classes and functions in Python, which can be parsed by Sphinx.
Attributes and variables in Python can be documented inline; attribute docstrings will be rendered under the docstring for the class.
Avoid using reflection in Python when functionality can be easily achieved without it.
When using try-except blocks in Python, limit the except to the smallest set of errors possible.
When using try-except blocks to handle multiple possible variable types in Python, keep the body of the try as small as possible, using the else block to implement the logic.
Files:
tensorrt_llm/_torch/pyexecutor/executor_request_queue.py
tests/unittest/_torch/test_executor_request_queue.py
tensorrt_llm/scheduling_params.py
🧠 Learnings (1)
tensorrt_llm/_torch/pyexecutor/executor_request_queue.py (3)
Learnt from: yiqingy0
PR: #5198
File: jenkins/mergeWaiveList.py:0-0
Timestamp: 2025-07-22T08:33:49.109Z
Learning: In the TensorRT-LLM waive list merging system, removed lines are always located at the end of the merge waive lists, which is why the mergeWaiveList.py script uses reverse traversal - it's an optimization for this specific domain constraint.
Learnt from: amitz-nv
PR: #5616
File: tensorrt_llm/executor/worker.py:375-384
Timestamp: 2025-07-17T09:01:27.402Z
Learning: In tensorrt_llm/executor/worker.py, the LoRA adapter cache optimization logic that checks is_adapter_in_cpu_cache()
and conditionally passes None for weights/config has a known race condition issue that cannot be solved with simple error handling or verification checks. This is a known limitation that requires a more comprehensive solution.
Learnt from: CR
PR: NVIDIA/TensorRT-LLM#0
File: CODING_GUIDELINES.md:0-0
Timestamp: 2025-07-29T23:14:44.448Z
Learning: Applies to **/*.py : For interfaces that may be used outside a file, prefer docstrings over comments in Python.
🧬 Code Graph Analysis (2)
cpp/tensorrt_llm/common/tllmException.cpp (1)
cpp/include/tensorrt_llm/common/tllmException.h (1)
RequestSpecificException
(63-78)
tensorrt_llm/_torch/pyexecutor/executor_request_queue.py (3)
tests/unittest/_torch/test_executor_request_queue.py (1)
all_ranks_num_active_requests
(354-355)tensorrt_llm/_torch/distributed/communicator.py (1)
tp_size
(46-47)tensorrt_llm/functional.py (1)
sum
(3253-3275)
🪛 Ruff (0.12.2)
tests/unittest/_torch/test_executor_request_queue.py
672-673: Avoid equality comparisons to True
; use ...:
for truth checks
Replace comparison
(E712)
679-680: Avoid equality comparisons to True
; use ...:
for truth checks
Replace comparison
(E712)
687-688: Avoid equality comparisons to True
; use ...:
for truth checks
Replace comparison
(E712)
696-697: Avoid equality comparisons to False
; use not ...:
for false checks
Replace comparison
(E712)
tensorrt_llm/scheduling_params.py
11-11: Line too long (123 > 120)
(E501)
⏰ Context from checks skipped due to timeout of 90000ms. You can increase the timeout in your CodeRabbit configuration to a maximum of 15 minutes (900000ms). (1)
- GitHub Check: Pre-commit Check
🔇 Additional comments (1)
cpp/include/tensorrt_llm/common/tllmException.h (1)
35-78
: Well-structured exception class and error code enum.The implementation follows C++ best practices with proper use of
[[nodiscard]]
,noexcept
, explicit constructor, and strongly-typed enum. The error code categorization with reserved ranges is a good design choice.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
LGTM from the llmapi perspective.
/bot run --disable-fail-fast |
PR_Github #13500 [ run ] triggered by Bot |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
I would also cherry-pick this commit to fix the issue with attention dp balancing: pcastonguay@72b2d00
#include <stdexcept> | ||
#include <string> | ||
|
||
#define NEW_TLLM_EXCEPTION(...) \ | ||
tensorrt_llm::common::TllmException(__FILE__, __LINE__, tensorrt_llm::common::fmtstr(__VA_ARGS__).c_str()) | ||
|
||
#define NEW_TLLM_REQUEST_SPECIFIC_EXCEPTION_WITH_ERROR_CODE(requestID, errorCode, ...) \ |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Is this needed in this MR? If not, let's move to a separate MR about error handling. We should try to make MR well-scoped.
PR_Github #13500 [ run ] completed with state |
Summary by CodeRabbit
New Features
SchedulingParams
class for specifying attention data parallel rank and relax options.Bug Fixes
Tests
Documentation
scheduling_params
parameter in generation methods.Description
In this PR, I add the feature of scheduling attention dp requests. In this implementation, we have three choice.
The workflow is as follows: each rank selects the requests routed to it. It then attempts to retrieve requests until it reaches the
max_num_active_requests
constraint. If a request has theattention_dp_relax
flag enabled, the scheduler will try to assign it to other ranks to improve overall throughput. In contrast, requests without the relax flag will be placed back at the front of the waiting queue to await the next round of scheduling.Test Coverage
GitHub Bot Help
/bot [-h] ['run', 'kill', 'skip', 'reuse-pipeline'] ...
Provide a user friendly way for developers to interact with a Jenkins server.
Run
/bot [-h|--help]
to print this help message.See details below for each supported subcommand.
run [--reuse-test (optional)pipeline-id --disable-fail-fast --skip-test --stage-list "A10-PyTorch-1, xxx" --gpu-type "A30, H100_PCIe" --test-backend "pytorch, cpp" --add-multi-gpu-test --only-multi-gpu-test --disable-multi-gpu-test --post-merge --extra-stage "H100_PCIe-TensorRT-Post-Merge-1, xxx" --detailed-log --debug(experimental)]
Launch build/test pipelines. All previously running jobs will be killed.
--reuse-test (optional)pipeline-id
(OPTIONAL) : Allow the new pipeline to reuse build artifacts and skip successful test stages from a specified pipeline or the last pipeline if no pipeline-id is indicated. If the Git commit ID has changed, this option will be always ignored. The DEFAULT behavior of the bot is to reuse build artifacts and successful test results from the last pipeline.--disable-reuse-test
(OPTIONAL) : Explicitly prevent the pipeline from reusing build artifacts and skipping successful test stages from a previous pipeline. Ensure that all builds and tests are run regardless of previous successes.--disable-fail-fast
(OPTIONAL) : Disable fail fast on build/tests/infra failures.--skip-test
(OPTIONAL) : Skip all test stages, but still run build stages, package stages and sanity check stages. Note: Does NOT update GitHub check status.--stage-list "A10-PyTorch-1, xxx"
(OPTIONAL) : Only run the specified test stages. Examples: "A10-PyTorch-1, xxx". Note: Does NOT update GitHub check status.--gpu-type "A30, H100_PCIe"
(OPTIONAL) : Only run the test stages on the specified GPU types. Examples: "A30, H100_PCIe". Note: Does NOT update GitHub check status.--test-backend "pytorch, cpp"
(OPTIONAL) : Skip test stages which don't match the specified backends. Only support [pytorch, cpp, tensorrt, triton]. Examples: "pytorch, cpp" (does not run test stages with tensorrt or triton backend). Note: Does NOT update GitHub pipeline status.--only-multi-gpu-test
(OPTIONAL) : Only run the multi-GPU tests. Note: Does NOT update GitHub check status.--disable-multi-gpu-test
(OPTIONAL) : Disable the multi-GPU tests. Note: Does NOT update GitHub check status.--add-multi-gpu-test
(OPTIONAL) : Force run the multi-GPU tests in addition to running L0 pre-merge pipeline.--post-merge
(OPTIONAL) : Run the L0 post-merge pipeline instead of the ordinary L0 pre-merge pipeline.--extra-stage "H100_PCIe-TensorRT-Post-Merge-1, xxx"
(OPTIONAL) : Run the ordinary L0 pre-merge pipeline and specified test stages. Examples: --extra-stage "H100_PCIe-TensorRT-Post-Merge-1, xxx".--detailed-log
(OPTIONAL) : Enable flushing out all logs to the Jenkins console. This will significantly increase the log volume and may slow down the job.--debug
(OPTIONAL) : Experimental feature. Enable access to the CI container for debugging purpose. Note: Specify exactly one stage in thestage-list
parameter to access the appropriate container environment. Note: Does NOT update GitHub check status.For guidance on mapping tests to stage names, see
docs/source/reference/ci-overview.md
and the
scripts/test_to_stage_mapping.py
helper.kill
kill
Kill all running builds associated with pull request.
skip
skip --comment COMMENT
Skip testing for latest commit on pull request.
--comment "Reason for skipping build/test"
is required. IMPORTANT NOTE: This is dangerous since lack of user care and validation can cause top of tree to break.reuse-pipeline
reuse-pipeline
Reuse a previous pipeline to validate current commit. This action will also kill all currently running builds associated with the pull request. IMPORTANT NOTE: This is dangerous since lack of user care and validation can cause top of tree to break.