[TRTLLM-7412][feat] Turn off spec decode when the rolling average acceptance length drops below threshold. #7283
Conversation
📝 WalkthroughWalkthroughAdds rolling acceptance-based speculative-decoding controls: new DecodingBaseConfig fields, a SpeculationGate class (two module locations), integration into PyTorchModelEngine to track permanent-disable state, executor changes to consult Drafter with resource constraints and to update/speculation-disable via the gate, and unit-test updates for the new Drafter signature and logic. Changes
Sequence Diagram(s)sequenceDiagram
autonumber
actor Client
participant PyExec as PyExecutor
participant Engine as PyTorchModelEngine
participant Drafter as Drafter
Client->>PyExec: _prepare_and_schedule_batch(active_requests)
alt Engine.speculation_permanently_disabled == True
PyExec->>Engine: enable_spec_decode = False
else
PyExec->>Drafter: should_use_spec_decode(active_requests, max_batch_size, max_num_tokens, max_draft_len)
Drafter-->>PyExec: decision (True/False)
PyExec->>Engine: enable_spec_decode = decision
opt decision == True
PyExec->>Drafter: _prepare_draft_requests(...)
end
end
PyExec-->>Client: scheduled batch
sequenceDiagram
autonumber
participant PyExec as PyExecutor
participant Engine as PyTorchModelEngine
participant Gate as SpeculationGate
PyExec->>PyExec: _handle_responses(request_done)
alt Engine.is_spec_decode && not Engine.speculation_permanently_disabled && Gate exists
PyExec->>Gate: record_avg_decoded(avg_decoded_tokens_per_iter, request_id)
Gate-->>PyExec: (disabled_now, avg_accept)
alt disabled_now == True
PyExec->>Engine: enable_spec_decode = False
PyExec->>Engine: speculation_permanently_disabled = True
end
end
PyExec-->>PyExec: continue routing/termination
Estimated code review effort🎯 3 (Moderate) | ⏱️ ~25 minutes Possibly related PRs
Suggested reviewers
Tip 🔌 Remote MCP (Model Context Protocol) integration is now available!Pro plan users can now connect to remote MCP servers from the Integrations page. Connect with popular remote MCPs such as Notion and Linear to add more context to your reviews and chats. ✨ Finishing Touches
🧪 Generate unit tests
Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out. 🪧 TipsChatThere are 3 ways to chat with CodeRabbit:
SupportNeed help? Create a ticket on our support page for assistance with any issues or questions. CodeRabbit Commands (Invoked using PR/Issue comments)Type Other keywords and placeholders
Status, Documentation and Community
|
![coderabbitai[bot]](https://avatars.githubusercontent.com/in/347564?s=60&v=4){kind=small}
There was a problem hiding this comment.
Actionable comments posted: 0
Caution
Some comments are outside the diff and can’t be posted inline due to platform limitations.
⚠️ Outside diff range comments (5)
tensorrt_llm/llmapi/llm_args.py (1)
1-1: Add NVIDIA copyright header (2025).Repository guidelines require the NVIDIA header on all source files.
Apply this diff at the top of the file (replace with the repo-standard header if different):
+# Copyright (c) 2025, NVIDIA CORPORATION. +# +# 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.tensorrt_llm/_torch/speculative/drafter.py (1)
1-1: Add NVIDIA copyright header (2025).Apply this diff at the top of the file:
+# Copyright (c) 2025, NVIDIA CORPORATION. +# +# 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.tests/unittest/_torch/speculative/test_dynamic_spec_decode.py (1)
1-1: Add NVIDIA copyright header (2025).Apply this diff at the top of the file:
+# Copyright (c) 2025, NVIDIA CORPORATION. +# +# 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.tensorrt_llm/_torch/pyexecutor/model_engine.py (1)
1-1: Add NVIDIA copyright header (2025).Apply this diff at the top of the file:
+# Copyright (c) 2025, NVIDIA CORPORATION. +# +# 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.tensorrt_llm/_torch/pyexecutor/py_executor.py (1)
1-1: Add NVIDIA copyright header (2025).Apply this diff at the top of the file:
+# Copyright (c) 2025, NVIDIA CORPORATION. +# +# 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.
🧹 Nitpick comments (5)
tensorrt_llm/llmapi/llm_args.py (1)
356-365: Validate acceptance gating knobs (‘both-or-none’, positive values).Prevent silent misconfigurations and NPEs: require both fields together and enforce acceptance_window > 0 and acceptance_threshold > 0.
Add a validator to DecodingBaseConfig:
@model_validator(mode="after") def validate_acceptance_gating(self): if (self.acceptance_window is None) ^ (self.acceptance_threshold is None): raise ValueError("Both acceptance_window and acceptance_threshold must be set together.") if self.acceptance_window is not None: if self.acceptance_window <= 0: raise ValueError("acceptance_window must be > 0") if self.acceptance_threshold <= 0: raise ValueError("acceptance_threshold must be > 0") return selftensorrt_llm/_torch/speculative/drafter.py (1)
30-55: Solid resource-aware gating; add guard for non-positive max_concurrency.Logic looks good. Treat max_concurrency <= 0 as “always off” explicitly.
Apply this diff:
if self.max_concurrency is None: return True + if isinstance(self.max_concurrency, int) and self.max_concurrency <= 0: + return Falsetests/unittest/_torch/speculative/test_dynamic_spec_decode.py (1)
91-145: Good coverage for the new budgeting logic. Consider parametrize to reduce repetition.The scenarios are clear; you can condense with pytest.mark.parametrize for readability.
tensorrt_llm/_torch/pyexecutor/model_engine.py (1)
295-308: Instantiate gate only for meaningful values; warn on partial config.Avoid creating a gate for zero/invalid thresholds and surface partial configs.
Apply this diff:
- if self.acceptance_window and self.acceptance_threshold is not None: - self.speculation_gate = SpeculationGate(self.acceptance_window, - self.acceptance_threshold) + if (self.acceptance_window or 0) > 0 and (self.acceptance_threshold or 0) > 0: + self.speculation_gate = SpeculationGate( + self.acceptance_window, self.acceptance_threshold + ) + elif (self.acceptance_window is not None) ^ (self.acceptance_threshold is not None): + logger.warning( + "SpeculationGate requires both acceptance_window (>0) and " + "acceptance_threshold (>0); partial config will be ignored." + )tensorrt_llm/_torch/pyexecutor/py_executor.py (1)
1688-1711: Make gating update immediate; fix long lines; guard None avg.
- Set self.use_spec_decode = False when permanently disabling to take effect next iteration without recompute.
- Wrap long log lines (Ruff E501).
- Skip gate update when avg_decoded is None.
Apply this diff:
- logger.info( - f"[PyExecutor] _handle_responses: request_done={request_done}, request.py_request_id={request.py_request_id}" - ) + logger.info( + "[PyExecutor] _handle_responses: " + f"request_done={request_done}, " + f"request.py_request_id={request.py_request_id}" + ) try: if self.model_engine.is_spec_decode and not self.model_engine.speculation_permanently_disabled: - logger.info( - f"[PyExecutor] _handle_responses: self.model_engine.is_spec_decode={self.model_engine.is_spec_decode}, self.model_engine.speculation_permanently_disabled={self.model_engine.speculation_permanently_disabled}" - ) + logger.info( + "[PyExecutor] _handle_responses: " + f"is_spec_decode={self.model_engine.is_spec_decode}, " + f"permanently_disabled={self.model_engine.speculation_permanently_disabled}" + ) if self.model_engine.speculation_gate is not None: - avg_decoded = getattr( - request, 'avg_decoded_tokens_per_iter', None) - disabled_now, _ = self.model_engine.speculation_gate.record_avg_decoded( - avg_decoded, - request_id=getattr(request, 'py_request_id', - None)) - if disabled_now: + avg_decoded = getattr(request, 'avg_decoded_tokens_per_iter', None) + if avg_decoded is not None: + disabled_now, _ = self.model_engine.speculation_gate.record_avg_decoded( + avg_decoded, request_id=getattr(request, 'py_request_id', None) + ) + else: + disabled_now = False + if disabled_now: self.model_engine.speculation_permanently_disabled = True self.model_engine.enable_spec_decode = False + # Ensure executor stops drafting immediately on subsequent loop + self.use_spec_decode = False
📜 Review details
Configuration used: Path: .coderabbit.yaml
Review profile: CHILL
Plan: Pro
💡 Knowledge Base configuration:
- MCP integration is disabled by default for public repositories
- Jira integration is disabled by default for public repositories
- Linear integration is disabled by default for public repositories
You can enable these sources in your CodeRabbit configuration.
📒 Files selected for processing (5)
tensorrt_llm/_torch/pyexecutor/model_engine.py(2 hunks)tensorrt_llm/_torch/pyexecutor/py_executor.py(2 hunks)tensorrt_llm/_torch/speculative/drafter.py(1 hunks)tensorrt_llm/llmapi/llm_args.py(1 hunks)tests/unittest/_torch/speculative/test_dynamic_spec_decode.py(2 hunks)
🧰 Additional context used
📓 Path-based instructions (2)
**/*.py
📄 CodeRabbit inference engine (CODING_GUIDELINES.md)
**/*.py: Code must target Python 3.8+
Indent Python code with 4 spaces; do not use tabs
Preserve module namespaces when importing; import modules/packages and access members via the module (e.g., from package.subpackage import foo; foo.SomeClass())
Python file names should be snake_case
Python class names should be PascalCase
Python functions/methods and local variables should be snake_case; variables beginning with a number should be prefixed with k_ (e.g., k_99th_percentile)
Global variables should be UPPER_SNAKE_CASE prefixed with G_ (e.g., G_MY_GLOBAL); constants should be UPPER_SNAKE_CASE
Avoid shadowing variables from outer scopes; initialize all externally visible members in init
Prefer docstrings for interfaces used outside a file; comments should be reserved for in-function or file-local interfaces
Use Google-style docstrings for classes and functions; attributes and variables may be documented inline with trailing string literals
Avoid reflection when simpler, explicit code suffices (e.g., avoid dict(**locals()) patterns)
In try/except, catch the narrowest exceptions possible
For duck-typing patterns, keep the try body minimal and move logic to else to avoid masking unrelated failures
Files:
tensorrt_llm/_torch/pyexecutor/model_engine.pytensorrt_llm/_torch/speculative/drafter.pytensorrt_llm/llmapi/llm_args.pytensorrt_llm/_torch/pyexecutor/py_executor.pytests/unittest/_torch/speculative/test_dynamic_spec_decode.py
**/*.{c,cc,cpp,cxx,h,hh,hpp,hxx,cu,cuh,py}
📄 CodeRabbit inference engine (CODING_GUIDELINES.md)
Prepend the NVIDIA copyright header (current year) to all source files (.cpp, .h, .cu, .py, etc.)
Files:
tensorrt_llm/_torch/pyexecutor/model_engine.pytensorrt_llm/_torch/speculative/drafter.pytensorrt_llm/llmapi/llm_args.pytensorrt_llm/_torch/pyexecutor/py_executor.pytests/unittest/_torch/speculative/test_dynamic_spec_decode.py
🧠 Learnings (1)
📚 Learning: 2025-08-19T12:45:11.997Z
Learnt from: amitz-nv
PR: NVIDIA/TensorRT-LLM#7033
File: tensorrt_llm/_torch/pyexecutor/model_engine.py:0-0
Timestamp: 2025-08-19T12:45:11.997Z
Learning: In tensorrt_llm/_torch/pyexecutor/model_engine.py, DoRA (Delta Orthogonal Rank Adaptation) functionality was removed from the PyTorch flow to eliminate issues with inverted DoRA detection logic. The original is_dora condition was checking if scaling_vec_pointer == 0, which was potentially incorrect.
Applied to files:
tensorrt_llm/_torch/pyexecutor/py_executor.py
🧬 Code graph analysis (3)
tensorrt_llm/_torch/speculative/drafter.py (1)
tensorrt_llm/_torch/pyexecutor/llm_request.py (1)
LlmRequest(282-424)
tensorrt_llm/_torch/pyexecutor/py_executor.py (3)
tensorrt_llm/_torch/speculative/drafter.py (1)
should_use_spec_decode(30-54)tensorrt_llm/_torch/pyexecutor/cuda_graph_runner.py (1)
enable_spec_decode(52-53)tensorrt_llm/logger.py (1)
warning(131-132)
tests/unittest/_torch/speculative/test_dynamic_spec_decode.py (1)
tensorrt_llm/_torch/speculative/drafter.py (3)
Drafter(9-54)prepare_draft_tokens(16-27)should_use_spec_decode(30-54)
🪛 Ruff (0.12.2)
tensorrt_llm/_torch/pyexecutor/py_executor.py
1690-1690: Line too long (129 > 120)
(E501)
1695-1695: Line too long (235 > 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)
tests/unittest/_torch/speculative/test_dynamic_spec_decode.py (1)
54-65: Mock shape aligns with new signature. LGTM.tensorrt_llm/_torch/pyexecutor/model_engine.py (1)
65-65: Import of SpeculationGate looks right.tensorrt_llm/_torch/pyexecutor/py_executor.py (1)
887-897: Respect permanent-disable first; resource-aware Drafter call is correct.
There was a problem hiding this comment.
Actionable comments posted: 5
🧹 Nitpick comments (7)
tensorrt_llm/_torch/pyexecutor/speculation_gate.py (7)
4-4: Import style: prefer module namespace for logger (repo guideline)To preserve module namespaces, consider importing the module and accessing the member via the module (e.g.,
import tensorrt_llm.logger as trt_logger; trt_logger.logger.info(...)). If this diverges from existing local convention, feel free to ignore.
8-11: Docstring formatting and content (Google style)Docstring uses tabs and lacks Args/Behavior details. Propose a concise Google-style docstring.
Apply this diff:
class SpeculationGate: - """ - Tracks rolling average of accepted draft tokens per iteration over the last N completed requests. - Permanently disables speculation when average falls below a threshold. - """ + """ + Tracks a rolling average of accepted draft tokens per iteration across the last N completed requests. + + When the rolling average falls below a threshold, speculative decoding is permanently disabled + until `reset()` is called. + """
30-38: Method docstring: clarify contract and return valuesMake the API crystal-clear for call sites and tests.
Apply this diff:
- """ - Record a completed request's avg_decoded_tokens_per_iter. - Returns (disabled_now, current_avg_accept) where disabled_now is True only when the call causes disable. - """ + """ + Record the per-request average decoded tokens per iteration. + + Args: + avg_decoded_tokens_per_iter: Average tokens decoded per iteration for the request. + Interpreted as accepted_len = max(0, value - 1). None or invalid values are treated as 0. + request_id: Optional request identifier for logging. + + Returns: + (disabled_now, current_avg_accept) + disabled_now: True only on the call that causes permanent disable. + current_avg_accept: The rolling average once at least `window` samples have been observed; + otherwise None. + """
44-47: Redundant None/<=0 checks for window/thresholdConstructor enforces valid values; these early returns are dead code. Remove for clarity, or make the ctor accept Optional and keep the checks.
Apply this diff to remove them:
- if self.window is None or self.threshold is None: - return False, None - if self.window <= 0: - return False, None
30-91: Potential data race if called from multiple threadsIf
record_avg_decodedcan be invoked concurrently (e.g., multiple response-handling threads), updates to deque/sum/counters need a lock.Apply this diff if multi-threaded:
+import threading @@ def __init__(self, window: int, threshold: float): @@ self.disabled = False + self._lock = threading.Lock() @@ - def record_avg_decoded( + def record_avg_decoded( self, avg_decoded_tokens_per_iter: Optional[float], *, request_id: Optional[int] = None) -> Tuple[bool, Optional[float]]: @@ - logger.debug("[SpeculationGate] record_avg_decoded avg=%s req_id=%s", - avg_decoded_tokens_per_iter, request_id) - if self.disabled: - return False, None + logger.debug("[SpeculationGate] record_avg_decoded avg=%s req_id=%s", + avg_decoded_tokens_per_iter, request_id) + with self._lock: + if self.disabled: + return False, None @@ - self.acceptance_history.append(accepted_len) - self.acceptance_sum += accepted_len + with self._lock: + self.acceptance_history.append(accepted_len) + self.acceptance_sum += accepted_len if len(self.acceptance_history) > self.window: removed = self.acceptance_history.popleft() self.acceptance_sum -= removed @@ - self.num_completed_for_acceptance += 1 + with self._lock: + self.num_completed_for_acceptance += 1 @@ - if self.num_completed_for_acceptance >= self.window: - avg_accept = self.acceptance_sum / len(self.acceptance_history) + with self._lock: + if self.num_completed_for_acceptance >= self.window: + avg_accept = self.acceptance_sum / len(self.acceptance_history)
24-29: Optional: log reset() for traceabilityLightweight DEBUG log helps correlate state transitions.
Apply this diff:
def reset(self) -> None: self.acceptance_history.clear() self.acceptance_sum = 0.0 self.num_completed_for_acceptance = 0 self.disabled = False + logger.debug("[SpeculationGate] reset() called; state cleared and re-enabled")
1-92: Unit tests to add (happy path + edge cases)Recommend adding tests to pin behavior: warmup gate, disable trigger, skip/invalid inputs, reset(), and idempotence after disable.
- window=3, threshold=0.5: feed [1.0, 1.2, 1.1] → stays enabled with avg≈0.1 < 0.5? should disable; verify disabled_now on 3rd call.
- window=2, threshold=0.0: any inputs → never disable.
- Include None/NaN/negative → treated as 0; ensure no NaN propagation.
- After disable, subsequent calls return (False, None) and don’t mutate history/sum.
- reset() re-enables and clears stats.
I can draft these tests if helpful.
📜 Review details
Configuration used: Path: .coderabbit.yaml
Review profile: CHILL
Plan: Pro
💡 Knowledge Base configuration:
- MCP integration is disabled by default for public repositories
- Jira integration is disabled by default for public repositories
- Linear integration is disabled by default for public repositories
You can enable these sources in your CodeRabbit configuration.
📒 Files selected for processing (1)
tensorrt_llm/_torch/pyexecutor/speculation_gate.py(1 hunks)
🧰 Additional context used
📓 Path-based instructions (2)
**/*.py
📄 CodeRabbit inference engine (CODING_GUIDELINES.md)
**/*.py: Code must target Python 3.8+
Indent Python code with 4 spaces; do not use tabs
Preserve module namespaces when importing; import modules/packages and access members via the module (e.g., from package.subpackage import foo; foo.SomeClass())
Python file names should be snake_case
Python class names should be PascalCase
Python functions/methods and local variables should be snake_case; variables beginning with a number should be prefixed with k_ (e.g., k_99th_percentile)
Global variables should be UPPER_SNAKE_CASE prefixed with G_ (e.g., G_MY_GLOBAL); constants should be UPPER_SNAKE_CASE
Avoid shadowing variables from outer scopes; initialize all externally visible members in init
Prefer docstrings for interfaces used outside a file; comments should be reserved for in-function or file-local interfaces
Use Google-style docstrings for classes and functions; attributes and variables may be documented inline with trailing string literals
Avoid reflection when simpler, explicit code suffices (e.g., avoid dict(**locals()) patterns)
In try/except, catch the narrowest exceptions possible
For duck-typing patterns, keep the try body minimal and move logic to else to avoid masking unrelated failures
Files:
tensorrt_llm/_torch/pyexecutor/speculation_gate.py
**/*.{c,cc,cpp,cxx,h,hh,hpp,hxx,cu,cuh,py}
📄 CodeRabbit inference engine (CODING_GUIDELINES.md)
Prepend the NVIDIA copyright header (current year) to all source files (.cpp, .h, .cu, .py, etc.)
Files:
tensorrt_llm/_torch/pyexecutor/speculation_gate.py
🪛 Ruff (0.12.2)
tensorrt_llm/_torch/pyexecutor/speculation_gate.py
40-40: Line too long (146 > 120)
(E501)
56-56: Line too long (198 > 120)
(E501)
66-66: Line too long (128 > 120)
(E501)
71-71: Line too long (203 > 120)
(E501)
77-77: Line too long (143 > 120)
(E501)
82-82: Line too long (173 > 120)
(E501)
87-87: Line too long (148 > 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 (2)
tensorrt_llm/_torch/pyexecutor/speculation_gate.py (2)
69-75: Confirm counting semantics for None/invalid samplesCurrently, requests with None/invalid avg contribute 0 and count toward the N-sample warmup. Verify this matches the product requirement; alternatively, skip such samples to avoid premature disables.
If you decide to skip them, apply:
- self.num_completed_for_acceptance += 1 + # Count only valid samples toward warmup. + if avg_decoded_tokens_per_iter is not None and math.isfinite(float(avg_decoded_tokens_per_iter)) \ + and float(avg_decoded_tokens_per_iter) >= 0.0: + self.num_completed_for_acceptance += 1 + else: + logger.debug("[SpeculationGate] skipping invalid sample for warmup") - if self.num_completed_for_acceptance >= self.window: + if self.num_completed_for_acceptance >= self.window and len(self.acceptance_history) > 0:
74-91: LGTM on core gating logicWindowed average, permanent disable, and return contract look correct and align with the PR intent.
There was a problem hiding this comment.
Actionable comments posted: 4
🧹 Nitpick comments (8)
tensorrt_llm/_torch/speculative/speculation_gate.py (4)
39-47: Return early branches fine; reduce log verbosity to debugThese paths run per request and log at INFO. Switch to DEBUG to avoid log flood.
- logger.info( + logger.debug( f"[SpeculationGate] record_avg_decoded called with avg_decoded_tokens_per_iter={avg_decoded_tokens_per_iter}, request_id={request_id}" )
49-58: Acceptance metric assumption may not hold across algorithmsaccepted_len = max(0, avg_decoded-1) assumes 1 target token + rest accepted. Verify this is consistent for all speculative modes you gate; otherwise, take the actual “accepted” metric from request stats and record that.
I can wire this to a more direct “accepted per iter” field if available.
60-72: Long log lines >120 cols (ruff E501)Break lines or lower to DEBUG to comply with style.
- logger.info( - f"[SpeculationGate] Rolling window: removed old value {removed:.3f}, window size={len(self.acceptance_history)}" - ) + logger.debug( + "[SpeculationGate] Rolling window: removed old value %.3f, window size=%d", + removed, len(self.acceptance_history) + )
74-91: Simplify condition; rely on history length instead of a separate counternum_completed_for_acceptance is redundant. Using len(self.acceptance_history) improves clarity and avoids divergence if code changes pop logic later.
- self.num_completed_for_acceptance += 1 - logger.info( - f"[SpeculationGate] Rolling stats: completed={self.num_completed_for_acceptance}/{self.window}, current_sum={self.acceptance_sum:.3f}, history={[f'{x:.3f}' for x in self.acceptance_history]}" - ) - - if self.num_completed_for_acceptance >= self.window: + logger.debug( + "[SpeculationGate] Rolling stats: completed=%d/%d, current_sum=%.3f", + len(self.acceptance_history), self.window, self.acceptance_sum + ) + if len(self.acceptance_history) >= self.window:tests/unittest/_torch/speculative/test_spec_gate.py (3)
15-21: Guard on model availability; skip gracefullyIf models are missing at llm_models_root(), generate a clear skip instead of failing later in model load.
models_path = llm_models_root() + if not os.path.isdir(models_path): + pytest.skip(f"Models path not found: {models_path}")
45-52: This test doesn’t exercise gating disablementYou set acceptance_window/threshold but run only 2 prompts (window=3), so the gate never triggers. Add a lightweight unit test for SpeculationGate logic that doesn’t need GPUs.
I can add a new fast test (no CUDA) that feeds synthetic averages to trigger disable and asserts state transitions.
Additional file (new):# tests/unittest/_torch/speculative/test_speculation_gate_unit.py import pytest from tensorrt_llm._torch.speculative.speculation_gate import SpeculationGate def test_gate_triggers_disable(): g = SpeculationGate(window=3, threshold=0.6) # accepted lens: 0.2, 0.4, 0.5 -> avg 0.366 < 0.6 => disable outs = [g.record_avg_decoded(a) for a in (1.2, 1.4, 1.5)] assert outs[-1][0] is True assert g.disabled is True
71-75: Remove prints in testsUse assertion messages instead of print noise.
- print(f"text_spec: {text_spec}") - print(f"text_ref: {text_ref}") - # The spec decode algorithm currently guarantees identical results - assert text_spec == text_ref + assert text_spec == text_ref, f"Mismatch:\n spec={text_spec}\n ref={text_ref}"tensorrt_llm/_torch/pyexecutor/py_executor.py (1)
1689-1713: Shorten long INFO logs (ruff E501)Break long f-strings or switch to structured logging.
- logger.info( - f"[PyExecutor] _handle_responses: request_done={request_done}, request.py_request_id={request.py_request_id}" - ) + logger.info("[PyExecutor] request_done=%s, req_id=%s", + request_done, request.py_request_id)
📜 Review details
Configuration used: Path: .coderabbit.yaml
Review profile: CHILL
Plan: Pro
💡 Knowledge Base configuration:
- MCP integration is disabled by default for public repositories
- Jira integration is disabled by default for public repositories
- Linear integration is disabled by default for public repositories
You can enable these sources in your CodeRabbit configuration.
📒 Files selected for processing (5)
tensorrt_llm/_torch/pyexecutor/model_engine.py(2 hunks)tensorrt_llm/_torch/pyexecutor/py_executor.py(3 hunks)tensorrt_llm/_torch/speculative/speculation_gate.py(1 hunks)tensorrt_llm/llmapi/llm_args.py(1 hunks)tests/unittest/_torch/speculative/test_spec_gate.py(1 hunks)
🚧 Files skipped from review as they are similar to previous changes (1)
- tensorrt_llm/_torch/pyexecutor/model_engine.py
🧰 Additional context used
📓 Path-based instructions (3)
**/*.{cpp,cc,cxx,cu,py,h,hpp,hh,hxx,cuh}
📄 CodeRabbit inference engine (CODING_GUIDELINES.md)
Use spaces only; indent 4 spaces
Files:
tests/unittest/_torch/speculative/test_spec_gate.pytensorrt_llm/llmapi/llm_args.pytensorrt_llm/_torch/pyexecutor/py_executor.pytensorrt_llm/_torch/speculative/speculation_gate.py
**/*.py
📄 CodeRabbit inference engine (CODING_GUIDELINES.md)
**/*.py: Python code must target Python 3.8+
Indent with 4 spaces; do not use tabs
Preserve module namespaces in imports; prefer from package.subpackage import module then module.Symbol
Python file names use snake_case (e.g., some_file.py)
Class names use PascalCase
Function and method names use snake_case
Local variable names use snake_case; if starting with a number, prefix k (e.g., k_99th_percentile)
Global variables use G_ prefix and UPPER_SNAKE_CASE (e.g., G_MY_GLOBAL)
Constants use UPPER_SNAKE_CASE
Avoid shadowing variables from an outer scope
Initialize all externally visible members of a class in init
For interfaces used outside a file, prefer docstrings; reserve comments for internal code or local interfaces
Use Google-style docstrings for classes and functions; document attributes/variables inline as shown
Avoid reflection when simple, explicit code suffices (e.g., prefer def make_complex(x,y) over locals()/dict tricks)
Catch the narrowest exceptions possible in try/except
For duck-typing try/except, keep try body minimal and use else for main logic
Files:
tests/unittest/_torch/speculative/test_spec_gate.pytensorrt_llm/llmapi/llm_args.pytensorrt_llm/_torch/pyexecutor/py_executor.pytensorrt_llm/_torch/speculative/speculation_gate.py
**/*.{cpp,cc,cxx,cu,h,hpp,hh,hxx,cuh,py}
📄 CodeRabbit inference engine (CODING_GUIDELINES.md)
Prepend NVIDIA copyright header (current year) to all source files (.cpp, .h, .cu, .py, etc.)
Files:
tests/unittest/_torch/speculative/test_spec_gate.pytensorrt_llm/llmapi/llm_args.pytensorrt_llm/_torch/pyexecutor/py_executor.pytensorrt_llm/_torch/speculative/speculation_gate.py
🧠 Learnings (1)
📚 Learning: 2025-07-28T17:06:08.621Z
Learnt from: moraxu
PR: NVIDIA/TensorRT-LLM#6303
File: tests/integration/test_lists/qa/examples_test_list.txt:494-494
Timestamp: 2025-07-28T17:06:08.621Z
Learning: In TensorRT-LLM testing, it's common to have both CLI flow tests (test_cli_flow.py) and PyTorch API tests (test_llm_api_pytorch.py) for the same model. These serve different purposes: CLI flow tests validate the traditional command-line workflow, while PyTorch API tests validate the newer LLM API backend. Both are legitimate and should coexist.
Applied to files:
tests/unittest/_torch/speculative/test_spec_gate.py
🧬 Code graph analysis (2)
tests/unittest/_torch/speculative/test_spec_gate.py (2)
tensorrt_llm/llmapi/llm_args.py (4)
CudaGraphConfig(106-163)EagleDecodingConfig(455-493)KvCacheConfig(980-1075)speculative_model_dir(1398-1399)tensorrt_llm/_torch/pyexecutor/py_executor.py (1)
shutdown(353-366)
tensorrt_llm/_torch/pyexecutor/py_executor.py (3)
tensorrt_llm/_torch/speculative/drafter.py (1)
should_use_spec_decode(30-54)tensorrt_llm/_torch/speculative/speculation_gate.py (1)
record_avg_decoded(30-91)tensorrt_llm/logger.py (1)
warning(131-132)
🪛 Ruff (0.12.2)
tensorrt_llm/_torch/pyexecutor/py_executor.py
1691-1691: Line too long (129 > 120)
(E501)
1696-1696: Line too long (235 > 120)
(E501)
tensorrt_llm/_torch/speculative/speculation_gate.py
16-16: Undefined name Deque
(F821)
40-40: Line too long (146 > 120)
(E501)
56-56: Line too long (198 > 120)
(E501)
66-66: Line too long (128 > 120)
(E501)
71-71: Line too long (203 > 120)
(E501)
77-77: Line too long (143 > 120)
(E501)
82-82: Line too long (173 > 120)
(E501)
87-87: Line too long (148 > 120)
(E501)
🔇 Additional comments (3)
tensorrt_llm/llmapi/llm_args.py (1)
369-391: Validators OK; align semantics with docs
- Window=0 disables feature per comments; this is good. Consider clamping very small thresholds (e.g., negative rejected already) or documenting that 0 means “never disable.”
If you want a warning on the “no-op” case (window==0 with threshold set), I can add one. -->
tests/unittest/_torch/speculative/test_spec_gate.py (1)
1-14: Missing NVIDIA copyright headerAdd the standard header per repo guidelines.
+# Copyright (c) 2025, NVIDIA CORPORATION. All rights reserved.⛔ Skipped due to learnings
Learnt from: CR PR: NVIDIA/TensorRT-LLM#0 File: CODING_GUIDELINES.md:0-0 Timestamp: 2025-08-29T06:18:00.220Z Learning: Applies to **/*.{cpp,cc,cxx,cu,h,hpp,hh,hxx,cuh,py} : Prepend NVIDIA copyright header (current year) to all source files (.cpp, .h, .cu, .py, etc.)Learnt from: moraxu PR: NVIDIA/TensorRT-LLM#6303 File: tests/integration/test_lists/qa/examples_test_list.txt:494-494 Timestamp: 2025-07-28T17:06:08.621Z Learning: In TensorRT-LLM testing, it's common to have both CLI flow tests (test_cli_flow.py) and PyTorch API tests (test_llm_api_pytorch.py) for the same model. These serve different purposes: CLI flow tests validate the traditional command-line workflow, while PyTorch API tests validate the newer LLM API backend. Both are legitimate and should coexist.Learnt from: galagam PR: NVIDIA/TensorRT-LLM#6487 File: tests/unittest/_torch/auto_deploy/unit/singlegpu/test_ad_trtllm_bench.py:1-12 Timestamp: 2025-08-06T13:58:07.506Z Learning: In TensorRT-LLM, test files (files under tests/ directories) do not require NVIDIA copyright headers, unlike production source code files. Test files typically start directly with imports, docstrings, or code.tensorrt_llm/_torch/pyexecutor/py_executor.py (1)
1029-1031: Retain empty list for req.py_draft_tokens
Downstream code uniformly treats py_draft_tokens as a list (falsy checks, extends, appends), so None isn’t expected and would break list operations.Likely an incorrect or invalid review comment.
zheyuf
changed the title
zheyuf
changed the title
zheyuf
requested review from
syuoni,
nv-yilinf and
mikeiovine
and removed request for
syuoni and
nv-yilinf
zheyuf
force-pushed
the
roll_avg
branch
3 times, most recently
from
73f2065 to
a760ddc
Compare
|
/bot run --disable-fail-fast |
|
PR_Github #19715 [ run ] triggered by Bot |
|
PR_Github #19715 [ run ] completed with state |
|
/bot run --disable-fail-fast |
|
PR_Github #19749 [ run ] triggered by Bot |
|
PR_Github #19749 [ run ] completed with state |
|
/bot run --disable-fail-fast |
|
PR_Github #19823 [ run ] triggered by Bot |
|
PR_Github #19823 [ run ] completed with state |
|
/bot run --disable-fail-fast |
|
PR_Github #19847 [ run ] triggered by Bot |
|
PR_Github #19847 [ run ] completed with state |
|
/bot run |
|
PR_Github #19884 [ run ] triggered by Bot |
|
PR_Github #19884 [ run ] completed with state |
|
/bot run --disable-fail-fast |
|
PR_Github #19988 [ run ] triggered by Bot |
|
PR_Github #19988 [ run ] completed with state |
… test in test_dynamic_spec_decode(patch is not called at all). Signed-off-by: Zheyu Fu <[email protected]>
… threshold. Signed-off-by: Zheyu Fu <[email protected]>
Signed-off-by: Zheyu Fu <[email protected]>
…r. Also clean. Signed-off-by: Zheyu Fu <[email protected]>
Signed-off-by: Zheyu Fu <[email protected]>
Signed-off-by: Zheyu Fu <[email protected]>
Signed-off-by: Zheyu Fu <[email protected]>
|
/bot run --disable-fail-fast |
|
PR_Github #20124 [ run ] triggered by Bot |
|
PR_Github #20124 [ run ] completed with state |
Summary by CodeRabbit
Description
This pull request (PR) depends on PR#7511 and should be merged after it.
Feature requested from Microsoft. Keep a rolling average of the acceptance length over the last N requests (specified via a DecodingConfig). Turn off speculative decoding permanently when the rolling average drops below some user-specified threshold. This should only kick in after at least N requests have completed since it's going to fluctuate a lot at the beginning.
Test Coverage
Added unit tests in test_spec_gate.py, which contains an end-to-end test and several functional test only for the class SpeculationGate.
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-listparameter 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.mdand the
scripts/test_to_stage_mapping.pyhelper.kill
killKill all running builds associated with pull request.
skip
skip --comment COMMENTSkip 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-pipelineReuse 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.