fix: prepare DeepSeek-V4 draft sliding tables - #17178
Open
zcxGGmu wants to merge 1 commit into
Open
Conversation
Signed-off-by: zq <zhouquan1511@163.com>
Contributor
WalkthroughDeepSeek-V4 metadata preparation now computes sliding-window block tables for the target and optional draft KV-cache managers before base TRT-LLM preparation. A unit test verifies the calls, arguments, and ordering. ChangesDeepSeek-V4 metadata preparation
Estimated code review effort: 2 (Simple) | ~10 minutes Suggested reviewers: 🚥 Pre-merge checks | ✅ 5✅ Passed checks (5 passed)
✨ Finishing Touches🧪 Generate unit tests (beta)
Comment |
![coderabbitai[bot]](https://avatars.githubusercontent.com/in/347564?s=60&v=4){kind=small}
Contributor
There was a problem hiding this comment.
Actionable comments posted: 2
🤖 Prompt for all review comments with AI agents
Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.
Inline comments:
In
`@tests/unittest/_torch/attention/sparse/deepseek_v4/test_deepseek_v4_metadata.py`:
- Around line 26-55: Add
tests/unittest/_torch/attention/sparse/deepseek_v4/test_deepseek_v4_metadata.py
to the appropriate CI and QA entries in
tests/integration/test_lists/test-db/l0_b200.yml so
test_prepare_computes_draft_sliding_block_tables_before_base_prepare is included
in automated coverage.
- Around line 26-43: Update
test_prepare_computes_draft_sliding_block_tables_before_base_prepare with
explicit pytest.MonkeyPatch and TrtllmAttentionMetadata parameter annotations
and a None return annotation. Rename the local stop_at_base_prepare helper to
_stop_at_base_prepare, annotate its self parameter with TrtllmAttentionMetadata
and its return type as None, and use the renamed helper in the monkeypatch.
🪄 Autofix (Beta)
Fix all unresolved CodeRabbit comments on this PR:
- Push a commit to this branch (recommended)
- Create a new PR with the fixes
ℹ️ Review info
⚙️ Run configuration
Configuration used: Path: .coderabbit.yaml
Review profile: CHILL
Plan: Enterprise
Run ID: 552abcac-19c8-4abe-b2c5-2074371be7db
📒 Files selected for processing (2)
tensorrt_llm/_torch/attention_backend/sparse/deepseek_v4/deepseek_v4.pytests/unittest/_torch/attention/sparse/deepseek_v4/test_deepseek_v4_metadata.py
Comment on lines
+26
to
+43
| def test_prepare_computes_draft_sliding_block_tables_before_base_prepare(monkeypatch): | ||
| """DeepSeek-V4 MTP draft KV managers need their sliding tables prepared. | ||
|
|
||
| The base TRT-LLM metadata prepare path copies block offsets from both the | ||
| target and draft managers. DeepSeek-V4's copy path consumes precomputed | ||
| sliding-window tables, so the draft manager must compute them before the | ||
| base prepare reaches copy_batch_block_offsets(). | ||
| """ | ||
| metadata = object.__new__(DeepseekV4TrtllmAttentionMetadata) | ||
| metadata.kv_cache_manager = MagicMock() | ||
| metadata.draft_kv_cache_manager = MagicMock() | ||
| metadata.request_ids = [11, 12, 13] | ||
| metadata.num_contexts = 2 | ||
|
|
||
| def stop_at_base_prepare(self): | ||
| raise RuntimeError("base prepare reached") | ||
|
|
||
| monkeypatch.setattr(TrtllmAttentionMetadata, "prepare", stop_at_base_prepare) |
Contributor
There was a problem hiding this comment.
📐 Maintainability & Code Quality | 🟠 Major | ⚡ Quick win
🧩 Analysis chain
🏁 Script executed:
#!/bin/bash
set -e
printf '%s\n' '--- target file ---'
sed -n '1,180p' tests/unittest/_torch/attention/sparse/deepseek_v4/test_deepseek_v4_metadata.py
printf '%s\n' '--- relevant guidelines ---'
if [ -f CODING_GUIDELINES.md ]; then
rg -n -A4 -B4 'Annotate every function|Prefix non-public|pytest|test' CODING_GUIDELINES.md | head -160
fi
printf '%s\n' '--- nearby test-list references ---'
rg -n 'deepseek_v4|test_deepseek_v4_metadata' tests/integration/test_lists tests 2>/dev/null | head -160Repository: NVIDIA/TensorRT-LLM
Length of output: 27204
🏁 Script executed:
#!/bin/bash
set -e
sed -n '1,180p' tests/unittest/_torch/attention/sparse/deepseek_v4/test_deepseek_v4_metadata.py
printf '%s\n' '--- guidelines ---'
if [ -f CODING_GUIDELINES.md ]; then
rg -n -A4 -B4 'Annotate every function|Prefix non-public|pytest|test' CODING_GUIDELINES.md | head -160
fi
printf '%s\n' '--- test-list references ---'
rg -n 'deepseek_v4|test_deepseek_v4_metadata' tests/integration/test_lists tests 2>/dev/null | head -160Repository: NVIDIA/TensorRT-LLM
Length of output: 26941
🏁 Script executed:
#!/bin/bash
set -e
python3 - <<'PY'
import ast
from pathlib import Path
path = Path("tests/unittest/_torch/attention/sparse/deepseek_v4/test_deepseek_v4_metadata.py")
tree = ast.parse(path.read_text())
for node in ast.walk(tree):
if isinstance(node, (ast.FunctionDef, ast.AsyncFunctionDef)):
args = [*node.args.posonlyargs, *node.args.args, *node.args.kwonlyargs]
missing_args = [arg.arg for arg in args if arg.annotation is None]
print(
f"{node.name}: return_annotation={node.returns is not None}, "
f"missing_argument_annotations={missing_args}"
)
if node.name == "test_prepare_computes_draft_sliding_block_tables_before_base_prepare":
for child in node.body:
if isinstance(child, ast.FunctionDef):
print(
f"nested_helper={child.name}: "
f"self_annotated={bool(child.args.args and child.args.args[0].annotation)}, "
f"return_annotation={child.returns is not None}"
)
PY
printf '%s\n' '--- exact guideline matches ---'
rg -n -i 'annotat(e|ion).*every function|non-public|private names|leading underscore' CODING_GUIDELINES.md
printf '%s\n' '--- tracked-file status and focused diff ---'
git status --short -- tests/unittest/_torch/attention/sparse/deepseek_v4/test_deepseek_v4_metadata.py tests/integration/test_lists
git diff --stat -- tests/unittest/_torch/attention/sparse/deepseek_v4/test_deepseek_v4_metadata.py tests/integration/test_lists
git diff --unified=3 -- tests/unittest/_torch/attention/sparse/deepseek_v4/test_deepseek_v4_metadata.py
printf '%s\n' '--- exact test-list membership ---'
rg -n -F 'unittest/_torch/attention/sparse/deepseek_v4/test_deepseek_v4_metadata.py' tests/integration/test_lists || trueRepository: NVIDIA/TensorRT-LLM
Length of output: 496
🏁 Script executed:
#!/bin/bash
set -e
python3 - <<'PY'
import ast
from pathlib import Path
path = Path("tests/unittest/_torch/attention/sparse/deepseek_v4/test_deepseek_v4_metadata.py")
tree = ast.parse(path.read_text())
for node in ast.walk(tree):
if isinstance(node, (ast.FunctionDef, ast.AsyncFunctionDef)):
args = [*node.args.posonlyargs, *node.args.args, *node.args.kwonlyargs]
missing_args = [arg.arg for arg in args if arg.annotation is None]
print(
f"{node.name}: return_annotation={node.returns is not None}, "
f"missing_argument_annotations={missing_args}"
)
if node.name == "test_prepare_computes_draft_sliding_block_tables_before_base_prepare":
for child in node.body:
if isinstance(child, ast.FunctionDef):
print(
f"nested_helper={child.name}: "
f"self_annotated={bool(child.args.args and child.args.args[0].annotation)}, "
f"return_annotation={child.returns is not None}"
)
PY
printf '%s\n' '--- exact guideline matches ---'
rg -n -i 'annotat(e|ion).*every function|non-public|private names|leading underscore' CODING_GUIDELINES.md
printf '%s\n' '--- focused status and diff ---'
git status --short -- tests/unittest/_torch/attention/sparse/deepseek_v4/test_deepseek_v4_metadata.py tests/integration/test_lists
git diff --stat -- tests/unittest/_torch/attention/sparse/deepseek_v4/test_deepseek_v4_metadata.py tests/integration/test_lists
git diff --unified=3 -- tests/unittest/_torch/attention/sparse/deepseek_v4/test_deepseek_v4_metadata.py
printf '%s\n' '--- test-list membership ---'
rg -n -F 'unittest/_torch/attention/sparse/deepseek_v4/test_deepseek_v4_metadata.py' tests/integration/test_lists || trueRepository: NVIDIA/TensorRT-LLM
Length of output: 496
🏁 Script executed:
#!/bin/bash
set -e
printf '%s\n' '--- focused status and diff ---'
git status --short -- tests/unittest/_torch/attention/sparse/deepseek_v4/test_deepseek_v4_metadata.py tests/integration/test_lists
git diff --stat -- tests/unittest/_torch/attention/sparse/deepseek_v4/test_deepseek_v4_metadata.py tests/integration/test_lists
git diff --unified=5 -- tests/unittest/_torch/attention/sparse/deepseek_v4/test_deepseek_v4_metadata.py
printf '%s\n' '--- exact test-list membership ---'
rg -n -F 'unittest/_torch/attention/sparse/deepseek_v4/test_deepseek_v4_metadata.py' tests/integration/test_lists || true
printf '%s\n' '--- nearby CI list entries ---'
sed -n '68,90p' tests/integration/test_lists/test-db/l0_b200.ymlRepository: NVIDIA/TensorRT-LLM
Length of output: 2045
🏁 Script executed:
#!/bin/bash
set -e
python3 - <<'PY'
import ast
from pathlib import Path
source = Path(
"tests/unittest/_torch/attention/sparse/deepseek_v4/test_deepseek_v4_metadata.py"
).read_text()
tree = ast.parse(source)
test = next(
node for node in tree.body
if isinstance(node, ast.FunctionDef)
and node.name == "test_prepare_computes_draft_sliding_block_tables_before_base_prepare"
)
helper = next(node for node in test.body if isinstance(node, ast.FunctionDef))
print("test_function_present:", test.name)
print("helper_present:", helper.name)
print("test_annotation_state:", test.returns is not None, [a.annotation is not None for a in test.args.args])
print("helper_annotation_state:", helper.returns is not None, [a.annotation is not None for a in helper.args.args])
list_path = Path("tests/integration/test_lists/test-db/l0_b200.yml")
list_text = list_path.read_text()
print("parent_directory_covers_file:",
" - unittest/_torch/attention\n" in list_text)
print("exact_file_entry:",
"unittest/_torch/attention/sparse/deepseek_v4/test_deepseek_v4_metadata.py"
in list_text)
PY
python3 - <<'PY'
import pytest
print("pytest.MonkeyPatch_available:", hasattr(pytest, "MonkeyPatch"))
print("pytest.MonkeyPatch:", pytest.MonkeyPatch)
PYRepository: NVIDIA/TensorRT-LLM
Length of output: 540
Annotate the new test functions and mark the local helper as non-public.
Annotate test_prepare_computes_draft_sliding_block_tables_before_base_prepare() and _stop_at_base_prepare() with pytest.MonkeyPatch, TrtllmAttentionMetadata, and None as appropriate. Rename stop_at_base_prepare to _stop_at_base_prepare.
Test coverage: test_prepare_computes_draft_sliding_block_tables_before_base_prepare() was added and is covered by unittest/_torch/attention in tests/integration/test_lists/test-db/l0_b200.yml. Verdict: sufficient.
🤖 Prompt for AI Agents
Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.
In
`@tests/unittest/_torch/attention/sparse/deepseek_v4/test_deepseek_v4_metadata.py`
around lines 26 - 43, Update
test_prepare_computes_draft_sliding_block_tables_before_base_prepare with
explicit pytest.MonkeyPatch and TrtllmAttentionMetadata parameter annotations
and a None return annotation. Rename the local stop_at_base_prepare helper to
_stop_at_base_prepare, annotate its self parameter with TrtllmAttentionMetadata
and its return type as None, and use the renamed helper in the monkeypatch.
Source: Coding guidelines
Comment on lines
+26
to
+55
| def test_prepare_computes_draft_sliding_block_tables_before_base_prepare(monkeypatch): | ||
| """DeepSeek-V4 MTP draft KV managers need their sliding tables prepared. | ||
|
|
||
| The base TRT-LLM metadata prepare path copies block offsets from both the | ||
| target and draft managers. DeepSeek-V4's copy path consumes precomputed | ||
| sliding-window tables, so the draft manager must compute them before the | ||
| base prepare reaches copy_batch_block_offsets(). | ||
| """ | ||
| metadata = object.__new__(DeepseekV4TrtllmAttentionMetadata) | ||
| metadata.kv_cache_manager = MagicMock() | ||
| metadata.draft_kv_cache_manager = MagicMock() | ||
| metadata.request_ids = [11, 12, 13] | ||
| metadata.num_contexts = 2 | ||
|
|
||
| def stop_at_base_prepare(self): | ||
| raise RuntimeError("base prepare reached") | ||
|
|
||
| monkeypatch.setattr(TrtllmAttentionMetadata, "prepare", stop_at_base_prepare) | ||
|
|
||
| with pytest.raises(RuntimeError, match="base prepare reached"): | ||
| DeepseekV4TrtllmAttentionMetadata.prepare(metadata) | ||
|
|
||
| metadata.kv_cache_manager.compute_sliding_block_tables.assert_called_once_with( | ||
| metadata.request_ids, | ||
| metadata.num_contexts, | ||
| ) | ||
| metadata.draft_kv_cache_manager.compute_sliding_block_tables.assert_called_once_with( | ||
| metadata.request_ids, | ||
| metadata.num_contexts, | ||
| ) |
Contributor
There was a problem hiding this comment.
🎯 Functional Correctness | 🟡 Minor | ⚡ Quick win
🧩 Analysis chain
🏁 Script executed:
#!/bin/bash
set -euo pipefail
# Locate candidate CI and QA test-list files.
fd -t f . tests/integration/test_lists | sort
# Check whether this test module or DeepSeek-V4 test group is registered.
rg -n -i -C 2 \
'test_deepseek_v4_metadata|deepseek_v4' \
tests/integration/test_lists || true
# Locate available CBTS scope evidence.
fd -t f 'cbts_touchmap.sqlite' . || true
fd -t f '*coverage*' . | rg -i 'cbts|touchmap' || trueRepository: NVIDIA/TensorRT-LLM
Length of output: 12813
🏁 Script executed:
#!/bin/bash
set -euo pipefail
printf '%s\n' '--- changed files ---'
git diff --name-status
printf '%s\n' '--- test file diff ---'
git diff -- tests/unittest/_torch/attention/sparse/deepseek_v4/test_deepseek_v4_metadata.py
printf '%s\n' '--- matching test-list entries ---'
rg -n -F \
'unittest/_torch/attention/sparse/deepseek_v4/test_deepseek_v4_metadata.py' \
tests/integration/test_lists || true
printf '%s\n' '--- DeepSeek-V4 test-list entries ---'
rg -n -i -C 1 \
'unittest/_torch/attention/sparse/deepseek_v4' \
tests/integration/test_lists || true
printf '%s\n' '--- CBTS artifacts tracked or present ---'
git ls-files | rg -i '(^|/)(cbts_touchmap\.sqlite|.*cbts.*coverage.*|.*coverage.*cbts.*)$' || true
find . -type f \( -name 'cbts_touchmap.sqlite' -o -iname '*cbts*coverage*' -o -iname '*coverage*cbts*' \) -printRepository: NVIDIA/TensorRT-LLM
Length of output: 2077
Add the test module to tests/integration/test_lists/test-db/l0_b200.yml. test_deepseek_v4_metadata.py is absent from the CI and QA lists. Coverage summary: one test added; coverage verdict: insufficient.
🤖 Prompt for AI Agents
Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.
In
`@tests/unittest/_torch/attention/sparse/deepseek_v4/test_deepseek_v4_metadata.py`
around lines 26 - 55, Add
tests/unittest/_torch/attention/sparse/deepseek_v4/test_deepseek_v4_metadata.py
to the appropriate CI and QA entries in
tests/integration/test_lists/test-db/l0_b200.yml so
test_prepare_computes_draft_sliding_block_tables_before_base_prepare is included
in automated coverage.
Source: Path instructions
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
TrtllmAttentionMetadata.prepare()reachescopy_batch_block_offsets().Fixes #17024
Test Plan
python -m ruff check tensorrt_llm/_torch/attention_backend/sparse/deepseek_v4/deepseek_v4.py tests/unittest/_torch/attention/sparse/deepseek_v4/test_deepseek_v4_metadata.pypython -m py_compile tensorrt_llm/_torch/attention_backend/sparse/deepseek_v4/deepseek_v4.py tests/unittest/_torch/attention/sparse/deepseek_v4/test_deepseek_v4_metadata.pyimpi.dll/msmpi.dll) and compiled TensorRT-LLM bindings.Dev Engineer Review
DeepseekV4TrtllmAttentionMetadata.preparenow computes draft sliding-window block tables before the base metadata path copies draft block offsets._num_tablesinitialization failure in DeepSeek-V4 speculative decoding.QA Engineer Review
test_prepare_computes_sliding_tables_before_base_prepare.tests/integration/test_lists/,test-db/, orqa/entry was added or modified.