tensorrt_llm/_torch/pyexecutor/resource_manager.py (1)

555-558: Validate fraction range.

Consider asserting lower bound too to catch 0 or negative values: 0.0 < free_mem_fraction < 1.0.

-        assert free_mem_fraction < 1.0, f"Invalid freeMemFraction, freeMemFraction {free_mem_fraction} must be smaller than 1.0"
+        assert 0.0 < free_mem_fraction < 1.0, (
+            f"Invalid freeMemFraction: {free_mem_fraction}. Must be in (0.0, 1.0)."
+        )

docs/source/reference/support-matrix.md (1)

162-162: SM designations are valid; small terminology tweak optional.

“Blackwell (SM100/SM120)” aligns with CUDA 12.8 notes (SM_100/SM_120). Consider using “NVFP4” instead of “FP4” for precision naming consistency with code. (docs.nvidia.com)

docs/source/overview.md (1)

27-29: Nit: spacing and terminology.

• Remove the stray space before the comma after the B200 link.
• Prefer “NVFP4” (brand) once, then “FP4” thereafter, to match code/constants.

- [NVIDIA B200 GPUs](https://www.nvidia.com/en-us/data-center/dgx-b200/) , when used with TensorRT-LLM, enable seamless loading of model weights in the new [FP4 format](https://developer.nvidia.com/blog/introducing-nvfp4-for-efficient-and-accurate-low-precision-inference/#what_is_nvfp4), allowing you to automatically leverage optimized FP4 kernels for efficient and accurate low-precision inference.
+ [NVIDIA B200 GPUs](https://www.nvidia.com/en-us/data-center/dgx-b200/), when used with TensorRT-LLM, enable seamless loading of model weights in the new [NVFP4 format](https://developer.nvidia.com/blog/introducing-nvfp4-for-efficient-and-accurate-low-precision-inference/#what_is_nvfp4), allowing you to automatically leverage optimized FP4 kernels for efficient and accurate low-precision inference.

tensorrt_llm/_torch/attention_backend/trtllm.py (1)

846-851: Zero only the written slices to avoid unnecessary work.

Full-buffer fill_ can be sizable; zero the relevant regions instead and prefer zero_() for clarity.

-        self.kv_block_ids_per_seq.fill_(0)
-        self.kv_block_ids_per_seq[:self.num_seqs, :num_blocks].copy_(
+        self.kv_block_ids_per_seq[:self.num_seqs, :num_blocks].zero_()
+        self.kv_block_ids_per_seq[:self.num_seqs, :num_blocks].copy_(
             block_ids_per_seq, non_blocking=True)
-        self.block_ids_per_seq.fill_(0)
-        self.block_ids_per_seq[:self.num_generations, :num_blocks].copy_(
+        self.block_ids_per_seq[:self.num_generations, :num_blocks].zero_()
+        self.block_ids_per_seq[:self.num_generations, :num_blocks].copy_(
             block_ids_per_seq[self.num_contexts:], non_blocking=True)

tests/unittest/_torch/speculative/test_eagle3.py (1)

331-332: Optional: avoid running unittest main when skipping heavy tests.

Since pytest drives CI, dropping the __main__ block reduces accidental local execution. Keep if needed for local workflows.

tensorrt_llm/models/gemma/model.py (1)

141-146: Guard tuple-unpacking of hidden_states under reduce_fusion.

The FIXME hints potential shape mismatches. Add a defensive check to avoid crashes when upstream doesn’t return a (hidden_states, residual) tuple for some ranks/configs.

-        if default_net(
-        ).plugin_config.reduce_fusion and self.local_layer_idx > 0:
-            hidden_states, residual = hidden_states  #FIXME:AN need to check if appropriate residual value is hidden state is pulled out.
+        if default_net().plugin_config.reduce_fusion and self.local_layer_idx > 0:
+            if isinstance(hidden_states, tuple):
+                hidden_states, residual = hidden_states
+            else:
+                # Fallback to non-fused contract
+                residual = hidden_states
+                hidden_states = self.input_layernorm(hidden_states)
         else:
             residual = hidden_states
             hidden_states = self.input_layernorm(hidden_states)

tests/unittest/llmapi/test_mpi_session.py (2)

60-66: Resolve path once; reuse for cwd to avoid recomputing.

Minor cleanup: reuse cur_dir for cwd instead of recomputing os.path.dirname(os.path.abspath(__file__)).

-    with Popen(command,
+    with Popen(command,
                env=os.environ,
                stdout=PIPE,
                stderr=PIPE,
                bufsize=1,
                start_new_session=True,
                universal_newlines=True,
-               cwd=os.path.dirname(os.path.abspath(__file__))) as process:
+               cwd=cur_dir) as process:

---

62-62: Silence Ruff S101 for test assertion (intended in tests).

Keep the fast-fail assert but add a local noqa to satisfy Ruff.

-    assert os.path.exists(test_file), f"Test file {test_file} does not exist"
+    assert os.path.exists(test_file), f"Test file {test_file} does not exist"  # noqa: S101

tensorrt_llm/_torch/pyexecutor/_util.py (1)

181-184: Lazy-init dummy requests here is good; also guard in configure_kv_cache_capacity.

Some callers might jump straight to configure_kv_cache_capacity(); add a safety init there to avoid None usage.

@@
-        py_executor.set_gather_responses(True)
+        # Ensure dummy requests exist even if try_prepare_estimation() was skipped.
+        if self._dummy_reqs is None:
+            self._dummy_reqs = self._create_dummy_context_requests(
+                max(1, self._max_seq_len - 1)
+            )
+        py_executor.set_gather_responses(True)

docs/source/llm-api/index.md (3)

68-68: Fix markdownlint MD026: remove trailing punctuation from heading

Drop the period at the end of this heading to satisfy MD026.

-### MPI_ABORT was invoked on rank 1 in communicator MPI_COMM_WORLD with errorcode 1.
+### MPI_ABORT was invoked on rank 1 in communicator MPI_COMM_WORLD with errorcode 1

---

78-79: Grammar/typo fixes in guidance

s/an contextmanager/a context manager/ and s/methed/method/.

-  2. Use LLM as an contextmanager, with the following code: `with LLM(...) as llm: ...`, the shutdown methed will be invoked automatically once it goes out of the `with`-statement block.
+  2. Use LLM as a context manager, with the following code: `with LLM(...) as llm: ...`. The shutdown method will be invoked automatically once it exits the `with` block.

---

82-93: Clarify the docker networking workaround scope

The mpi4py workaround suggests avoiding --net=host; note explicitly that --ipc=host is the required flag, and keep --net default. This helps users not set both.

-The root cause may be related to `mpi4py`. There is a [workaround](https://github.com/mpi4py/mpi4py/discussions/491#discussioncomment-12660609) suggesting a change from `--net=host` to `--ipc=host`, or setting the following environment variables:
+The root cause may be related to `mpi4py`. A documented [workaround](https://github.com/mpi4py/mpi4py/discussions/491#discussioncomment-12660609) is to avoid `--net=host` and use `--ipc=host` instead, or set the following environment variables:

jenkins/L0_Test.groovy (1)

2105-2107: Avoid magic number 7 for post-merge splits

Make split count a named constant to keep Jenkins and test-db YAML in sync and ease future changes.

-    multiNodesSBSAConfigs += (1..7).collectEntries { i ->
-        ["GB200-8_GPUs-2_Nodes-PyTorch-Post-Merge-${i}".toString(), ["gb200-multi-node", "l0_gb200_multi_nodes", i, 7, 8, 2]]
-    }
+    // Use a single source of truth for split count.
+    final int POST_MERGE_SPLITS = 7
+    multiNodesSBSAConfigs += (1..POST_MERGE_SPLITS).collectEntries { i ->
+        ["GB200-8_GPUs-2_Nodes-PyTorch-Post-Merge-${i}".toString(),
+         ["gb200-multi-node", "l0_gb200_multi_nodes", i, POST_MERGE_SPLITS, 8, 2]]
+    }

If preferred, I can place POST_MERGE_SPLITS next to other @field constants.

tensorrt_llm/llmapi/llm_utils.py (1)

769-771: Don’t delete attributes from llm_args

Deleting the mpi_session attribute mutates the schema and risks surprises for later users. Prefer setting it to None.

-                    if self.llm_args.mpi_session:
-                        del self.llm_args.mpi_session
+                    if getattr(self.llm_args, "mpi_session", None) is not None:
+                        self.llm_args.mpi_session = None

Configuration used: Path: .coderabbit.yaml

Review profile: CHILL

Plan: Pro

Learnt from: EmmaQiaoCh
PR: NVIDIA/TensorRT-LLM#7370
File: tests/unittest/trt/model_api/test_model_quantization.py:24-27
Timestamp: 2025-08-29T14:07:45.863Z
Learning: In TensorRT-LLM's CI infrastructure, pytest skip markers (pytest.mark.skip) are properly honored even when test files have __main__ blocks that call test functions directly. The testing system correctly skips tests without requiring modifications to the __main__ block execution pattern.

Learnt from: thorjohnsen
PR: NVIDIA/TensorRT-LLM#6910
File: cpp/tensorrt_llm/batch_manager/kvCacheManager.cpp:0-0
Timestamp: 2025-08-14T21:04:50.248Z
Learning: In KV cache onboarding logic during prefill in cpp/tensorrt_llm/batch_manager/kvCacheManager.cpp, when calculating which blocks fall within the attention window, use getTokensPerBlock() to advance token indices rather than block->getUniqueTokens().size(), because the calculation needs to consider the post-prefill state where blocks will be filled to capacity, not their current token count.

Learnt from: eopXD
PR: NVIDIA/TensorRT-LLM#6768
File: cpp/tensorrt_llm/batch_manager/kvCacheManager.cpp:2010-2045
Timestamp: 2025-08-21T09:41:49.347Z
Learning: In cpp/tensorrt_llm/batch_manager/kvCacheManager.cpp, updateSequenceCacheBlockOffsets is specifically for updating bookkeeping when blocks are added during the context phase, not for refreshing offsets after detach operations. During detach operations, GenerationRequest::removeFrontBlock handles the necessary cache block bookkeeping internally.

Learnt from: ChristinaZ
PR: NVIDIA/TensorRT-LLM#7068
File: cpp/tensorrt_llm/kernels/moeTopKFuncs.cuh:169-172
Timestamp: 2025-08-20T07:43:36.447Z
Learning: In TensorRT-LLM MOE kernels, when processing up to 128 experts across 32 threads, each thread handles at most 4 experts (N < 5 constraint), where N represents candidates per thread rather than total system capacity.

Learnt from: eopXD
PR: NVIDIA/TensorRT-LLM#6768
File: cpp/tensorrt_llm/batch_manager/kvCacheManager.cpp:577-579
Timestamp: 2025-08-20T06:56:02.889Z
Learning: In cpp/tensorrt_llm/batch_manager/kvCacheManager.cpp, maxSequenceLength is now enforced as a non-optional argument in the BlockManager constructor, so concerns about std::nullopt defaulting to 0 are not applicable. When windowSize > maxSequenceLength, a warning should be added instead of handling optional parameter cases.

Learnt from: Funatiq
PR: NVIDIA/TensorRT-LLM#6754
File: tests/integration/test_lists/test-db/l0_a30.yml:41-47
Timestamp: 2025-08-13T11:07:11.772Z
Learning: In TensorRT-LLM test configuration files like tests/integration/test_lists/test-db/l0_a30.yml, TIMEOUT values are specified in minutes, not seconds.