[pull] main from NVIDIA:main

docs/source/legacy/performance/perf-benchmarking.md

@@ -706,7 +706,7 @@ kv_cache_dtype: "fp8"
 ```
 
 ```{tip}
-The two valid values for `kv_cache_dtype` are `auto` and `fp8`.
+The valid values for `kv_cache_dtype` are `auto`, `fp8`, and `nvfp4`.
 ```
 
 ## Low Latency Benchmark

jenkins/L0_MergeRequest.groovy

@@ -342,15 +342,12 @@ def mergeWaiveList(pipeline, globalVars)
 
     try {
         // Get TOT waive list
-        LLM_TOT_ROOT = "llm-tot"
         targetBranch = env.gitlabTargetBranch ? env.gitlabTargetBranch : globalVars[TARGET_BRANCH]
         echo "Target branch: ${targetBranch}"
         withCredentials([string(credentialsId: 'default-llm-repo', variable: 'DEFAULT_LLM_REPO')]) {
-            trtllm_utils.checkoutSource(DEFAULT_LLM_REPO, targetBranch, LLM_TOT_ROOT, false, true)
+            trtllm_utils.checkoutFile(DEFAULT_LLM_REPO, targetBranch, "tests/integration/test_lists/waives.txt", ".")
         }
-        targetBranchTOTCommit = sh (script: "cd ${LLM_TOT_ROOT} && git rev-parse HEAD", returnStdout: true).trim()
-        echo "Target branch TOT commit: ${targetBranchTOTCommit}"
-        sh "cp ${LLM_TOT_ROOT}/tests/integration/test_lists/waives.txt ./waives_TOT_${targetBranchTOTCommit}.txt"
+        sh "mv waives.txt waives_TOT.txt"
 
         // Get waive list diff in current MR
         def diff = getMergeRequestOneFileChanges(pipeline, globalVars, "tests/integration/test_lists/waives.txt")
@@ -362,7 +359,7 @@ def mergeWaiveList(pipeline, globalVars)
         sh """
             python3 mergeWaiveList.py \
             --cur-waive-list=waives_CUR_${env.gitlabCommit}.txt \
-            --latest-waive-list=waives_TOT_${targetBranchTOTCommit}.txt \
+            --latest-waive-list=waives_TOT.txt \
             --diff-file=diff_content.txt \
             --output-file=waives.txt
         """

tensorrt_llm/_torch/auto_deploy/custom_ops/README.md

@@ -131,4 +131,4 @@ The table below lists the operators grouped by category.
 | Operator Name | Description |
 |--------------|-------------|
 | `torch.ops.auto_deploy.triton_utils_fused_gather_scatter` | Triton fused gather + scatter for overlap scheduling input_ids reordering |
-| `torch.ops.auto_deploy.gather_logits_before_lm_head` | Gather hidden states using logits indices before LM head |
+| `torch.ops.auto_deploy.gather_tokens` | Gather hidden states using token indices before LM head |

tensorrt_llm/_torch/auto_deploy/custom_ops/utils/__init__.py

@@ -16,11 +16,13 @@
 """Utility operations.
 
 This module provides utility functions and helpers:
+- block_table_ragged: Block table <-> ragged tensor conversions
 - torch_gather_logits: Logit gathering operations
 - triton_utils: Triton utility functions and helpers
 """
 
 __all__ = [
+    "block_table_ragged",
     "torch_gather_logits",
     "triton_utils",
 ]