[TRTLLM-16621][feat] Support DeepSeek-V4.1-Flash

.gitignore

@@ -81,6 +81,7 @@ ad-test-workspace/
 */tllm_debug/**
 *.patch
 !cpp/tensorrt_llm/deep_ep/*.patch
+!3rdparty/patches/*.patch
 !3rdparty/vendor_patches/*.patch
 examples/disaggregated/slurm/benchmark/logs/
 scripts/attribution/data/checksum_to_paths.yml

3rdparty/fetch_content.json

@@ -39,10 +39,9 @@
     {
       "name": "deepgemm",
       "git_repository": "https://github.com/deepseek-ai/DeepGEMM",
-      "git_tag": "8b1392b978f5a03c828dd1711090d7fb50958b8a",
+      "git_tag": "b64107f2b9599ca76445b7f62eedf66bae1d095b",
       "git_submodules_recurse": true,
-      "source_subdir": "dont-add-this-project-with-add-subdirectory",
-      "patch_file": "patches/deepgemm_configurable_barrier_timeout.patch"
+      "source_subdir": "dont-add-this-project-with-add-subdirectory"
     },
     {
       "name": "eigen",
@@ -57,6 +56,12 @@
       "git_submodules_recurse": true,
       "source_subdir": "dont-add-this-project-with-add-subdirectory"
     },
+    {
+      "name": "fmt",
+      "git_repository": "https://github.com/fmtlib/fmt",
+      "git_tag": "553ec11ec06fbe0beebfbb45f9dc3c9eabd83d28",
+      "source_subdir": "dont-add-this-project-with-add-subdirectory"
+    },
     {
       "name": "googlebenchmark",
       "git_repository": "https://github.com/google/benchmark",

cpp/tensorrt_llm/batch_manager/kv_cache_manager_v2/blockRadixTree.cpp

@@ -482,7 +482,8 @@ std::vector<SharedPtr<Block>> Block::clearStaleBlocksAfterPageUnlink(
 
     // If this is a full-attention block or a sink block: evict subtree.
     // Mirrors Python: pages = remove_subtree(self)
-    if (alc && (!alc->windowSize.has_value() || block.ordinal() < BlockOrdinal{alc->numSinkBlocks}))
+    if (alc && alc->requiresPrefixCoverage()
+        && (!alc->windowSize.has_value() || block.ordinal() < BlockOrdinal{alc->numSinkBlocks}))
     {
         pruneStart = block.prev;
         detachedBlocks.push_back(removeSubtree(block));
@@ -887,6 +888,10 @@ std::vector<BlockRadixTree::MatchResult> BlockRadixTree::pruneMatch(
         bool shortened = false;
         for (auto [lcId, attn] : attnLcs)
         {
+            if (!attn->requiresPrefixCoverage())
+            {
+                continue;
+            }
             auto const staleRange = attn->getStaleRange(numTok, mTokensPerBlock);
             int const staleBeg = staleRange.beg.value();
             int const staleEnd = staleRange.end.value();

cpp/tensorrt_llm/batch_manager/kv_cache_manager_v2/config.h

@@ -137,6 +137,19 @@ inline void validateNoDuplicateBufferRoles(std::vector<BufferConfig> const& buff
 
 } // namespace detail
 
+//! Controls whether an attention lifecycle participates in prefix matching and page reuse.
+enum class AttentionReusePolicy
+{
+    //! Commit reusable pages and require live-window coverage when matching a prefix.
+    REQUIRED = 0,
+    //! Keep request-local working pages; never publish or reuse them through the radix tree.
+    PRIVATE = 1,
+    //! Commit reusable pages without constraining the REQUIRED prefix match. The caller selects
+    //! available groups at resume; missing or unselected groups use fresh pages for recomputation.
+    //! These pages may be evicted independently without removing the REQUIRED prefix.
+    OPTIONAL = 2,
+};
+
 struct AttentionLayerConfig
 {
     static constexpr LayerType type = LayerType::ATTENTION;
@@ -150,6 +163,9 @@ struct AttentionLayerConfig
     // nullopt or 0 = no sink tokens.
     std::optional<int> numSinkTokens;
 
+    // Private working pages never constrain prefix matching or enter the tree.
+    AttentionReusePolicy reusePolicy = AttentionReusePolicy::REQUIRED;
+
     [[nodiscard]] std::optional<int> windowSize() const noexcept
     {
         return slidingWindowSize;