[unified-memory] Drop the vacated 'dense' qualifier and the restating comments (#37170)

python/sglang/kernels/ops/kvcache/kv_read_table.py

@@ -21,8 +21,8 @@ the result into `out`:
                 for c < ceil(seq_lens[b] / ps)          -- the row's LIVE prefix
 
 `v2p` is the pool's virtual->physical page table and `multiplier` scales a
-physical page into the id space the per-layer views use (1 when they are not
-dense). Since only the page number is rewritten, a token-level consumer can
+physical page into the id space the per-layer views use (1 when one page maps
+to one row-block). Since only the page number is rewritten, a token-level consumer can
 rebuild flat ids as `entry * ps + offset`.
 
 PREFIX-ONLY per row: columns past the live prefix are never written, so a

python/sglang/srt/layers/attention/trtllm_mla_backend.py

@@ -396,7 +396,7 @@ class TRTLLMMLABackend(FlashInferMLAAttnBackend):
         self.decode_cuda_graph_kv_indices = torch.full(
             (max_bs, max_blocks_per_seq), -1, dtype=torch.int32, device=self.device
         )
-        # Unified pool: capture-stable buffer for the DENSE KV write loc, filled
+        # Unified pool: capture-stable buffer for the kernel-facing KV write loc, filled
         # out-of-graph in init_forward_metadata_out_graph so the in-graph
         # set_mla_kv_buffer captures no translate.
         if self.kv_index_translator.is_translating:
@@ -635,7 +635,7 @@ class TRTLLMMLABackend(FlashInferMLAAttnBackend):
             # Replay-prep receives the RAW (unpadded) out_cache_loc
             # (build_replay_fb_view), but the captured write kernel consumes the
             # full captured tier of this buffer. Zero the tail so pad rows write
-            # to the dense sink (row 0) instead of stale dense locs left by
+            # to the sink (row 0) instead of stale kernel-facing locs left by
             # earlier larger replays — a stale tail scatters pad-row garbage into
             # live KV pages. Mirrors the runner's PaddingPolicy.ZERO on its own
             # out_cache_loc slot.

python/sglang/srt/mem_cache/memory_pool.py

@@ -1697,8 +1697,8 @@ class KVCache(abc.ABC):
         self.size = size
         self.page_size = page_size
         # Row-blocks one page holds in this pool's kernel-facing id space; >1
-        # only where the per-layer views are dense (the unified pool), and then
-        # a write loc must have been translated into that space first.
+        # only for the unified pool's per-layer views, and then a write loc must
+        # have been translated into that space first.
         self.kernel_page_blocks = 1
         self.dtype = dtype
         self.device = device
@@ -2790,7 +2790,7 @@ class MHATokenToKVPool(KVCache):
         num_rows = int(loc_2d.numel())
         if cache_k.shape[0] != num_rows or cache_v.shape[0] != num_rows:
             raise ValueError(
-                "dense KV rows must match loc_2d size: "
+                "KV rows must match loc_2d size: "
                 f"{tuple(cache_k.shape)=} {tuple(cache_v.shape)=} {tuple(loc_2d.shape)=}."
             )
 

python/sglang/srt/mem_cache/unified_memory_pool.py

@@ -1192,7 +1192,7 @@ def init_unified_mamba_pools(
         pre_alloc_size=decode_pre_alloc_size,
     )
     if use_mla_backend:
-        # start_layer stays 0: HybridLinearKVPool patches layer ids to the dense
+        # start_layer stays 0: HybridLinearKVPool patches layer ids to the contiguous
         # 0..N-1 index via _transfer_id_context before every MLA pool call.
         unified_full_kv_pool = UnifiedMLATokenToKVPool(
             unified_buffer=shared_pool,