[Nixl] Heterogeneous TP support FlashInfer (#20189)

Signed-off-by: NickLucche <nlucches@redhat.com>

vllm/distributed/kv_transfer/kv_connector/v1/nixl_connector.py

@@ -715,7 +715,7 @@ class NixlConnectorWorker:
         # are non-contiguous (it's not locally guaranteed that they will be)
         # Disadvantage is that the encoded NixlAgentMetadata is now larger
         # (roughly 8KB vs 5KB).
-        # Conversely for FlashInfer, K and V are transferred in the same tensor
+        # Conversely for FlashInfer, K and V are registered in the same region
         # to better exploit the memory layout (ie num_blocks is the first dim).
         split_k_and_v = not (self.use_mla or self._use_pallas_v1
                              or self._use_flashinfer)
@@ -758,12 +758,21 @@ class NixlConnectorWorker:
         assert tensor_size_bytes % self.num_blocks == 0
         self.block_len = tensor_size_bytes // self.num_blocks
         self.slot_size_bytes = self.block_len // self.block_size
+        self.device_kv_caches = kv_caches
+        self.dst_num_blocks[self.engine_id] = self.num_blocks
         if self._use_flashinfer:
             assert self.slot_size_bytes % 2 == 0
             self.slot_size_bytes /= 2
-        self.device_kv_caches = kv_caches
-        self.dst_num_blocks[self.engine_id] = self.num_blocks
 
+            # NOTE (NickLucche) When FlashInfer is used, memory is registered
+            # with joint KV for each block. This minimizes the overhead in
+            # registerMem allowing faster descs queries. In order to be able to
+            # split on kv_heads dim as required by heterogeneous TP, one must
+            # be able to index K/V separately. Hence the we double the number
+            # of 'virtual' regions here and halve `block_len` below.
+            self.num_regions *= 2
+
+        kv_block_len = self.get_backend_aware_kv_block_len()
         # Register local/src descr for NIXL xfer.
         blocks_data = []
         for base_addr in seen_base_addresses:
@@ -776,8 +785,18 @@ class NixlConnectorWorker:
                 block_offset = block_id * self.block_len
                 addr = base_addr + block_offset
                 # (addr, len, device id)
-                # TODO: does device_id matter to DRAM?
+                blocks_data.append((addr, kv_block_len, self.tp_rank))
-                blocks_data.append((addr, self.block_len, self.tp_rank))
+
+            if self._use_flashinfer:
+                # Separate and interleave K/V regions to maintain the same
+                # descs ordering. This is needed for selecting contiguous heads
+                # when split across TP ranks.
+                for block_id in range(self.num_blocks):
+                    block_offset = block_id * self.block_len
+                    addr = base_addr + block_offset
+                    # Register addresses for V cache (K registered first).
+                    v_addr = addr + kv_block_len
+                    blocks_data.append((v_addr, kv_block_len, self.tp_rank))
         logger.debug("Created %s blocks for src engine %s and rank %s",
                      len(blocks_data), self.engine_id, self.tp_rank)
 
@@ -903,7 +922,7 @@ class NixlConnectorWorker:
             remote_block_size = nixl_agent_meta.block_len // (
                 self.slot_size_bytes * tp_ratio)
             if self._use_flashinfer:
-                # Account for joint KV in FlashInfer.
+                # With flashinfer, KV are sent in the same message.
                 remote_block_size //= 2
             if tp_ratio > 1:
                 # Heterogeneous TP expects same kv_cache_layout.
@@ -929,10 +948,10 @@ class NixlConnectorWorker:
         # rank. With heterogeneous TP, prepare the descriptors by splitting the
         # P KV cache along kv_head dim, of D worker's kv_head size (D>P).
         # Eg. PTP1 DTP2 => P0 KV:[block0-KV_0 | block0-KV_1..].
-        # Only register the remote's descriptors if current rank pulls from it.
         self.kv_caches_base_addr[
             engine_id] = nixl_agent_meta.kv_caches_base_addr
-        rank_offset = self.tp_rank % tp_ratio * self.block_len \
+        kv_block_len = self.get_backend_aware_kv_block_len()
+        rank_offset = self.tp_rank % tp_ratio * kv_block_len \
             if not (self.use_mla or is_kv_replicated) else 0
         # Register all remote blocks, but only the corresponding kv heads.
         for base_addr in nixl_agent_meta.kv_caches_base_addr:
@@ -943,7 +962,16 @@ class NixlConnectorWorker:
                 # self.block_len == remote_block_len//tp_ratio bytes.
                 addr = base_addr + block_offset + rank_offset
                 # (addr, len, device id)
-                blocks_data.append((addr, self.block_len, remote_tp_rank))
+                blocks_data.append((addr, kv_block_len, remote_tp_rank))
+
+            if self._use_flashinfer:
+                # With FlashInfer index V separately to allow head splitting.
+                for block_id in range(nixl_agent_meta.num_blocks):
+                    block_offset = block_id * nixl_agent_meta.block_len
+                    addr = base_addr + block_offset + rank_offset
+                    v_addr = addr + nixl_agent_meta.block_len // 2
+                    blocks_data.append((v_addr, kv_block_len, remote_tp_rank))
+
         logger.debug(
             "Created %s blocks for dst engine %s with remote rank %s and "
             "local rank %s", len(blocks_data), engine_id, remote_tp_rank,
@@ -1249,6 +1277,22 @@ class NixlConnectorWorker:
                 descs_ids.append(reg_id * num_blocks + block_id)
         return descs_ids
 
+    def get_backend_aware_kv_block_len(self):
+        """
+        Get the block length for one K/V element (K and V have the same size).
+
+        For FA and other backends, this is equal to the length of the whole 
+        block, as K and V are in separate regions.
+        For FlashInfer, this is half the length of the whole block, as K and V
+        share the same region.
+        """
+        if self._use_flashinfer:
+            # For indexing only half (either just the K or V part).
+            block_len = self.block_len // 2
+        else:
+            block_len = self.block_len
+        return block_len
+
 
 @contextlib.contextmanager
 def zmq_ctx(socket_type: Any, addr: str) -> Iterator[zmq.Socket]: