[Docs] Fix accidental cross-references in docstrings

vllm/_custom_ops.py

@@ -3355,8 +3355,8 @@ def swap_blocks(
     and their matching destination block number on the destination tensor.
     block_mapping is expected to be a tensor of shape (num_blocks_to_copy, 2)
     where each block_mapping[i] represents a single copy operation, copying
-    block #block_mapping[i][0] from the source tensor
-    to block #block_mapping[i][1] on the destination tensor.
+    block `block_mapping[i][0]` from the source tensor
+    to block `block_mapping[i][1]` on the destination tensor.
     block_mapping should have dtype int64.
 
     The source and the destination tensors can be either on cpu or gpu,

vllm/models/kimi_k3/amd/specialized_decode/mono/attention/back.py

@@ -103,7 +103,7 @@ def sent_on(c, which="ar"):
 
 def sent_ix(src, col, t, s, width):
     """(OPT_SENT) Word index of (src rank, pair at ``col``, token t), blocked
-    [src][16-col group][token][8 pairs]; shared by push and reduce."""
+    `[src][16-col group][token][8 pairs]`; shared by push and reduce."""
     return ((src * (width // 16) + col // 16) * s + t) * 8 + (col % 16) // 2
 
 

vllm/models/kimi_k3/amd/specialized_decode/mono/layer.py

@@ -215,7 +215,7 @@ def sent_bytes() -> int:
 
 def rlogp_half_bytes() -> int:
     """(OPT_RLOG_PARTS) One epoch parity's RLOGP, tagged pairs
-    [src rank][part][token][router row], SENT_ROWS rows."""
+    `[src rank][part][token][router row]`, SENT_ROWS rows."""
     return TP_ * k2b.R_PARTS * SENT_ROWS * (k2b.E // TP_) * 8
 
 

vllm/models/kimi_k3/amd/specialized_decode/mono/stages/moe.py

@@ -531,7 +531,7 @@ def _route_out(c, t, red, sel, rank, me, ms):
 @traced
 def route_parts(c, t):
     """(OPT_RLOG_PARTS) Token t's router partials (every rank's 4 K
-    parts, RLOGP) -> LDS ``vals`` [part][expert], wave q < R_PARTS polling part
+    parts, RLOGP) -> LDS `vals[part][expert]`, wave q < R_PARTS polling part
     q (one batch: the 14 experts of its lane); ``stage_route`` sums each
     expert's parts in part order, as the router's part 0 did."""
     s, lane, wave = c["S"], c["lane"], c["wave"]

vllm/models/minimax_m3/amd/mono/layout.py

@@ -110,8 +110,9 @@ def _layout(items, align=256):
 
 
 def sym_layout(npes: int):
-    """Per-rank symmetric buffer: attn / ffn partial regions, [src rank][token][HIDDEN],
-    and the peers' index-selection candidates (``sel_cp``)."""
+    """Per-rank symmetric buffer: attn / ffn partial regions,
+    `[src rank][token][HIDDEN]`, and the peers' index-selection candidates
+    (``sel_cp``)."""
     rows = npes * MAX_TOKENS * HIDDEN // 2
     # a_ag / ffn_ag: the finished attention / FFN all-reduce rows, each written by
     # its owner rank (the reduce-scatter + all-gather form, S >= AG_RS_FROM)

vllm/models/muse_spark/amd/mono/kernels/attn_frag.py

@@ -30,8 +30,8 @@
 
 Shapes (a TP8 rank): 16 q heads, 2 kv heads (QPK = 8), head dim 64, 8 query tokens a request,
 NR requests (K3 S8: 1, S16: 2; HA S32: 4). K / V bf16 in the layer's vLLM LBHNC cache
-[block][2][16][K 64 | V 64] (block 16); the page table is the layer group's block_table
-(int32 [NR][stride]); seq_used = vLLM's seq_lens (int32 [NR], keys incl. the 8 new ones).
+`[block][2][16][K 64 | V 64]` (block 16); the page table is the layer group's block_table
+(int32 `[NR][stride]`); seq_used = vLLM's seq_lens (int32 [NR], keys incl. the 8 new ones).
 q [8 NR, 16, 64] bf16 (row stride 1024, without the QK scale); out attn_o [8 NR, 1024] bf16.
 SWA: vLLM's window (2047, 0), keys qpos - key < 2048, and only keys in the window are read
 (vLLM frees the blocks before it: their block-table entries are stale).
@@ -633,8 +633,8 @@ def vllm_scale_word(scale):
 
 def vllm_desc(q, kv_cache, block_table, seq_lens, out, part, nreq, scale):
     """The 10 descriptor words of one layer: q [8 nreq.., 16 x 64] bf16 rows,
-    kv_cache the layer's [block][2][16][128] bf16 view (block stride = kv_cache.stride(0)), block_table int32
-    [>= nreq][stride] of the layer's group, seq_lens int32 [>= nreq], out attn_o [rows, 1024] bf16, part the scratch
+    kv_cache the layer's `[block][2][16][128]` bf16 view (block stride = kv_cache.stride(0)), block_table int32
+    `[>= nreq][stride]` of the layer's group, seq_lens int32 [>= nreq], out attn_o [rows, 1024] bf16, part the scratch
     (int32, part_bytes()), nreq the requests of the step (graph padding included: seq_len 0 = no keys, zero rows)."""
     import torch