Add files using upload-large-folder tool

README.md

@@ -54,6 +54,6 @@ The exact sources the image was built from — `code/` in this repo is byte-iden
 | tt-metal | a local checkout — commit not published *(dirty tree — the image includes uncommitted changes)* |
 | vLLM | [`v0.26.0`](https://github.com/vllm-project/vllm/releases/tag/v0.26.0) |
 | vllm-tt-plugin | a local checkout — commit not published |
-| `code/` digest | `39e6b36b56f864a9` (sha256, first 16 hex digits) |
-| built | 2026-09-27T10:15:01+00:00 by tt-model 0.1.0 |
+| `code/` digest | `3e629a797061595a` (sha256, first 16 hex digits) |
+| built | 2026-09-29T16:25:00+00:00 by tt-model 0.1.0 |
 

code/models/autoports/qwen_qwen3_5_122b_a10b/tt/moe_dispatch.py

@@ -21,6 +21,8 @@ from __future__ import annotations
 
 import os
 
+import torch
+
 import ttnn
 
 TILE = 32
@@ -47,7 +49,12 @@ def dispatch_rows(T: int, k: int, E: int, group: int = TILE) -> int:
 #   "sort_u16": ttnn.sort on uint16 ids (DRAM multi-core factory; hang-free but ~4 ms per 512-token chunk);
 #   "sort_fp32": the original fp32 ttnn.sort (cross-core-exchange factory; its sort_barrier deadlocks intermittently on
 #   one device of the (1,4) mesh - the 48-layer warm-up hang, doc/full_model/triage/warm_hang_*.txt).
-DISPATCH_MODE = os.environ.get("QWEN35_DISPATCH_MODE", "counting_h")  # "counting" | "sort_u16" | "sort_fp32"
+#   "counting_mm" (default, stage 12): counting sort whose running count is a BLOCK-TRIANGULAR MATMUL (fp32 accumulate, exact
+#   integers) instead of ttnn.cumsum on the [Rp, E] one-hot: the dim-2 cumsum ran on ~4 cores and needed two full-tensor
+#   permutes (25 of the 53 ms MoE device time per 2048-token layer, stage-12 profile); the matmuls use the whole grid.
+DISPATCH_MODE = os.environ.get(
+    "QWEN35_DISPATCH_MODE", "counting_mm"
+)  # "counting_mm" | "counting_h" | "counting" | "sort_u16" | "sort_fp32"
 SORT_ON_UINT16 = DISPATCH_MODE == "sort_u16"
 
 
@@ -85,13 +92,71 @@ class SortedDispatch:
                 "zeros_g": ttnn.zeros((1, 1, 1, G), dtype=u32, layout=RM, device=dev),
                 "zeros_rp": ttnn.zeros((1, 1, 1, Rp), dtype=u32, layout=RM, device=dev),
             }
+            if DISPATCH_MODE == "counting_mm":
+                # block-triangular running count: rows in blocks of Bk (the smallest tile multiple giving at most 32
+                # blocks, so the block prefix is one 32 x 32 tile); Rp is zero-padded up to nb * Bk when Bk does not
+                # divide it (e.g. T=500: Rp=4000 -> Bk=128, 32 blocks, Rp_pad=4096). L = lower-triangular ones incl. the
+                # diagonal (within-block inclusive count), P = strictly-lower-triangular ones over the (up to 32) blocks
+                Bk = TILE * (-(-Rp // (32 * TILE)))
+                nb = -(-Rp // Bk)
+                assert nb <= 32 and nb * Bk >= Rp, (Rp, Bk, nb)
+                c["Rp_pad"] = nb * Bk
+                L = torch.tril(torch.ones(Bk, Bk, dtype=torch.float32))
+                P = torch.tril(torch.ones(32, 32, dtype=torch.float32), diagonal=-1)
+                c["Bk"] = Bk
+                c["nb"] = nb
+                c["L_tri"] = ttnn.from_torch(L.reshape(1, 1, Bk, Bk), dtype=f32, layout=ttnn.TILE_LAYOUT, device=dev)
+                c["P_excl"] = ttnn.from_torch(P.reshape(1, 1, 32, 32), dtype=f32, layout=ttnn.TILE_LAYOUT, device=dev)
+                c["ones_row_bk"] = ttnn.full((1, 1, TILE, Bk), 1.0, dtype=f32, layout=ttnn.TILE_LAYOUT, device=dev)
             if len(self._const) >= CONST_CACHE_ENTRIES:  # bounded: drop the oldest length's constants
                 oldest = next(iter(self._const))
                 for t in self._const.pop(oldest).values():
-                    ttnn.deallocate(t)
+                    if not isinstance(t, int):
+                        ttnn.deallocate(t)
             self._const[T] = c
         return c
 
+    def _running_count_mm(self, oh, consts, Rp, E, f32_ckc):
+        """Inclusive running count of the one-hot rows along the token axis (== ttnn.cumsum(oh, dim=2)) as matmuls:
+        within each block of Bk rows ``L_tri @ oh_block`` (batched over the blocks), plus the exclusive prefix of the
+        block totals (``P_excl @ totals``) broadcast over the block's rows. All operands are 0/1 or small integers, the
+        accumulation is fp32, so every count is exact."""
+        Bk, nb, Rp_pad = consts["Bk"], consts["nb"], consts["Rp_pad"]
+        f32 = ttnn.float32
+        dram = ttnn.DRAM_MEMORY_CONFIG
+        oh_src = oh
+        if Rp_pad > Rp:  # zero rows appended: they add nothing to any count and are sliced off below
+            oh_src = ttnn.pad(oh, [(0, 0), (0, 0), (0, Rp_pad - Rp), (0, 0)], value=0.0)
+        oh_b = ttnn.reshape(oh_src, (1, nb, Bk, E))  # tile-aligned view: Bk % 32 == 0
+        cs_local = ttnn.matmul(consts["L_tri"], oh_b, compute_kernel_config=f32_ckc, dtype=f32, memory_config=dram)
+        if nb > 1:
+            tot = ttnn.matmul(consts["ones_row_bk"], oh_b, compute_kernel_config=f32_ckc, dtype=f32, memory_config=dram)
+            tot = ttnn.slice(tot, (0, 0, 0, 0), (1, nb, 1, E))  # [1,nb,1,E]: block totals (row 0 of every 32-row block)
+            tot_rows = ttnn.reshape(ttnn.to_layout(tot, ttnn.ROW_MAJOR_LAYOUT), (1, 1, nb, E))
+            ttnn.deallocate(tot)
+            if nb < 32:
+                tot_rows = ttnn.pad(tot_rows, [(0, 0), (0, 0), (0, 32 - nb), (0, 0)], value=0.0)
+            tot_rows = ttnn.to_layout(tot_rows, ttnn.TILE_LAYOUT)
+            pre = ttnn.matmul(consts["P_excl"], tot_rows, compute_kernel_config=f32_ckc, dtype=f32, memory_config=dram)
+            ttnn.deallocate(tot_rows)  # [1,1,32,E]: exclusive block prefix in rows 0..nb-1
+            pre_rows = ttnn.reshape(
+                ttnn.to_layout(ttnn.slice(pre, (0, 0, 0, 0), (1, 1, nb, E)), ttnn.ROW_MAJOR_LAYOUT), (1, nb, 1, E)
+            )
+            ttnn.deallocate(pre)
+            pre_rows = ttnn.to_layout(pre_rows, ttnn.TILE_LAYOUT)
+            cs = ttnn.add(cs_local, pre_rows, memory_config=dram)  # broadcast over the Bk rows of every block
+            ttnn.deallocate(cs_local)
+            ttnn.deallocate(pre_rows)
+        else:
+            cs = cs_local
+        cs = ttnn.reshape(cs, (1, 1, Rp_pad, E))
+        if Rp_pad > Rp:
+            ttnn.deallocate(oh_src)
+            full = cs
+            cs = ttnn.slice(full, (0, 0, 0, 0), (1, 1, Rp, E), memory_config=dram)
+            ttnn.deallocate(full)
+        return cs
+
     @staticmethod
     def _f32(t):
         return t if t.dtype == ttnn.float32 else ttnn.typecast(t, ttnn.float32)
@@ -134,7 +199,7 @@ class SortedDispatch:
         f32_ckc = ttnn.WormholeComputeKernelConfig(
             math_fidelity=ttnn.MathFidelity.HiFi4, math_approx_mode=False, fp32_dest_acc_en=True, packer_l1_acc=False
         )
-        if DISPATCH_MODE in ("counting", "counting_h"):
+        if DISPATCH_MODE in ("counting", "counting_h", "counting_mm"):
             # ---- counting sort (no ttnn.sort): every token-major row i keeps its position; its destination is
             # offsets[e_i] + (rank of i among the rows of expert e_i). The rank is the running count of the one-hot
             # along the token axis, picked through the one-hot. All integers stay exact: counts/cumsum in fp32,
@@ -209,7 +274,10 @@ class SortedDispatch:
                 off_i = ttnn.add(ttnn.multiply(sel_hi, 256.0), sel_lo)  # [1,1,Rp,1]
                 ttnn.deallocate(sel_hi)
                 ttnn.deallocate(sel_lo)
-                cs = ttnn.cumsum(oh, dim=2, dtype=f32)  # [1,1,Rp,E]
+                if DISPATCH_MODE == "counting_mm":
+                    cs = self._running_count_mm(oh, consts, Rp, E, f32_ckc)  # [1,1,Rp,E] fp32, == cumsum(oh, dim=2)
+                else:
+                    cs = ttnn.cumsum(oh, dim=2, dtype=f32)  # [1,1,Rp,E]
                 cs_hi = ttnn.floor(ttnn.multiply(cs, 1.0 / 256.0))
                 cs_lo = ttnn.subtract(cs, ttnn.multiply(cs_hi, 256.0))
                 ttnn.deallocate(cs)
@@ -337,6 +405,9 @@ class SortedDispatch:
         ttnn.deallocate(grp_of_row)
         if src_ids is not sorted_ids:
             ttnn.deallocate(src_ids)
+        # expert id of every group as the uint16 ROW_MAJOR stick the group-parallel sparse matmul (`indices=`) reads;
+        # padding groups carry expert 0 (their rows are zero-weighted and never read back)
+        group_ids = ttnn.reshape(ttnn.typecast(grp_expert, ttnn.uint16), (1, G))  # uint32 RM -> uint16 RM [1, G]
         grp_f = self._f32(ttnn.to_layout(grp_expert, ttnn.TILE_LAYOUT))
         ttnn.deallocate(grp_expert)
         group_mask = ttnn.eq(ttnn.reshape(grp_f, (1, G, 1, 1)), self.arange_E)  # [1,G,1,E] fp32 one-hot
@@ -362,6 +433,7 @@ class SortedDispatch:
             "src_tok": src_tok_rm,
             "weights": weights,
             "group_mask": group_mask,
+            "group_ids": group_ids,
             "undispatch": undispatch,
             "Rd": Rd,
             "G": G,

code/models/autoports/qwen_qwen3_5_122b_a10b/tt/optimized_decoder.py

@@ -74,6 +74,13 @@ class OptimizationConfig:
     sorted_gate_up_in0: int = 32
     sorted_down_cores: int = 64
     sorted_down_in0: int = 32
+    # stage 12: group-parallel sparse matmul (MatmulMultiCoreReuseProgramConfig + `indices`): the expert groups are split
+    # across the whole core grid (each core reads its groups' rows and the whole [K, N] expert block) instead of every
+    # core walking every group with only the small per-device N split. in0_block_w per projection (K tiles per step).
+    sorted_grouped: bool = os.environ.get("QWEN35_SORTED_GROUPED", "1") == "1"
+    grouped_gate_up_in0: int = 4
+    grouped_down_in0: int = 4
+    grouped_subblock_w: int = 8
     # --- decode matmul program configs (round 2, from the stage-2 profile) ---
     # router gate matmul [32,H]x[H,256]: "ttnn" (auto), "derived" (TTMoEGate 2D-mcast derivation), "1d8" (1D, 8 cores)
     router_decode_pc: str = "derived"
@@ -397,14 +404,22 @@ class Qwen35MoEOptimized(Qwen35MoEFused):
         x_disp = gather_rows(x4, m["src_tok"], Rd, gs)  # [1,G,gs,H]
         ttnn.deallocate(m["src_tok"])
         narrow = self.tile_gu.tile_shape[1] != 32
+        grouped = bool(getattr(opt, "sorted_grouped", False))
+        if grouped:
+            # indexed/gather mode: the id list drives the weight reads, `sparsity` is unused (kept for the op signature)
+            sparse_kw = {"indices": m["group_ids"]}
+            cfg_gu, cfg_dn = self._grouped_configs(gs, H, Nl)
+        else:
+            sparse_kw = {"nnz": G}
+            cfg_gu, cfg_dn = self._cfg_sorted_gu, self._cfg_sorted_dn
         out_gu = None if narrow else self._alloc_dev((1, G, gs, 2 * Nl), act)
         fused = ttnn.sparse_matmul(
             x_disp,
             self.up_gate,
             sparsity=m["group_mask"],
-            nnz=G,
+            **sparse_kw,
             memory_config=ttnn.DRAM_MEMORY_CONFIG,
-            program_config=self._cfg_sorted_gu,
+            program_config=cfg_gu,
             **({"output_tile": self.tile_gu} if narrow else {"optional_output_tensor": out_gu}),
             compute_kernel_config=self.routed_ckc,
             dtype=act,
@@ -421,22 +436,68 @@ class Qwen35MoEOptimized(Qwen35MoEFused):
             din,
             self.down,
             sparsity=m["group_mask"],
-            nnz=G,
+            **sparse_kw,
             optional_output_tensor=out_dn,
             is_input_a_sparse=False,
             is_input_b_sparse=True,
             memory_config=ttnn.DRAM_MEMORY_CONFIG,
-            program_config=self._cfg_sorted_dn,
+            program_config=cfg_dn,
             compute_kernel_config=self.routed_ckc,
             dtype=act,
         )  # compact [1,G,32,H]
         ttnn.deallocate(din)
         ttnn.deallocate(m["group_mask"])
+        ttnn.deallocate(m["group_ids"])
         y = undispatch_rows(down, None, m["undispatch"], T, k, H)
         ttnn.deallocate(down)
         ttnn.deallocate(m["undispatch"])
         return y
 
+    def _grouped_configs(self, gs, H, Nl):
+        """MatmulMultiCoreReuseProgramConfig pair (gate/up, down) for the group-parallel sparse matmul: whole groups per
+        core (per_core_M = gs/32, per_core_N = all N tiles), K stepped in in0_block_w tiles, subblock 1 x w."""
+        if getattr(self, "_cfg_grouped", None) is None:
+            opt = self.opt
+            grid = self.device.compute_with_storage_grid_size()
+
+            def cfg(k, n_tiles, in0_cap, w_tile_bytes):
+                k_tiles = k // 32
+                ib = max(1, min(int(in0_cap), k_tiles))
+                while k_tiles % ib:
+                    ib -= 1
+                # L1 budget of the grouped factory per core: double-buffered in0 [M, ib] + in1 [ib, N] blocks and one
+                # bf16 output block. Shrink the K block until it fits (a too-large block is a program-build TT_THROW).
+                m_tiles = gs // 32
+                while (
+                    ib > 1
+                    and 2 * m_tiles * ib * 2048 + 2 * n_tiles * ib * w_tile_bytes + m_tiles * n_tiles * 2048 > 1_300_000
+                ):
+                    ib -= 1
+                    while k_tiles % ib:
+                        ib -= 1
+                sw = max(1, min(int(opt.grouped_subblock_w), n_tiles))
+                while n_tiles % sw:
+                    sw -= 1
+                if getattr(self.routed_ckc, "fp32_dest_acc_en", False):
+                    while sw > 4:
+                        sw -= 1
+                return ttnn.MatmulMultiCoreReuseProgramConfig(
+                    compute_with_storage_grid_size=ttnn.CoreCoord(grid.x, grid.y),
+                    in0_block_w=ib,
+                    out_subblock_h=1,
+                    out_subblock_w=sw,
+                    per_core_M=gs // 32,
+                    per_core_N=n_tiles,
+                )
+
+            gu_tiles = -(-(2 * Nl) // self.tile_gu.tile_shape[1])
+            tile_bytes = {ttnn.bfloat4_b: 576, ttnn.bfloat8_b: 1088}
+            self._cfg_grouped = (
+                cfg(H, gu_tiles, opt.grouped_gate_up_in0, tile_bytes.get(self.up_gate.dtype, 2048)),
+                cfg(Nl, H // 32, opt.grouped_down_in0, tile_bytes.get(self.down.dtype, 2048)),
+            )
+        return self._cfg_grouped
+
     def _route(self, x4):
         if self._gate_pc_decode is not None:
             self.gate.matmul_program_config = self._gate_pc_decode if _is_single_tile_rows(x4) else None

image/index.json

@@ -1 +1 @@
-{"schemaVersion":2,"mediaType":"application/vnd.oci.image.index.v1+json","manifests":[{"mediaType":"application/vnd.oci.image.index.v1+json","digest":"sha256:36b6aa8bb6e401b97ed0ef5db2b8ab13ea4829fca22611767a3484281fb8efd3","size":856,"annotations":{"io.containerd.image.name":"docker.io/tt-model/qwen3-5-122b-a10b-blackholex4:36b6aa8bb6e4","org.opencontainers.image.ref.name":"36b6aa8bb6e4"}}]}
+{"schemaVersion":2,"mediaType":"application/vnd.oci.image.index.v1+json","manifests":[{"mediaType":"application/vnd.oci.image.index.v1+json","digest":"sha256:5eba8ab305a32cfd16dd1b0e4a95e16511eb5d65a4bb15480145d32343ab87a7","size":856,"annotations":{"io.containerd.image.name":"docker.io/tt-model/qwen3-5-122b-a10b-blackholex4:5eba8ab305a3","org.opencontainers.image.ref.name":"5eba8ab305a3"}}]}

image/manifest.json

@@ -1 +1 @@
-[{"Config":"blobs/sha256/40268308fd04bae76d29e0b898ecee44f4adf7258ce5ff5682dca2f69453844b","RepoTags":["tt-model/qwen3-5-122b-a10b-blackholex4:36b6aa8bb6e4"],"Layers":["blobs/sha256/20c3783cc497b5b0df1fc5f92bd64c3d2fbb24057c88692c8fda205b6ea8a2f2","blobs/sha256/409bccaa1e5edce776e437ba1fd2dfeae72411c39808ad5b8fe4b2c44d0fdecf","blobs/sha256/ca1b48345588152c337c06522d0e8b9ead94ab85c617a2e0ed86b2cf1eed1a4c","blobs/sha256/6c896dff2fad0954649352d1f1d3a3e349c4e91f9f1333abc162f48447e9def8","blobs/sha256/7a326200e30db3906bfdf9663e5a6a0d15a5239a0fbe5621aed6f010d189eca4","blobs/sha256/4a3c0365cbab8679ce6df3d94017005b357e605bf0fed2af2ba8269fa6f611f6","blobs/sha256/8b4dab4fc6a31529d446fe5f773581efa6eafc05003fcffb35c9f8f28d219495","blobs/sha256/6ed79b6483351c2a4a3d7ccf0a4990d34c369c4bda02694c98865309634459cc","blobs/sha256/5158307e08b9efcf148b0aec8495563eadf530a700f6692d011cfd24924e429d","blobs/sha256/316c35d44617a9e384bc1966e5174fbfab8bdce8101a383c849131e079239071","blobs/sha256/82d50a35e90f4d6f6522e0d37fedfd907c7bcdbcab6a5142d3422ecb513e6892","blobs/sha256/530b11ccc64d093ac59b0feb79f1bca320c5fa2f26d284c4096b9b1fbd93d763","blobs/sha256/35aa8e21d3b8665d01e3e535118ee4c404f34f53fa869645d7dc30ad7d87fd7e","blobs/sha256/e5a76bf01ac570834f7f699d6cf93e9c3b97b517387266b34762fd2213bb6d21","blobs/sha256/78db12aa8510946a74e526e5c2dd0e0eba787445024ee56b408f29bd669929b4","blobs/sha256/84be0373b501ba7dbb7167427ede32fe3238b196488edd1000deeae3e332d3ec","blobs/sha256/82764f7e6fa6951ff87f8e5801559d6b573c1364b0fef4fcc6059f82855c3768","blobs/sha256/4296bafc732ff3ecc1b73a7e513d6517fa335d414224bb23f626c90cfee07dce","blobs/sha256/4f4fb700ef54461cfa02571ae0db9a0dc1e0cdb5577484a6d75e68dc38e8acc1","blobs/sha256/f95a67c2d0990406b69139c87121bd185efba6faff383171494bf48c6e159df0","blobs/sha256/2d1af980d91a9f536b2d1459050125533da8506be50c5f2d0f39e9d9d19a8d80","blobs/sha256/8659a6f4cd6b4b76e8c1c7fb974361ae92689d02b8f3086679de1318aacf838e"]}]
+[{"Config":"blobs/sha256/fa8c0c6f2e72626f40f8729615d5ee939cc3aaaabfaaaaba13f66c5dbc291c7c","RepoTags":["tt-model/qwen3-5-122b-a10b-blackholex4:5eba8ab305a3"],"Layers":["blobs/sha256/20c3783cc497b5b0df1fc5f92bd64c3d2fbb24057c88692c8fda205b6ea8a2f2","blobs/sha256/78815c488d38ee90bad5e99fec6c732ab7da586982ce87ddedccc9ff0a9aa816","blobs/sha256/035bde8561dd66df090fd80af6ebc54431c4c261619fe21403d8f290eeafc870","blobs/sha256/2eddd918ff6c02569610b5da36ed3f178c8f0ee1ddf0603a3a6c569e192e23c9","blobs/sha256/c854f847de74423cd1f1af2b563665f89337c0d2804b41267365e151f9f0461e","blobs/sha256/4820d8f261faee9be0a6acd0450e16b8d61234cf6487b22b6bcf0c62fdfafae8","blobs/sha256/9f04092f7a343f595eb45f3e4a104139020bb8adccdabd44d2fdaf387d839c96","blobs/sha256/aae6bb6c9aedbe9e83a25b770c86c3001011fea226706fcdf9ce82236b3b6a9e","blobs/sha256/14f7cbfb67bf9b4d82b081176395cd0e8c2a571d282ebf92e005f9f4bdc37f95","blobs/sha256/f3fcfe98dabc1c03cab56f823e2234331c1f5f1ede82c632047eb02668e8c037","blobs/sha256/97cd366873414f4c22d9ef218bde71134a3a0b8ed5e365805065292ed8973b35","blobs/sha256/bce398621f84b70dda369f251b3081f845d7e227051233751d3385b50c94c3fd","blobs/sha256/26a79ed154b4569c6305b329add28623bd8f8b2a38e0c477e64ed1faf8257a7b","blobs/sha256/a7cfd2eb195796a51d66401b624da093e82ef15c02192d6e564b1ce7df512562","blobs/sha256/eb8a185622509cc112e65d9d9648570d7f398e503944dd99dc140662ea67944c","blobs/sha256/e363d35c5fff386aaa991575332fc542a5ffc22036ad2d77796ed723f2d17068","blobs/sha256/7a019a88b8e38115a9885e604af3c441d1934e2c13d09133a4161d3ab450239b","blobs/sha256/9b72b5c978c392fcbd09e2962d3e6c099d8eee5d305cdd6e61c9137db3c44cb4","blobs/sha256/4f4fb700ef54461cfa02571ae0db9a0dc1e0cdb5577484a6d75e68dc38e8acc1","blobs/sha256/c80293f68187a74e7f1c75d3775f7481deb58205d67d41814a35ed3ff3ba96f7","blobs/sha256/ecb21a85d0f8992726659678c047f7521a7a4b5b60ba17ab7644e6c2aa04729e","blobs/sha256/83c8dc56544d4dd8b1e3abd53cdac3e01d14657d11f1e131aaf37c3a8309ce07"]}]

requirements.lock

@@ -4,12 +4,12 @@ aiohttp==3.14.3
 aiosignal==1.4.0
 annotated-doc==0.0.5
 annotated-types==0.8.0
-anthropic==1.8.0
+anthropic==1.9.0
 anyio==4.15.1
 apache-tvm-ffi==0.1.10
 astor==0.8.1
 attrs==26.1.0
-blake3==1.0.9
+blake3==1.0.10
 cachetools==7.2.0
 cbor2==6.1.4
 certifi==2026.7.22
@@ -43,9 +43,9 @@ fastapi-cli==0.0.32
 fastapi-cloud-cli==0.26.0
 fastar==0.12.0
 fastsafetensors==0.4.0
-filelock==4.0.4
+filelock==4.0.6
 flashinfer-python==0.6.14
-fonttools==4.66.0
+fonttools==4.66.1
 frozenlist==1.8.0
 fsspec==2026.9.0
 googleapis-common-protos==1.75.4
@@ -60,7 +60,7 @@ httpx==0.28.1
 httpx2==2.13.1
 huggingface_hub==1.33.0
 humming-kernels==0.1.10
-identify==2.6.19
+identify==2.6.20
 idna==3.20
 ijson==3.5.1
 interegular==0.3.3
@@ -87,7 +87,7 @@ mistral_common==1.12.0
 ml_dtypes==0.5.4
 model-hosting-container-standards==0.1.16
 mpmath==1.3.0
-msgspec==0.21.1
+msgspec==0.22.0
 multidict==6.9.1
 networkx==3.7
 ninja==1.13.2
@@ -109,7 +109,7 @@ nvidia-cutlass-dsl-libs-cu13==4.6.0
 nvidia-ml-py==13.615.71
 nvidia-nvvm==13.4.92
 nvtx==0.2.15
-openai==3.19.2
+openai==3.20.0
 openai-harmony==0.0.8
 opencv-python-headless==4.11.0.86
 opentelemetry-api==1.45.0
@@ -128,7 +128,7 @@ packaging==26.3
 pandas==3.0.6
 partial-json-parser==0.2.1.1.post7
 pillow==12.3.0
-platformdirs==4.12.0
+platformdirs==4.12.1
 pre_commit==4.6.2
 prometheus-fastapi-instrumentator==8.1.0
 prometheus_client==0.26.0
@@ -145,7 +145,7 @@ pydantic-settings==2.15.0
 pydantic_core==2.46.5
 pyelftools==0.33
 Pygments==2.21.0
-PyJWT==2.15.0
+PyJWT==2.15.1
 pyluwen==0.9.0
 pynvvideocodec==2.0.4
 pyparsing==3.3.3
@@ -158,7 +158,7 @@ PyYAML==6.0.3
 pyzmq==27.2.0
 quack-kernels==0.6.3
 referencing==0.37.0
-regex==2026.9.10
+regex==2026.9.29
 requests==2.34.2
 rich==15.0.0
 rich-toolkit==0.20.5
@@ -167,14 +167,14 @@ rpds-py==2026.6.3
 safetensors==0.8.0
 seaborn==0.13.2
 sentencepiece==0.2.2
-sentry-sdk==2.70.0
+sentry-sdk==2.71.0
 setproctitle==1.3.7
 setuptools==80.10.2
 setuptools-scm==8.1.0
 shellingham==1.5.4
 six==1.17.0
 sniffio==1.3.1
-sse-starlette==3.4.11
+sse-starlette==3.5.0
 starlette==1.7.0
 supervisor==4.3.0
 sympy==1.14.0
@@ -198,7 +198,7 @@ truststore==0.10.4
 tt-smi==6.6.0
 tt-tools-common==1.6.0
 tt-umd==0.9.11
-ttnn==0.65.2.dev9873+g191f7f672f3
+ttnn==0.65.2.dev9875
 typer==0.27.2
 typing-inspection==0.4.4
 typing_extensions==4.16.0

tt_kernel_manifest.json

@@ -1,12 +1,12 @@
 {
   "schema_version": "5.1",
   "name": "qwen3-5-122b-a10b-blackholex4",
-  "tt_metal_version": "0.65.2.dev9873+g191f7f672f3",
+  "tt_metal_version": "0.65.2.dev9875",
   "arch": "blackhole",
   "device_count": 4,
   "producer": {
     "tt_kernel_version": "0.1.0",
-    "created_at": "2026-09-27T10:16:30.250376+00:00",
+    "created_at": "2026-09-29T16:31:17.784121+00:00",
     "hostname": "qb2-120-p11t05"
   },
   "weights": {
@@ -27,8 +27,8 @@
     "image": {
       "registry": "hf",
       "repository": "qwen3-5-122b-a10b-blackholex4",
-      "tag": "tt-model/qwen3-5-122b-a10b-blackholex4:36b6aa8bb6e4",
-      "digest": "sha256:36b6aa8bb6e401b97ed0ef5db2b8ab13ea4829fca22611767a3484281fb8efd3"
+      "tag": "tt-model/qwen3-5-122b-a10b-blackholex4:5eba8ab305a3",
+      "digest": "sha256:5eba8ab305a32cfd16dd1b0e4a95e16511eb5d65a4bb15480145d32343ab87a7"
     },
     "kind": "vllm-plugin",
     "runtime": {
@@ -120,27 +120,27 @@
       "import models.autoports.qwen_qwen3_5_122b_a10b.tt.decode_buckets as b; assert b.widths(32) == [1, 2, 4, 8, 16, 32]"
     ],
     "built": {
-      "image": "tt-model/qwen3-5-122b-a10b-blackholex4:36b6aa8bb6e4",
+      "image": "tt-model/qwen3-5-122b-a10b-blackholex4:5eba8ab305a3",
       "repo": "raahemnabeel/qwen3-5-122b-a10b-blackholex4",
       "tt_model_version": "0.1.0",
-      "created_at": "2026-09-27T10:15:01+00:00",
+      "created_at": "2026-09-29T16:25:00+00:00",
       "tt_metal": {
-        "sha": "191f7f672f31d73915950f369af4dda6ab5345aa",
-        "describe": "v0.79.0-dev20260909-70-g191f7f672f-dirty",
+        "sha": "6da58d94a83546bb02886bc8e2658919f76f1254",
+        "describe": "v0.79.0-dev20260909-72-g6da58d94a8",
         "dirty": true,
-        "scm_version": "0.65.2.dev9873+g191f7f672f3",
+        "scm_version": "0.65.2.dev9875",
         "mode": "local",
         "remote": "https://github.com/tenstorrent/tt-metal.git",
         "branch": "qwen3_5_122b_a10b",
         "pushed": false
       },
-      "code_sha256": "39e6b36b56f864a9fa8fc3a209759a1984a472385f4c83ab7e5723d0a52f6935",
+      "code_sha256": "3e629a797061595a45db0ee745c31907e553b12f8745ae6ae3f6bd769a1ff774",
       "plugin": {
         "sha": "92403f3b0db48ac86aabf28159ea675c47a420e8",
         "path": "/home/raahem/vllm-tt-plugin",
         "dirty": false
       },
-      "image_digest": "sha256:36b6aa8bb6e401b97ed0ef5db2b8ab13ea4829fca22611767a3484281fb8efd3"
+      "image_digest": "sha256:5eba8ab305a32cfd16dd1b0e4a95e16511eb5d65a4bb15480145d32343ab87a7"
     }
   }
 }