[HIP] [CK] [FlyDSL] [Kernel] Extend MXFP4 GEMM1 replacement to A4W4

aiter/aot/flydsl/fhmoe.py

@@ -92,6 +92,7 @@ def build_stage2_args(
         sorted_weights,
         num_valid_ids,
         token_num,
+        x_rows,
         n_in,
         k_in,
         blocks,
@@ -114,6 +115,7 @@ def build_stage2_args(
             sorted_weights,
             num_valid_ids,
             token_num,
+            x_rows,
             n_in,
             k_in,
             blocks,

aiter/aot/flydsl/moe.py

@@ -181,7 +181,7 @@ def parse_csv(csv_path: str):
                     }
                     if shared_expert_id >= 0:
                         job["shared_expert_id"] = shared_expert_id
-                    # Stage2 needs to know whether stage1 fuses fp4/fp8 quant —
+                    # Stage2 needs to know whether stage1 fuses fp4/fp8 quant --
                     # this changes the shape of a2_scale (sorted scale buffer
                     # vs separate quant call output).
                     if params["stage"] == 2:
@@ -352,7 +352,7 @@ def _make_a2_scale_for_stage2():
         """Stage2 a2_scale construction per fused_moe_2stages.
 
         When upstream stage1 fuses fp4/fp8 quant (``stage1_fuse_quant`` set),
-        stage2 receives stage1's ``out_scale_sorted`` buffer directly — that
+        stage2 receives stage1's ``out_scale_sorted`` buffer directly -- that
         buffer is padded to 256 rows and 8 cols.  Otherwise stage2 quantizes
         its own input and the resulting sorted scale uses 32-row alignment.
         """
@@ -397,7 +397,7 @@ def _make_a2_scale_for_stage2():
         return None
 
     def _make_w_scale(scale_storage_numel: int):
-        # mxfp4 e8m0 scale — viewed as uint8 by _view_safe before kernel launch.
+        # mxfp4 e8m0 scale -- viewed as uint8 by _view_safe before kernel launch.
         return torch.zeros(scale_storage_numel, dtype=torch.uint8, device=dev)
 
     def _make_a_user(a_dtype_user_shape):
@@ -669,7 +669,6 @@ def _make_a_user(a_dtype_user_shape):
             # inter_dim=384), in which case runtime compiles the legal fallback.
             tile_k = resolve_flydsl_stage2_tile_k(inter_dim, tile_k)
 
-            # Stage2 input is (token_num, topk, inter_dim) in a_dtype storage.
             if a_dtype == "fp4":
                 a_shape = (tokens, topk, inter_dim // 2)
             else:
@@ -755,6 +754,7 @@ def _make_a_user(a_dtype_user_shape):
                     sw_arg,
                     num_valid_ids,
                     tokens,
+                    tokens * topk,
                     _n_in,
                     _k_in,
                     m_blocks,
@@ -851,7 +851,7 @@ def _precompile_a16w4_to_cache(
     The port launch ABI (raw fx.Int64 device pointers) differs from the generic MX
     gemm (``_s1_args_fp4``), so it can't reuse ``_precompile_to_cache``'s arg
     builders.  Instead drive the SAME runtime launchers (``flydsl_a16w4_gemm{1,2}``)
-    the fused-MoE op uses, under ``COMPILE_ONLY=1`` — the cache key then matches
+    the fused-MoE op uses, under ``COMPILE_ONLY=1`` -- the cache key then matches
     runtime by construction (``waves_per_eu=None``, ``persist=False``,
     ``w_layout="standard"``, g2 tile downgrade are all applied inside the
     launcher).  The compiled artifact is keyed only on the kernel's constexpr

aiter/aot/flydsl/mxfp4_moe.py

@@ -1,7 +1,8 @@
 # SPDX-License-Identifier: MIT
 # Copyright (C) 2024-2026, Advanced Micro Devices, Inc. All rights reserved.
 
-"""AOT pre-compile for FlyDSL mxmoe and layout-v2 GEMM2 kernels.
+"""AOT pre-compile for the FlyDSL MXMOE a4w4/a8w4 port (GEMM1 / GEMM2) and the
+layout-v2 GEMM2 kernels.
 
 Parses flydsl_mxmoe_* and flydsl_moe2_layout_* rows from the existing model
 configs plus the active FMoE CSV, and warms the FlyDSL disk cache via the same
@@ -25,9 +26,12 @@
 # moe.py defers every ``flydsl_moe2_layout_`` name to this module, so a CSV the
 # glob misses gets no AOT job at all and JITs on the first inference call.
 DEFAULT_CSVS = sorted(
-    set(glob.glob(f"{_MODEL_CONFIG_DIR}/*_fp4_tuned_fmoe.csv"))
-    | set(glob.glob(f"{_MODEL_CONFIG_DIR}/*_a4w4_tuned_fmoe.csv"))
-    | set(glob.glob(f"{_MODEL_CONFIG_DIR}/*_a8w4_tuned_fmoe.csv"))
+    set(
+        glob.glob(f"{_MODEL_CONFIG_DIR}/*_fp4_tuned_fmoe.csv")
+        + glob.glob(f"{_MODEL_CONFIG_DIR}/*_fp8fp4_tuned_fmoe.csv")
+        + glob.glob(f"{_MODEL_CONFIG_DIR}/*_a4w4_tuned_fmoe.csv")
+        + glob.glob(f"{_MODEL_CONFIG_DIR}/*_a8w4_tuned_fmoe.csv")
+    )
 )
 _ACTIVE_FMOE_CSV = AITER_CONFIGS.AITER_CONFIG_FMOE_FILE
 if os.path.exists(_ACTIVE_FMOE_CSV) and _ACTIVE_FMOE_CSV not in DEFAULT_CSVS:
@@ -67,13 +71,27 @@ def _job_key(job: dict) -> tuple:
         return (
             1,
             job["BM"],
+            job["BN"],
+            job["BK"],
             job["use_nt"],
             job["inline_quant"],
+            job["prefetch_hidden"],
             job["D_HIDDEN"],
             job["D_INTER"],
             job["NE"],
             job["topk"],
             job["xcd_swizzle"],
+            job["a_dtype"],
+            job["out_dtype"],
+            job["act"],
+            job["situ_beta"],
+            job["situ_linear_beta"],
+            job["swiglu_limit"],
+            job["enable_bias"],
+            job["interleave"],
+            job["native_scale_layout"],
+            job["num_waves"],
+            job["k_wave"],
         )
     return (
         2,
@@ -89,12 +107,17 @@ def _job_key(job: dict) -> tuple:
 
 
 def parse_csv(csv_path: str):
-    """Parse an fp4 tuned CSV into unique mxmoe-port compile jobs (one per stage)."""
+    """Parse a tuned CSV into unique MXMOE-port compile jobs (one per stage)."""
+    from aiter.ops.flydsl.moe_common import (
+        DEFAULT_SITUV2_BETA,
+        DEFAULT_SITUV2_LINEAR_BETA,
+    )
     from aiter.ops.flydsl.mxfp4_gemm2_kernels import _epilog_of
     from aiter.ops.flydsl.mxfp4_kname import (
         _is_mxfp4_kname,
         _parse_mxfp4_g1_kname,
         _parse_mxfp4_g2_kname,
+        native_scale_layout_for,
         parse_flydsl_v2_gemm2_kernel,
     )
 
@@ -110,13 +133,15 @@ def _add(job):
 
     with open(csv_path, newline="") as f:
         for row in csv.DictReader(f):
+            token = int(row["token"])
             topk = int(row["topk"])
-            # Shape comes from CSV columns; v2 GEMM2 aligns K to its encoded BK.
             model_dim = int(row["model_dim"])
             expert = int(row["expert"])
             inter_dim = int(row["inter_dim"])
             d_inter = ((inter_dim + 255) // 256) * 256
             d_inter_real = inter_dim if inter_dim != d_inter else None
+
+            kn1 = (row.get("kernelName1") or "").strip()
             kn2 = (row.get("kernelName2") or "").strip()
             v2_g2 = parse_flydsl_v2_gemm2_kernel(kn2)
             if v2_g2 is not None:
@@ -127,23 +152,54 @@ def _add(job):
                 v2_d_inter = d_inter
                 v2_d_inter_real = d_inter_real
 
-            kn1 = (row.get("kernelName1") or "").strip()
             if _is_mxfp4_kname(kn1):
                 p1 = _parse_mxfp4_g1_kname(kn1)
-                _add(
-                    {
-                        "stage": 1,
-                        "kernel_name": kn1,
-                        "BM": p1["BM"],
-                        "use_nt": p1["use_nt"],
-                        "inline_quant": p1["inline_quant"],
-                        "D_HIDDEN": model_dim,
-                        "D_INTER": v2_d_inter,
-                        "NE": expert,
-                        "topk": topk,
-                        "xcd_swizzle": p1["xcd_swizzle"],
-                    }
-                )
+                situ_params = [(1.0, 1.0)]
+                if p1["act"] == "situv2":
+                    situ_params.append(
+                        (DEFAULT_SITUV2_BETA, DEFAULT_SITUV2_LINEAR_BETA)
+                    )
+                for situ_beta, situ_linear_beta in situ_params:
+                    _add(
+                        {
+                            "stage": 1,
+                            "kernel_name": kn1,
+                            "BM": p1["BM"],
+                            "BN": p1["BN"],
+                            "BK": p1["BK"],
+                            "use_nt": p1["use_nt"],
+                            "n_tokens": token,
+                            "inline_quant": p1["inline_quant"],
+                            "prefetch_hidden": p1.get("prefetch_hidden", False),
+                            "D_HIDDEN": model_dim,
+                            # Runtime derives D_INTER from the *stored* weight
+                            # width (``w1.shape[1] // 2``), and non-aligned
+                            # shards ship padded to a multiple of 256 with the
+                            # logical width carried separately in
+                            # ``w2.inter_real`` -- the same convention
+                            # ``is_mxfp4_moe_shape_supported`` uses. Keying on
+                            # the raw CSV ``inter_dim`` would miss those rows
+                            # (Kimi-K3: 384 stored as 512). ``d_inter`` is used
+                            # rather than the GEMM2 ``v2_d_inter`` because the
+                            # latter pads to a backend-specific tile_k.
+                            "D_INTER": d_inter,
+                            "NE": expert,
+                            "topk": topk,
+                            "xcd_swizzle": p1["xcd_swizzle"],
+                            "a_dtype": p1["a_dtype"],
+                            "out_dtype": p1["out_dtype"],
+                            "act": p1["act"],
+                            "situ_beta": situ_beta,
+                            "situ_linear_beta": situ_linear_beta,
+                            "swiglu_limit": 7.0,
+                            "enable_bias": p1["enable_bias"],
+                            "interleave": p1["interleave"],
+                            "native_scale_layout": native_scale_layout_for(p1["BM"]),
+                            "num_waves": p1.get("num_waves", 4),
+                            "k_wave": p1.get("k_wave", 1),
+                        }
+                    )
+
             if v2_g2 is not None:
                 bm = v2_g2["tile_m"]
                 inter_dim_pad = v2_d_inter - inter_dim
@@ -185,8 +241,6 @@ def _add(job):
                             "has_pad": inter_dim_pad > 0 or model_dim_pad > 0,
                             "out_dtype": out_dtype,
                             "enable_bias": enable_bias,
-                            # In the compiled kernel tag: must match the runtime
-                            # wrapper or the AOT entry is keyed differently.
                             "g2_spart": v2_g2["spart"],
                             "g2_bf16_lds": v2_g2["bf16_lds"],
                         }
@@ -248,15 +302,29 @@ def _compile_stage1(job):
         inter_sorted_quant=d,
         inter_sorted_shuffled_scale=d,
         hidden_states=d,
-        n_tokens=job["BM"],
+        n_tokens=job["n_tokens"],
         BM=job["BM"],
+        BN=job["BN"],
+        BK=job["BK"],
         use_nt=job["use_nt"],
         inline_quant=job["inline_quant"],
+        prefetch_hidden=job["prefetch_hidden"],
         NE=job["NE"],
         D_HIDDEN=job["D_HIDDEN"],
         D_INTER=job["D_INTER"],
         topk=job["topk"],
         xcd_swizzle=job["xcd_swizzle"],
+        a_dtype=job["a_dtype"],
+        out_dtype=job["out_dtype"],
+        act=job["act"],
+        situ_beta=job["situ_beta"],
+        situ_linear_beta=job["situ_linear_beta"],
+        swiglu_limit=job["swiglu_limit"],
+        bias=d if job["enable_bias"] else None,
+        interleave=job["interleave"],
+        native_scale_layout=job["native_scale_layout"],
+        num_waves=job["num_waves"],
+        k_wave=job["k_wave"],
         stream=0,
     )
 
@@ -371,6 +439,7 @@ def _compile_v2_stage2(job):
             job["BM"],
             job["topk"],
             job["N_OUT"],
+            model_dim_pad=job["model_dim_pad"],
             expert_mask=None,
             topk_ids=None,
             stream=0,
@@ -418,7 +487,8 @@ def main():
         type=str,
         nargs="+",
         default=DEFAULT_CSVS,
-        help="Path(s) to tuned FMoE CSVs; default: existing model configs plus active merged FMoE config",
+        help="Path(s) to tuned FMoE CSVs; default: fp4, fp8fp4, a4w4, and a8w4 "
+        "model configs plus the active merged FMoE config",
     )
     args = parser.parse_args()
 
@@ -436,7 +506,7 @@ def main():
     stage2_jobs = [j for j in all_jobs if j["stage"] == 2]
 
     print("=" * 72)
-    print("FlyDSL mxmoe a4w4 MoE-port AOT Pre-compilation")
+    print("FlyDSL MXMOE a4w4/a8w4 AOT Pre-compilation")
     print("=" * 72)
     for csv_path in csv_paths:
         print(f"  CSV:          {csv_path}")

aiter/configs/model_configs/glm5_fp4_tuned_fmoe.csv

@@ -32,9 +32,9 @@ gfx950,256,8192,6144,1024,257,9,ActivationType.Silu,torch.bfloat16,torch.float4_
 gfx950,256,16384,6144,1024,257,9,ActivationType.Silu,torch.bfloat16,torch.float4_e2m1fn_x2,torch.float4_e2m1fn_x2,QuantType.per_1x32,1,0,64,0,1083.8484,flydsl_moe1_afp4_wfp4_bf16_t64x128x256_w2_bnt0_fp4,0.9%,1206.3064,flydsl_moe2_layout_afp4_wfp4_bf16_t64x256x128_reduce_sbm64,1.2%,2290.1548,0,0,0,2430.52,2249.94,
 gfx950,256,32768,6144,1024,257,9,ActivationType.Silu,torch.bfloat16,torch.float4_e2m1fn_x2,torch.float4_e2m1fn_x2,QuantType.per_1x32,1,0,128,0,1996.7999,flydsl_moe1_afp4_wfp4_bf16_t128x128x256_w4_bnt0_fp4,0.9%,2417.8726,flydsl_moe2_layout_afp4_wfp4_bf16_t64x256x128_reduce_sbm128,1.2%,4414.6725,0,0,0,2521.72,1235.58,
 gfx950,256,1,6144,512,257,9,ActivationType.Silu,torch.bfloat16,torch.float4_e2m1fn_x2,torch.float4_e2m1fn_x2,QuantType.per_1x32,1,0,32,0,11.5491,flydsl_moe1_afp4_wfp4_bf16_t32x32x256_w4_kw4_fp4,1.0%,7.3233,flydsl_moe2_layout_afp4_wfp4_bf16_t32x128x128_atomic_sbm32,1.3%,18.8724,0,0,0,9.0,128514.38,
-gfx950,256,2,6144,512,257,9,ActivationType.Silu,torch.bfloat16,torch.float4_e2m1fn_x2,torch.float4_e2m1fn_x2,QuantType.per_1x32,1,0,16,0,28.7583,flydsl_mxmoe_g1_a4w4_16x256x256_f16in_nt,0.012391,0.0,flydsl_moe2_layout_afp4_wfp4_bf16_t16x128x256_atomic_sbm16,0.012391,28.7583,0,0,0,0.0,0.0,
-gfx950,256,4,6144,512,257,9,ActivationType.Silu,torch.bfloat16,torch.float4_e2m1fn_x2,torch.float4_e2m1fn_x2,QuantType.per_1x32,1,0,16,0,35.5442,flydsl_mxmoe_g1_a4w4_16x256x256_f16in_nt,0.012156,0.0,flydsl_moe2_layout_afp4_wfp4_bf16_t16x128x256_atomic_nt_sbm16,0.012156,35.5442,0,0,0,0.0,0.0,
-gfx950,256,8,6144,512,257,9,ActivationType.Silu,torch.bfloat16,torch.float4_e2m1fn_x2,torch.float4_e2m1fn_x2,QuantType.per_1x32,1,0,16,0,68.1308,flydsl_mxmoe_g1_a4w4_16x256x256_f16in_nt,0.012034,0.0,flydsl_moe2_layout_afp4_wfp4_bf16_t16x128x256_atomic_sbm16,0.012034,68.1308,0,0,0,0.0,0.0,
+gfx950,256,2,6144,512,257,9,ActivationType.Silu,torch.bfloat16,torch.float4_e2m1fn_x2,torch.float4_e2m1fn_x2,QuantType.per_1x32,1,0,16,0,16.3995,flydsl_mxmoe_g1_a4w4_16x128x256_f16in_nt,1.2%,5.622,flydsl_mxmoe_g2_a4w4_16x256x256_atomic,0.0%,22.0215,0,0,0,15.43,110137.51,
+gfx950,256,4,6144,512,257,9,ActivationType.Silu,torch.bfloat16,torch.float4_e2m1fn_x2,torch.float4_e2m1fn_x2,QuantType.per_1x32,1,0,16,0,23.1407,flydsl_mxmoe_g1_a4w4_16x256x256_f16in_nt,1.2%,10.9459,flydsl_mxmoe_g2_a4w4_16x256x256_atomic,0.0%,34.0866,0,0,0,19.93,71154.94,
+gfx950,256,8,6144,512,257,9,ActivationType.Silu,torch.bfloat16,torch.float4_e2m1fn_x2,torch.float4_e2m1fn_x2,QuantType.per_1x32,1,0,16,0,42.5523,flydsl_mxmoe_g1_a4w4_16x128x256_f16in_nt_xcd2,1.2%,17.0413,flydsl_mxmoe_g2_a4w4_16x256x256_atomic,0.0%,59.5936,0,0,0,22.8,40700.74,
 gfx950,256,16,6144,512,257,9,ActivationType.Silu,torch.bfloat16,torch.float4_e2m1fn_x2,torch.float4_e2m1fn_x2,QuantType.per_1x32,1,0,16,0,91.5029,flydsl_mxmoe_g1_a4w4_16x256x256_f16in_nt,0.011945,0.0,flydsl_moe2_layout_afp4_wfp4_bf16_t16x128x256_atomic_sbm16,0.011945,91.5029,0,0,0,0.0,0.0,
 gfx950,256,32,6144,512,257,9,ActivationType.Silu,torch.bfloat16,torch.float4_e2m1fn_x2,torch.float4_e2m1fn_x2,QuantType.per_1x32,1,0,16,0,135.3655,flydsl_mxmoe_g1_a4w4_16x256x256_f16in_nt,0.011955,0.0,flydsl_moe2_layout_afp4_wfp4_bf16_t16x128x256_atomic_nt_sbm16,0.011955,135.3655,0,0,0,0.0,0.0,
 gfx950,256,64,6144,512,257,9,ActivationType.Silu,torch.bfloat16,torch.float4_e2m1fn_x2,torch.float4_e2m1fn_x2,QuantType.per_1x32,1,0,16,0,181.4689,flydsl_mxmoe_g1_a4w4_16x256x256_f16in_nt,0.011692,0.0,flydsl_moe2_layout_afp4_wfp4_bf16_t16x128x128_atomic_nt_sbm16,0.011692,181.4689,0,0,0,0.0,0.0,
@@ -45,8 +45,8 @@ gfx950,256,1024,6144,512,257,9,ActivationType.Silu,torch.bfloat16,torch.float4_e
 gfx950,256,2048,6144,512,257,9,ActivationType.Silu,torch.bfloat16,torch.float4_e2m1fn_x2,torch.float4_e2m1fn_x2,QuantType.per_1x32,1,0,128,0,181.274,flydsl_moe1_afp4_wfp4_bf16_t128x128x256_w2_fp4,0.9%,185.3303,flydsl_moe2_layout_afp4_wfp4_bf16_t64x256x128_reduce_persist_sbm128,1.2%,366.6043,0,0,0,948.96,6718.7,
 gfx950,256,4096,6144,512,257,9,ActivationType.Silu,torch.bfloat16,torch.float4_e2m1fn_x2,torch.float4_e2m1fn_x2,QuantType.per_1x32,1,0,64,0,251.6512,flydsl_moe1_afp4_wfp4_bf16_t64x128x256_w3_bnt0_xcd4_fp4,0.9%,289.1168,flydsl_moe2_layout_afp4_wfp4_bf16_t64x256x128_reduce_persist_sbm64,1.2%,540.768,0,0,0,1286.66,4624.63,
 gfx950,256,8192,6144,512,257,9,ActivationType.Silu,torch.bfloat16,torch.float4_e2m1fn_x2,torch.float4_e2m1fn_x2,QuantType.per_1x32,1,0,64,0,361.1843,flydsl_moe1_afp4_wfp4_bf16_t64x128x256_w4_bnt0_fp4,0.9%,519.3092,flydsl_moe2_layout_afp4_wfp4_bf16_t64x256x256_reduce_sbm64,1.2%,880.4935,0,0,0,1580.44,2926.03,
-gfx950,256,16384,6144,512,257,9,ActivationType.Silu,torch.bfloat16,torch.float4_e2m1fn_x2,torch.float4_e2m1fn_x2,QuantType.per_1x32,1,0,128,0,1625.8568,flydsl_mxmoe_g1_a4w4_128x256x256,0.011729,0.0,flydsl_mxmoe_g2_a4w4_128x256x256_f4out,0.011729,1625.8568,0,0,0,0.0,0.0,
-gfx950,256,32768,6144,512,257,9,ActivationType.Silu,torch.bfloat16,torch.float4_e2m1fn_x2,torch.float4_e2m1fn_x2,QuantType.per_1x32,1,0,128,0,2997.8444,flydsl_mxmoe_g1_a4w4_128x256x256,0.011733,0.0,flydsl_mxmoe_g2_a4w4_128x256x256_f4out,0.011733,2997.8444,0,0,0,0.0,0.0,
+gfx950,256,16384,6144,512,257,9,ActivationType.Silu,torch.bfloat16,torch.float4_e2m1fn_x2,torch.float4_e2m1fn_x2,QuantType.per_1x32,1,0,128,0,504.2785,flydsl_mxmoe_g1_a4w4_128x256x256_xcd2,1.2%,573.7599,flydsl_mxmoe_g2_a4w4_128x256x256_f4out,0.0%,1078.0384,0,0,0,2581.67,2529.92,
+gfx950,256,32768,6144,512,257,9,ActivationType.Silu,torch.bfloat16,torch.float4_e2m1fn_x2,torch.float4_e2m1fn_x2,QuantType.per_1x32,1,0,128,0,938.3141,flydsl_mxmoe_g1_a4w4_128x256x256_xcd2,1.2%,1117.0695,flydsl_mxmoe_g2_a4w4_128x256x256,0.0%,2055.3836,0,0,0,2708.15,1473.85,
 gfx950,256,1,6144,256,257,9,ActivationType.Silu,torch.bfloat16,torch.float4_e2m1fn_x2,torch.float4_e2m1fn_x2,QuantType.per_1x32,1,0,32,0,10.4289,flydsl_moe1_afp4_wfp4_bf16_t32x32x256_kw4_fp4,0.9%,6.2885,flydsl_moe2_layout_afp4_wfp4_bf16_t32x128x256_atomic_nt_sbm32,1.3%,16.7174,0,0,0,5.08,72540.98,
 gfx950,256,2,6144,256,257,9,ActivationType.Silu,torch.bfloat16,torch.float4_e2m1fn_x2,torch.float4_e2m1fn_x2,QuantType.per_1x32,1,0,32,0,11.2231,flydsl_moe1_afp4_wfp4_bf16_t32x32x256_w2_bnt0_kw4_fp4,0.9%,7.6731,flydsl_moe2_layout_afp4_wfp4_bf16_t32x256x128_atomic_nt_sbm32,1.2%,18.8962,0,0,0,8.99,64177.72,
 gfx950,256,4,6144,256,257,9,ActivationType.Silu,torch.bfloat16,torch.float4_e2m1fn_x2,torch.float4_e2m1fn_x2,QuantType.per_1x32,1,0,16,0,29.5482,flydsl_mxmoe_g1_a4w4_16x256x256_f16in_nt,0.012428,0.0,flydsl_moe2_layout_afp4_wfp4_bf16_t16x256x256_atomic_sbm16,0.012428,29.5482,0,0,0,0.0,0.0,