files

2026-09-03 06:33:50 +08:00 · 2026-09-02 11:17:46 -05:00 · 2026-09-02 09:59:09 -05:00
369 changed files with 14725 additions and 7582 deletions
@@ -159,7 +159,21 @@ All four files must be updated together.

 ## Manifest

-Generate the base manifest after compilation, then namespace the release artifacts as v23:
+The current test branch uses manifest v25 and requests v25 only. Seed the new
+manifest from the previous catalog, then replace entries as artifacts are
+rebuilt with the pinned runtime:
+
+```bash
+cp /path/to/model_names_v24.json /path/to/model_names_v25.json
+```
+
+The current tinygrad pin is `f6fc4e3f2c3db5fae1e19cbfbc3ad9fc579a12ae`, from
+`openpilot` `origin/master` (`bump tg + TC_MIN_GLOBALS`). StarPilot's
+multi-model `modeld` remains in place; do not replace it with upstream's
+single-model `modeld`.
+
+For the older namespace migration workflow, generate the base manifest after
+compilation and namespace the release artifacts as v23:

 ```bash
 python3 scripts/model_rebuild_pipeline.py manifest \
@@ -175,9 +189,9 @@ python3 scripts/namespace_model_artifacts.py \

 The namespace command changes IDs such as `tr1422` to `tr14223`, renames the
 compiled and upload-ready files, and writes an ID map. It preserves display
-names and behavioral versions. The current model manager requests v23 only;
-the manifest is fetched from `Models/model_names_v23.json`, while v22 remains
-available for devices that have not updated yet.
+names and behavioral versions. The current model manager requests v25 only; the
+manifest is fetched from `Models/model_names_v25.json`. Devices still running
+the prior branch continue to request their existing manifest version.

 After importing newly compiled sources, normalize the release namespace before
 copying files into either resource repository:
@@ -205,4 +219,4 @@ Compilation validates JIT capture/replay, pickle round-trip, finite outputs, met
 4. Confirm `driverStateV2` on both supported camera resolutions.
 5. Test download, selection, deletion, randomization, migration, and fallback in both device UIs and Galaxy.

-The built-in RDF artifact is `selfdrive/modeld/models/driving_tinygrad.pkl`. If migration cannot download the selected v23 artifact, StarPilot switches to that built-in model.
+The built-in RDF artifact is `selfdrive/modeld/models/driving_tinygrad.pkl`. If migration cannot download the selected v25 artifact, StarPilot switches to that built-in model.
@@ -21,7 +21,7 @@ fi
 export QCOM_PRIORITY=12

 if [ -z "$AGNOS_VERSION" ]; then
-  export AGNOS_VERSION="19.6.17"
+  export AGNOS_VERSION="19.6.16"
 fi

 if [ -z "$AGNOS_ACCEPTED_VERSIONS" ]; then
@@ -8,9 +8,9 @@ from opendbc.car.lateral import apply_driver_steer_torque_limits, apply_steer_an
 from opendbc.car.common.conversions import Conversions as CV
 from opendbc.car.hyundai import hyundaicanfd, hyundaican
 from opendbc.car.hyundai.hyundaicanfd import CanBus
-from opendbc.car.hyundai.values import HyundaiFlags, HyundaiStarPilotFlags, Buttons, CarControllerParams, CAR, CANFD_ANGLE_LONGITUDINAL_CAR, \
+from opendbc.car.hyundai.values import HyundaiFlags, Buttons, CarControllerParams, CAR, CANFD_ANGLE_LONGITUDINAL_CAR, \
                                        CANFD_RADAR_LIVE_LONGITUDINAL_CAR, CANFD_ALT_BUTTONS_RESUME_CAR, kia_ev6_gt_line_longitudinal_tuning, \
-                                        KIA_EV6_GT_LINE_LONG_TUNING_TESTING_GROUND_ID, CANFD_DAW_SUPPRESSION_CAR
+                                        KIA_EV6_GT_LINE_LONG_TUNING_TESTING_GROUND_ID
 from opendbc.car.interfaces import CarControllerBase
 from opendbc.car.vehicle_model import VehicleModel
 from openpilot.common.params import Params
@@ -777,8 +777,6 @@ class CarController(CarControllerBase):
                                                  left_lane_warning, right_lane_warning, lka_icon))
      if self.CP.carFingerprint == CAR.KIA_RAY_EV:
        self._ray_lkas11_active = True
-      if getattr(self.FPCP, "flags", 0) & HyundaiStarPilotFlags.HAS_LKAS12:
-        can_sends.append(hyundaican.create_lkas12(self.packer, CS.lkas12))

    # Button messages
    if not self.long_active_ecu:
@@ -852,6 +850,9 @@ class CarController(CarControllerBase):
    )

    # steering control
+    # The first-generation Electrified GV70 expects the synthesized LKAS status
+    # payload. Forwarding its stock status bits leaves lane-safety state asserted
+    # while StarPilot is suppressing the stock LFA path.
    preserve_stock_lkas = bool(self.CP.flags & HyundaiFlags.CANFD_LKA_STEERING) and \
      not self.long_active_ecu and self.CP.carFingerprint != CAR.GENESIS_GV70_ELECTRIFIED_1ST_GEN and \
      preserve_stock_canfd_lkas_status(self.CP.carFingerprint)
@@ -894,9 +895,6 @@ class CarController(CarControllerBase):
      can_sends.extend(hyundaicanfd.create_inactive_angle_steering_messages(self.packer, self.CAN,
                                                                             inactive_steering_angle))

-    if self.CP.carFingerprint in CANFD_DAW_SUPPRESSION_CAR and getattr(CS, "stock_daw_msg", None):
-      can_sends.append(hyundaicanfd.create_suppress_daw(self.packer, self.CAN, CS.stock_daw_msg))
-
    # prevent LFA from activating on LKA steering cars by sending "no lane lines detected" to ADAS ECU
    suppress_lfa = bool(lka_steering)
    if angle_lkas_alt:
@@ -10,7 +10,6 @@ from opendbc.car.hyundai.hyundaicanfd import CanBus
 from opendbc.car.hyundai.values import HyundaiFlags, HyundaiStarPilotFlags, HyundaiStarPilotSafetyFlags, CAR, DBC, Buttons, CarControllerParams, \
                                       CANFD_ANGLE_LONGITUDINAL_CAR, CANFD_CORNER_RADAR_BSM_CAR, \
                                       CANFD_ALT_BUTTONS_RESUME_CAR, \
-                                       CANFD_DAW_SUPPRESSION_CAR, \
                                       hyundai_cancel_button_enables_cruise, ALT_BUS_LDA_BUTTON_CARS, ALT_BUS_LDA_BUTTON_SWL_STAT_CARS
 from opendbc.car.interfaces import CarStateBase

@@ -143,8 +142,6 @@ class CarState(CarStateBase):
    self.msg_364 = {}
    self.lfa_block_msg = {}
    self.stock_lkas_msg = {}
-    self.lkas12 = {}
-    self.stock_daw_msg = {}
    self.stock_lfa_msg = {}
    self.stock_lfahda_cluster_msg = {}
    self.stock_camera_lead_visible = False
@@ -443,8 +440,6 @@ class CarState(CarStateBase):
      self.lkas11 = {}
    else:
      self.lkas11 = copy.copy(cp_cam.vl["LKAS11"])
-    if getattr(self.FPCP, "flags", 0) & HyundaiStarPilotFlags.HAS_LKAS12:
-      self.lkas12 = copy.copy(cp_cam.vl["LKAS12"])
    self.clu11 = copy.copy(cp.vl["CLU11"])
    self.steer_state = cp.vl["MDPS12"]["CF_Mdps_ToiActive"]  # 0 NOT ACTIVE, 1 ACTIVE
    if not self.main_cruise_tracking:
@@ -615,8 +610,6 @@ class CarState(CarStateBase):
    if self.CP.carFingerprint in CANFD_ANGLE_LONGITUDINAL_CAR and cp.ts_nanos["FR_CMR_01_10ms"]["FR_CMR_Crc1Val"] > 0:
      hba_icon = int(cp.vl["FR_CMR_01_10ms"]["HBA_IndLmpReq"])
      self.hba_icon = hba_icon if hba_icon in (1, 2) else 0
-    if self.CP.carFingerprint in CANFD_DAW_SUPPRESSION_CAR and cp.ts_nanos["FR_CMR_01_10ms"]["FR_CMR_Crc1Val"] > 0:
-      self.stock_daw_msg = copy.copy(cp.vl["FR_CMR_01_10ms"])
    if cp.ts_nanos["BLINKER_STALKS"]["CHECKSUM_MAYBE"] > 0:
      self.stock_blinker_stalks_ts = cp.ts_nanos["BLINKER_STALKS"]["CHECKSUM_MAYBE"]

@@ -680,7 +673,7 @@ class CarState(CarStateBase):
    ]
    if CP.enableBsm:
      msgs.append(("BLINDSPOTS_REAR_CORNERS", 0))
-    if CP.carFingerprint in CANFD_ANGLE_LONGITUDINAL_CAR | CANFD_DAW_SUPPRESSION_CAR:
+    if CP.carFingerprint in CANFD_ANGLE_LONGITUDINAL_CAR:
      msgs.append(("BLINDSPOTS_FRONT_CORNER_2", 0))
      msgs.append(("FR_CMR_01_10ms", 0))
    if CP.flags & HyundaiFlags.EV:
@@ -737,7 +730,7 @@ class CarState(CarStateBase):

    parsers = {
      Bus.pt: CANParser(DBC[CP.carFingerprint][Bus.pt], msgs, 0),
-      Bus.cam: CANParser(DBC[CP.carFingerprint][Bus.pt], [("LKAS12", 0)], 2),
+      Bus.cam: CANParser(DBC[CP.carFingerprint][Bus.pt], [], 2),
    }
    if CP.carFingerprint in ALT_BUS_LDA_BUTTON_CARS:
      parsers[Bus.alt] = CANParser(DBC[CP.carFingerprint][Bus.pt], [("CLU13", 0)], 1)
@@ -81,10 +81,6 @@ def create_lkas11(packer, frame, CP, apply_torque, steer_req,
    values["CF_Lkas_LdwsActivemode"] = 2

  if CP.carFingerprint == CAR.KIA_RAY_EV:
-    if not enabled:
-      values["CF_Lkas_LdwsActivemode"] = lkas11["CF_Lkas_LdwsActivemode"]
-      values["CF_Lkas_LdwsSysState"] = lkas11["CF_Lkas_LdwsSysState"]
-      values["CF_Lkas_FcwOpt_USM"] = lkas11["CF_Lkas_FcwOpt_USM"]
    values["CF_Lkas_LdwsOpt_USM"] = 0
    values["CF_Lkas_Chksum"] = 0

@@ -106,19 +102,6 @@ def create_lkas11(packer, frame, CP, apply_torque, steer_req,
  return packer.make_can_msg("LKAS11", 0, values)


-def create_lkas12(packer, lkas12):
-  values = {s: lkas12[s] for s in (
-    "CF_Lkas_TsrSlifOpt",
-    "CF_LkasTsrStatus",
-    "CF_Lkas_TsrSpeed_Display_Clu",
-    "CF_LkasTsrSpeed_Display_Navi",
-    "CF_Lkas_TsrAddinfo_Display",
-    "CF_Lkas_Daw_USM",
-  ) if s in lkas12}
-  values["CF_LkasDawStatus"] = 0
-  return packer.make_can_msg("LKAS12", 0, values)
-
-
 def create_checksum_can_canfd_blended(packer, bus, addr, values):
  dat = packer.make_can_msg(addr, bus, values)[1]
  return hyundai_checksum(dat[1:8])
@@ -63,6 +63,61 @@ def _update_checksum(packer, address: int, dat: bytearray) -> None:
  _set_value(dat, sig_checksum, checksum)


+def _set_little_endian_bits(dat: bytearray, lsb: int, size: int, value: int) -> None:
+  """Write the legacy HDA-II field layout without changing the generated DBC aliases."""
+  value &= (1 << size) - 1
+  bit = lsb
+  remaining = size
+  while remaining:
+    byte = bit // 8
+    shift = bit % 8
+    chunk_size = min(remaining, 8 - shift)
+    mask = ((1 << chunk_size) - 1) << shift
+    dat[byte] = (dat[byte] & ~mask) | ((value & ((1 << chunk_size) - 1)) << shift)
+    value >>= chunk_size
+    bit += chunk_size
+    remaining -= chunk_size
+
+
+def _create_gv70_lka_status_msg(packer, CAN, message_name: str, bus: int, enabled: bool,
+                                lat_active: bool, apply_torque: int):
+  values = {
+    "LKA_MODE": 2,
+    "LKA_ICON": 2 if enabled else 1,
+    "TORQUE_REQUEST": apply_torque,
+    "STEER_REQ": 1 if lat_active else 0,
+    "LKA_ASSIST": 0,
+    "STEER_MODE": 0,
+    "DAMP_FACTOR": 100,
+  }
+  address, raw, _ = packer.make_can_msg(message_name, bus, values)
+  dat = bytearray(raw)
+
+  legacy_fields = (
+    (24, 3, 2),
+    (27, 3, 0),
+    (30, 2, 0),
+    (32, 2, 0),
+    (34, 2, 0),
+    (36, 2, 0),
+    (38, 3, 2 if enabled else 1),
+    (52, 2, 1 if lat_active else 0),
+    (54, 2, 0),
+    (56, 1, 0),
+    (60, 4, 0),
+    (80, 2, 0),
+  )
+  for lsb, size, value in legacy_fields:
+    _set_little_endian_bits(dat, lsb, size, value)
+
+  _set_little_endian_bits(dat, 64 if message_name == "LKAS" else 104, 8, 100)
+  if message_name == "LKAS":
+    _set_little_endian_bits(dat, 84, 3, 0)
+
+  _update_checksum(packer, address, dat)
+  return address, bytes(dat), bus
+
+
 def _create_angle_lfa_msg(packer, CAN, values, apply_angle: float, lat_active: bool, torque_reduction_gain: float):
  address = packer.dbc.name_to_msg["LFA"].address
  dat = packer.pack(address, values)
@@ -101,6 +156,13 @@ def create_steering_messages(packer, CP, CAN, enabled, lat_active, apply_torque,
  if lka_icon is None:
    lka_icon = 2 if enabled else 1

+  if CP.carFingerprint == CAR.GENESIS_GV70_ELECTRIFIED_1ST_GEN and CP.flags & HyundaiFlags.CANFD_LKA_STEERING:
+    ret = []
+    if CP.openpilotLongitudinalControl:
+      ret.append(_create_gv70_lka_status_msg(packer, CAN, "LFA", CAN.ECAN, enabled, lat_active, apply_torque))
+    ret.append(_create_gv70_lka_status_msg(packer, CAN, "LKAS", CAN.ACAN, enabled, lat_active, apply_torque))
+    return ret
+
  angle_lkas_alt = CP.flags & HyundaiFlags.CANFD_ANGLE_STEERING and CP.flags & HyundaiFlags.CANFD_LKA_STEERING_ALT

  control_values = {
@@ -243,12 +305,6 @@ def create_suppress_lfa(packer, CAN, lfa_block_msg, lka_steering_alt):
  return packer.make_can_msg(suppress_msg, CAN.ACAN, values)


-def create_suppress_daw(packer, CAN, daw_msg):
-  values = copy.copy(daw_msg)
-  values["DAW_WrnMsgSta"] = 0
-  return packer.make_can_msg("FR_CMR_01_10ms", CAN.ECAN, values)
-
-
 def create_buttons(packer, CP, CAN, cnt, btn=0, base_values=None, left_paddle=False, right_paddle=False):
  values = {k: v for k, v in base_values.items() if k not in ("CHECKSUM", "_CHECKSUM", "COUNTER")} if base_values else {}
  values.update({
@@ -570,41 +570,6 @@ class TestHyundaiFingerprint:
    assert not (CP.flags & HyundaiFlags.CANFD_LKA_STEERING)
    assert bool(CP.flags & HyundaiFlags.CANFD_CAMERA_SCC)

-  def test_ioniq_6_filters_stock_driver_attention_warning(self):
-    CP = CarParams.new_message()
-    CP.carFingerprint = CAR.HYUNDAI_IONIQ_6
-    CP.flags = int(HyundaiFlags.CANFD | HyundaiFlags.EV | HyundaiFlags.CANFD_LKA_STEERING |
-                   HyundaiFlags.CANFD_LKA_STEERING_ALT)
-    CP.openpilotLongitudinalControl = False
-
-    controller = CarController(DBC[CP.carFingerprint], CP)
-    controller.frame = 1
-    can_bus = CanBus(CP)
-    parser = CANParser(DBC[CP.carFingerprint][Bus.pt], [("FR_CMR_01_10ms", 0)], can_bus.ECAN)
-    stock_daw_msg = {
-      "FR_CMR_AlvCnt1Val": 7,
-      "DAW_SysSta": 6,
-      "DAW_WrnMsgSta": 1,
-      "HBA_IndLmpReq": 2,
-    }
-    cc = SimpleNamespace(enabled=True, latActive=True,
-                         actuators=SimpleNamespace(longControlState=LongCtrlState.off),
-                         leftBlinker=False, rightBlinker=False, hudControl=SimpleNamespace())
-    cs = SimpleNamespace(stock_lfa_msg=None, stock_lkas_msg={}, lfa_block_msg={}, stock_daw_msg=stock_daw_msg,
-                         out=SimpleNamespace(steeringAngleDeg=0.0, gearShifter=structs.CarState.GearShifter.drive))
-
-    msgs = controller.create_canfd_msgs(0, True, 0.0, 0.0, 0.0, 0.0, False, cc.hudControl, cs, cc,
-                                        get_test_toggles(), lka_icon=2, lfa_icon=2)
-    daw_msgs = [msg for msg in msgs if msg[0] == 0x11A]
-    assert len(daw_msgs) == 1
-
-    parser.update([(1, daw_msgs)])
-
-    assert parser.can_valid
-    assert parser.vl["FR_CMR_01_10ms"]["DAW_SysSta"] == 6
-    assert parser.vl["FR_CMR_01_10ms"]["DAW_WrnMsgSta"] == 0
-    assert parser.vl["FR_CMR_01_10ms"]["HBA_IndLmpReq"] == 2
-
  def test_ioniq_6_clears_torque_with_inactive_safety_request(self):
    ioniq_6_cp = SimpleNamespace(carFingerprint=CAR.HYUNDAI_IONIQ_6)
    other_cp = SimpleNamespace(carFingerprint=CAR.KIA_EV6)
@@ -745,7 +710,7 @@ class TestHyundaiFingerprint:
    assert parser.vl["LKAS11"]["CF_Lkas_LdwsActivemode"] == 0
    assert parser.vl["LKAS11"]["CF_Lkas_FcwOpt_USM"] == 0

-  def test_kia_ray_ev_preserves_stock_inactive_lkas_status(self):
+  def test_kia_ray_ev_preserves_stock_lkas_option(self):
    fingerprint = gen_empty_fingerprint()
    fingerprint[2][0x485] = 4
    CP = CarInterface.get_params(CAR.KIA_RAY_EV, fingerprint, [], False, False, False, None)
@@ -753,46 +718,15 @@ class TestHyundaiFingerprint:
    packer = CANPacker(DBC[CP.carFingerprint][Bus.pt])
    parser = CANParser(DBC[CP.carFingerprint][Bus.pt], [("LKAS11", 0)], 0)

-    lkas11 = parser.vl["LKAS11"]
-    lkas11.update({
-      "CF_Lkas_LdwsActivemode": 0,
-      "CF_Lkas_LdwsSysState": 1,
-      "CF_Lkas_FcwOpt_USM": 1,
-    })
    msg = hyundaican.create_lkas11(
-      packer, 0, CP, 0, True, False, lkas11, False, 4, False,
-      True, True, 0, 0, 2,
-    )
-    parser.update([(1, [msg])])
-
-    assert parser.vl["LKAS11"]["CF_Lkas_LdwsActivemode"] == 0
-    assert parser.vl["LKAS11"]["CF_Lkas_LdwsSysState"] == 1
-    assert parser.vl["LKAS11"]["CF_Lkas_LdwsOpt_USM"] == 0
-    assert parser.vl["LKAS11"]["CF_Lkas_FcwOpt_USM"] == 1
-
-  def test_kia_ray_ev_uses_active_lkas_status_when_enabled(self):
-    fingerprint = gen_empty_fingerprint()
-    fingerprint[2][0x485] = 4
-    CP = CarInterface.get_params(CAR.KIA_RAY_EV, fingerprint, [], False, False, False, None)
-    packer = CANPacker(DBC[CP.carFingerprint][Bus.pt])
-    parser = CANParser(DBC[CP.carFingerprint][Bus.pt], [("LKAS11", 0)], 0)
-
-    lkas11 = parser.vl["LKAS11"]
-    lkas11.update({
-      "CF_Lkas_LdwsActivemode": 0,
-      "CF_Lkas_LdwsSysState": 1,
-      "CF_Lkas_FcwOpt_USM": 1,
-    })
-    msg = hyundaican.create_lkas11(
-      packer, 0, CP, 0, True, False, lkas11, False, 4, True,
+      packer, 0, CP, 0, True, False, parser.vl["LKAS11"], False, 4, False,
      True, True, 0, 0, 2,
    )
    parser.update([(1, [msg])])

    assert parser.vl["LKAS11"]["CF_Lkas_LdwsActivemode"] == 3
-    assert parser.vl["LKAS11"]["CF_Lkas_LdwsSysState"] == 4
    assert parser.vl["LKAS11"]["CF_Lkas_LdwsOpt_USM"] == 0
-    assert parser.vl["LKAS11"]["CF_Lkas_FcwOpt_USM"] == 2
+    assert parser.vl["LKAS11"]["CF_Lkas_FcwOpt_USM"] == 1

  def test_kia_ray_ev_delays_first_lkas11(self):
    fingerprint = gen_empty_fingerprint()
@@ -862,37 +796,6 @@ class TestHyundaiFingerprint:
    palisade_2023 = CarInterface.get_params(CAR.HYUNDAI_PALISADE_2023, gen_empty_fingerprint(), [], True, False, False, None)
    assert palisade_2023.safetyConfigs[-1].safetyParam & HyundaiStarPilotSafetyFlags.HAS_LDA_BUTTON

-  def test_lkas12_da_warning_is_filtered_for_camera_fingerprint(self):
-    fingerprint = gen_empty_fingerprint()
-    fingerprint[2][0x53E] = 6
-    CP = CarInterface.get_params(CAR.HYUNDAI_SONATA_HYBRID, fingerprint, [], False, False, False, None)
-    FPCP = CarInterface.get_starpilot_params(CAR.HYUNDAI_SONATA_HYBRID, fingerprint, [], CP, get_test_toggles())
-    assert FPCP.flags & HyundaiStarPilotFlags.HAS_LKAS12
-
-    packer = CANPacker(DBC[CP.carFingerprint][Bus.pt])
-    stock = {
-      "CF_Lkas_TsrSlifOpt": 3,
-      "CF_LkasTsrStatus": 2,
-      "CF_Lkas_TsrSpeed_Display_Clu": 80,
-      "CF_LkasTsrSpeed_Display_Navi": 70,
-      "CF_Lkas_TsrAddinfo_Display": 1,
-      "CF_Lkas_Daw_USM": 0,
-      "CF_LkasDawStatus": 1,
-    }
-    msg = hyundaican.create_lkas12(packer, stock)
-    parser = CANParser(DBC[CP.carFingerprint][Bus.pt], [("LKAS12", 0)], 0)
-    parser.update([(1, [msg])])
-
-    assert parser.can_valid
-    assert parser.vl["LKAS12"]["CF_LkasDawStatus"] == 0
-    assert parser.vl["LKAS12"]["CF_Lkas_TsrSpeed_Display_Clu"] == 80
-
-    no_lkas12 = CarInterface.get_params(CAR.HYUNDAI_SONATA_HYBRID, gen_empty_fingerprint(), [], False, False, False, None)
-    no_lkas12_fpcp = CarInterface.get_starpilot_params(
-      CAR.HYUNDAI_SONATA_HYBRID, gen_empty_fingerprint(), [], no_lkas12, get_test_toggles(),
-    )
-    assert not (no_lkas12_fpcp.flags & HyundaiStarPilotFlags.HAS_LKAS12)
-
  def test_carnival_lka_button_does_not_enable_angle_steering_safety(self):
    fingerprint = gen_empty_fingerprint()
    fingerprint[0][0x391] = 8
@@ -2385,7 +2288,7 @@ class TestHyundaiFingerprint:
    assert parser.vl["LKAS_ALT"]["ADAS_ACIAnglTqRedcGainVal"] == pytest.approx(0.0)
    assert parser.vl["LKAS_ALT"]["ADAS_StrAnglReqVal"] == pytest.approx(8.5)

-  def test_gv70_electrified_uses_generic_lkas_status_payload(self):
+  def test_gv70_electrified_synthesizes_lkas_status_payload(self):
    CP = CarParams.new_message()
    CP.carFingerprint = CAR.GENESIS_GV70_ELECTRIFIED_1ST_GEN
    CP.flags = int(HyundaiFlags.CANFD | HyundaiFlags.EV | HyundaiFlags.CANFD_LKA_STEERING)
@@ -2428,11 +2331,9 @@ class TestHyundaiFingerprint:
    parser.update([(1, lkas_msgs)])
    assert parser.can_valid
    assert parser.vl["LKAS"]["HAS_LANE_SAFETY"] == 0
-    assert parser.vl["LKAS"]["DAMP_FACTOR"] == 0
+    assert parser.vl["LKAS"]["DAMP_FACTOR"] == 100
    assert parser.vl["LKAS"]["TORQUE_REQUEST"] == 0
    assert parser.vl["LKAS"]["STEER_REQ"] == 1
-    assert parser.vl["LKAS"]["STEER_MODE"] == 0
-    assert parser.vl["LKAS"]["NEW_SIGNAL_2"] == 0

    CP.openpilotLongitudinalControl = True
    lfa_parser = CANParser(DBC[CP.carFingerprint][Bus.pt], [("LFA", 0)], can_bus.ECAN)
@@ -124,7 +124,6 @@ class HyundaiStarPilotSafetyFlags(IntFlag):
 class HyundaiStarPilotFlags(IntFlag):
  SPEED_LIMIT_AVAILABLE = 1
  MAIN_CRUISE_STATE_TRACKING = 2 ** 2
-  HAS_LKAS12 = 2 ** 9


 class HyundaiFlags(IntFlag):
@@ -1204,7 +1203,6 @@ CANFD_SECURITYACCESS_CAR = {
 }
 CANFD_UNSUPPORTED_LONGITUDINAL_CAR = CAR.with_flags(HyundaiFlags.CANFD_NO_RADAR_DISABLE) - CANFD_SECURITYACCESS_CAR  # TODO: merge with UNSUPPORTED_LONGITUDINAL_CAR
 CANFD_ANGLE_LONGITUDINAL_CAR = {CAR.KIA_EV9, CAR.HYUNDAI_IONIQ_5_PE}
-CANFD_DAW_SUPPRESSION_CAR = {CAR.HYUNDAI_IONIQ_6}
 CANFD_ALT_BUTTONS_RESUME_CAR = {CAR.KIA_CARNIVAL_2025, CAR.KIA_CARNIVAL_HEV_4TH_GEN}
 CANFD_CORNER_RADAR_BSM_CAR = {CAR.HYUNDAI_IONIQ_6, CAR.HYUNDAI_IONIQ_5_PE, CAR.KIA_EV9}
 CANFD_RADAR_LIVE_LONGITUDINAL_CAR = {
@@ -240,9 +240,6 @@ class CarInterfaceBase(ABC):
          if 0x1FA in fingerprint[CAN.ECAN]:
            fp_ret.flags |= HyundaiStarPilotFlags.SPEED_LIMIT_AVAILABLE.value

-        if not (CP.flags & HyundaiFlags.CANFD) and 0x53E in fingerprint[2]:
-          fp_ret.flags |= HyundaiStarPilotFlags.HAS_LKAS12.value
-
        fp_ret.redneckCruiseAvailable = bool(CP.flags & HyundaiFlags.NON_SCC) and not bool(CP.flags & HyundaiFlags.CANFD_ALT_BUTTONS)
        if fp_ret.redneckCruiseAvailable and params.get_bool("RedneckCruise"):
          fp_ret.pcmCruiseSpeed = False
@@ -29,7 +29,6 @@ const LongitudinalLimits HYUNDAI_LONG_LIMITS = {
  {0x340, 0,                           8, .check_relay = true},   /* LKAS11 Bus 0                              */ \
  {0x4F1, scc_bus,                     4, .check_relay = false},  /* CLU11 Bus 0 (radar-SCC) or 2 (camera-SCC) */ \
  {0x485, 0,       (can_refresh) ? 8 : 4, .check_relay = true},   /* LFAHDA_MFC Bus 0                          */ \
-  {0x53E, 0,                           6, .check_relay = false},  /* LKAS12 replacement after camera advertises it */ \

 #define HYUNDAI_LONG_COMMON_TX_MSGS(scc_bus, can_refresh) \
  HYUNDAI_COMMON_TX_MSGS(scc_bus, can_refresh) \
@@ -141,12 +140,6 @@ static uint32_t hyundai_get_checksum(const CANPacket_t *msg) {
  return chksum;
 }

-static void hyundai_rx_all_hook(const CANPacket_t *msg) {
-  if ((msg->addr == 0x53EU) && (msg->bus == 2U) && (GET_LEN(msg) == 6U)) {
-    hyundai_has_lkas12 = true;
-  }
-}
-
 static uint32_t hyundai_compute_checksum(const CANPacket_t *msg) {
  uint8_t chksum = 0;
  if (msg->addr == 0x386U) {
@@ -291,10 +284,6 @@ static bool hyundai_tx_hook(const CANPacket_t *msg) {

  bool tx = true;

-  if ((msg->addr == 0x53EU) && !hyundai_has_lkas12) {
-    tx = false;
-  }
-
  // FCA11: Block any potential actuation. The blended HDA II layout uses
  // different static fields, but its explicit AEB/FCA request bits stay zero.
  if (msg->addr == 0x38DU) {
@@ -706,7 +695,6 @@ static safety_config hyundai_legacy_init(uint16_t param) {
 const safety_hooks hyundai_hooks = {
  .init = hyundai_init,
  .rx = hyundai_rx_hook,
-  .rx_all = hyundai_rx_all_hook,
  .tx = hyundai_tx_hook,
  .get_counter = hyundai_get_counter,
  .get_checksum = hyundai_get_checksum,
@@ -716,7 +704,6 @@ const safety_hooks hyundai_hooks = {
 const safety_hooks hyundai_legacy_hooks = {
  .init = hyundai_legacy_init,
  .rx = hyundai_rx_hook,
-  .rx_all = hyundai_rx_all_hook,
  .tx = hyundai_tx_hook,
  .get_counter = hyundai_get_counter,
  .get_checksum = hyundai_get_checksum,
@@ -16,7 +16,6 @@

 #define HYUNDAI_CANFD_LKA_STEERING_ALT_COMMON_TX_MSGS(a_can, e_can) \
  HYUNDAI_CANFD_CRUISE_BUTTON_TX_MSGS(e_can)                        \
-  {0x11A, e_can, 16, .check_relay = (e_can) == 0},  /* FR_CMR_01_10ms */ \
  {0x110, a_can, 32, .check_relay = (a_can) == 0, .disable_static_blocking = true},  /* LKAS_ALT */  \
  {0x362, a_can, 32, .check_relay = (a_can) == 0, .disable_static_blocking = true},  /* CAM_0x362 */ \

@@ -63,9 +63,6 @@ bool hyundai_cancel_button_enable = false;
 extern bool hyundai_can_refresh_msgs;
 bool hyundai_can_refresh_msgs = false;

-extern bool hyundai_has_lkas12;
-bool hyundai_has_lkas12 = false;
-
 extern bool hyundai_elantra_hev_2024;
 bool hyundai_elantra_hev_2024 = false;

@@ -109,7 +106,6 @@ void hyundai_common_init(uint16_t param) {
  hyundai_non_scc = GET_FLAG(param, HYUNDAI_PARAM_NON_SCC);
  hyundai_cancel_button_enable = GET_FLAG(param, HYUNDAI_PARAM_CANCEL_BTN_ENABLE);
  hyundai_can_refresh_msgs = GET_FLAG(param, HYUNDAI_PARAM_CAN_REFRESH_MSGS);
-  hyundai_has_lkas12 = false;
  hyundai_elantra_hev_2024 = hyundai_can_refresh_msgs && hyundai_hybrid_gas_signal && hyundai_camera_scc;
  hyundai_aol_main_lkas_sync = false;

@@ -427,18 +427,6 @@ def test_hyundai_starpilot_rx_sources():
  safety.safety_rx_hook(libsafety_py.make_CANPacket(0x421, 0, bytes(8)))


-def test_hyundai_lkas12_tx_requires_stock_camera_message():
-  safety = libsafety_py.libsafety
-  assert safety.set_safety_hooks(CarParams.SafetyModel.hyundai, 0) == 0
-  safety.init_tests()
-
-  lkas12 = libsafety_py.make_CANPacket(0x53E, 0, bytes(6))
-  assert not safety.safety_tx_hook(lkas12)
-
-  safety.safety_rx_hook(libsafety_py.make_CANPacket(0x53E, 2, bytes(6)))
-  assert safety.safety_tx_hook(lkas12)
-
-
 class TestHyundaiLongitudinalSafety(HyundaiLongitudinalBase, TestHyundaiSafety):
  TX_MSGS = [[0x340, 0], [0x4F1, 0], [0x485, 0], [0x420, 0], [0x421, 0], [0x50A, 0], [0x389, 0], [0x4A2, 0], [0x38D, 0], [0x483, 0], [0x7D0, 0]]

@@ -568,7 +568,7 @@ class TestHyundaiCanfdLKASteeringEV(TestHyundaiCanfdBase):
 # TODO: Handle ICE and HEV configurations once we see cars that use the new messages
 class TestHyundaiCanfdLKASteeringAltEV(TestHyundaiCanfdBase):

-  TX_MSGS = [[0x11A, 1], [0x110, 0], [0x1CF, 1], [0x362, 0]]
+  TX_MSGS = [[0x110, 0], [0x1CF, 1], [0x362, 0]]
  RELAY_MALFUNCTION_ADDRS = {0: (0x110, 0x362)}  # LKAS_ALT, CAM_0x362
  FWD_BLACKLISTED_ADDRS = {2: [0x110, 0x362]}

@@ -587,7 +587,7 @@ class TestHyundaiCanfdLKASteeringAltEV(TestHyundaiCanfdBase):

 class TestHyundaiCanfdLKASteeringAltButtonsICE(TestHyundaiCanfdLKASteeringAltEV):

-  TX_MSGS = [[0x11A, 1], [0x110, 0], [0x1CF, 1], [0x1A0, 1], [0x362, 0]]
+  TX_MSGS = [[0x110, 0], [0x1CF, 1], [0x1A0, 1], [0x362, 0]]
  GAS_MSG = ("ACCELERATOR_BRAKE_ALT", "ACCELERATOR_PEDAL_PRESSED")

  def setUp(self):
@@ -691,7 +691,7 @@ class TestHyundaiCanfdLKASteeringLongEV(HyundaiLongitudinalBase, TestHyundaiCanf

 class TestHyundaiCanfdLKASteeringAltAngleLongEV(HyundaiLongitudinalBase, TestHyundaiCanfdAngleSteering):

-  TX_MSGS = [[0x11A, 1], [0x110, 0], [0x1CF, 1], [0x362, 0], [0x51, 0], [0x100, 0], [0x730, 1], [0x12a, 1], [0x160, 1],
+  TX_MSGS = [[0x110, 0], [0x1CF, 1], [0x362, 0], [0x51, 0], [0x100, 0], [0x730, 1], [0x12a, 1], [0x160, 1],
             [0x1ba, 1], [0x1e0, 1], [0x1e5, 1], [0x31a, 1], [0x3b5, 1], [0x3c1, 1],
             [0x1a0, 1], [0x1ea, 1], [0x200, 1], [0x345, 1], [0x1da, 1]]

@@ -1,2 +1,2 @@
 extern const uint8_t gitversion[19];
-const uint8_t gitversion[19] = "DEV-5fe8b174-DEBUG";
+const uint8_t gitversion[19] = "DEV-3ebc6b99-DEBUG";
@@ -1 +1 @@
-DEV-5fe8b174-DEBUG
+DEV-3ebc6b99-DEBUG
@@ -56,14 +56,13 @@ def build_compile_env(*, supercombo: bool = False) -> dict[str, str]:
  existing_pythonpath = env.get("PYTHONPATH", "")
  env["PYTHONPATH"] = f"{REPO_ROOT}{os.pathsep}{existing_pythonpath}" if existing_pythonpath else str(REPO_ROOT)
  defaults = {
+    "DEBUG": "0",
    "FLOAT16": "1",
    "IMAGE": "1" if supercombo else "2",
    "JIT_BATCH_SIZE": "0",
    "NOLOCALS": "1",
    "OPENPILOT_HACKS": "1",
-  } | ({} if supercombo else {
-    "DEBUG": "0",
-  })
+  }
  for key, default in defaults.items():
    try:
      int(str(env.get(key)), 0)
@@ -31,7 +31,7 @@ OPENPILOT_REPO = "commaai/openpilot"
 RESOURCES_REPO = os.environ.get("STARPILOT_RESOURCES_REPO", "firestar5683/StarPilot-Resources")
 HF_BUCKET = os.environ.get("STARPILOT_HF_BUCKET", "StarPilot-Driving/StarPilot-Resources")
 RESOURCE_BRANCH = "Models"
-MANIFEST_VERSION = "v24"
+MANIFEST_VERSION = "v25"
 DEFAULT_BEHAVIOR_VERSION = "v16"
 DEVICE_ROOT = "/data/openpilot"
 REPOSITORY_FILE_LIMIT = 100_000_000
@@ -79,14 +79,14 @@ def test_runtime_scan_excludes_model_weights_but_flags_runtime_code():


 def test_update_manifest_replaces_one_entry(tmp_path: Path):
-  manifest = tmp_path / "model_names_v24.json"
+  manifest = tmp_path / "model_names_v25.json"
  manifest.write_text(json.dumps({"models": [{"id": "old"}]}) + "\n")
  info = parse_pasted_release(RELEASE_TEXT, "bmrlnapv4", "v16")
  path = update_manifest(
    tmp_path,
    info,
    {"size": 123, "sha256": "a" * 64},
-    "v24",
+    "v25",
  )
  payload = json.loads(path.read_text())
  assert len(payload["models"]) == 2
@@ -253,12 +253,12 @@ GENESIS_G70_FRICTION_CENTER_LAT = 0.28
 GENESIS_G70_FRICTION_CENTER_LAT_WIDTH = 0.10
 GENESIS_G70_FRICTION_CALM_JERK = 0.35
 GENESIS_G70_FRICTION_CALM_JERK_WIDTH = 0.10
-GENESIS_G70_FRICTION_JERK_DEADZONE_MAX = 0.39
+GENESIS_G70_FRICTION_JERK_DEADZONE_MAX = 0.30
 GENESIS_G70_FRICTION_JERK_DEADZONE_LAT = 0.30
 GENESIS_G70_FRICTION_JERK_DEADZONE_LAT_WIDTH = 0.08
 GENESIS_G70_FRICTION_JERK_DEADZONE_SPEED = 12.0
 GENESIS_G70_FRICTION_JERK_DEADZONE_SPEED_WIDTH = 3.5
-GENESIS_G70_CENTER_OUTPUT_TAPER_MAX = 0.22
+GENESIS_G70_CENTER_OUTPUT_TAPER_MAX = 0.16
 GENESIS_G70_CENTER_OUTPUT_TAPER_LAT = 0.30
 GENESIS_G70_CENTER_OUTPUT_TAPER_LAT_WIDTH = 0.10
 GENESIS_G70_CENTER_OUTPUT_TAPER_SPEED = 18.0
@@ -288,11 +288,11 @@ GENESIS_G70_CURVE_UNWIND_LAT = 0.25
 GENESIS_G70_CURVE_UNWIND_LAT_WIDTH = 0.12
 GENESIS_G70_CURVE_UNWIND_JERK = 0.08
 GENESIS_G70_CURVE_UNWIND_JERK_WIDTH = 0.08
-GENESIS_G70_UNWIND_FF_REDUCTION_MAX = 0.28
-GENESIS_G70_UNWIND_FF_OVERSHOOT = 0.18
-GENESIS_G70_UNWIND_FF_OVERSHOOT_WIDTH = 0.20
+GENESIS_G70_UNWIND_FF_REDUCTION_MAX = 0.36
+GENESIS_G70_UNWIND_FF_OVERSHOOT = 0.10
+GENESIS_G70_UNWIND_FF_OVERSHOOT_WIDTH = 0.12
 GENESIS_G70_UNWIND_FF_JERK = 0.10
-GENESIS_G70_UNWIND_FF_JERK_WIDTH = 0.13
+GENESIS_G70_UNWIND_FF_JERK_WIDTH = 0.10
 GENESIS_G70_UNWIND_FF_SPEED = 18.0
 GENESIS_G70_UNWIND_FF_SPEED_WIDTH = 3.0
 GENESIS_G70_HIGH_SPEED_ERROR_DAMPING_MAX = 0.15
@@ -302,6 +302,7 @@ GENESIS_G70_HIGH_SPEED_ERROR_DAMPING_ERROR = 0.18
 GENESIS_G70_HIGH_SPEED_ERROR_DAMPING_ERROR_WIDTH = 0.15
 GENESIS_G70_HIGH_SPEED_ERROR_DAMPING_JERK = 0.15
 GENESIS_G70_HIGH_SPEED_ERROR_DAMPING_JERK_WIDTH = 0.10
+GENESIS_G70_HIGH_SPEED_OVERSHOOT_PHASE_WEIGHT = 0.60
 GENESIS_G70_ANGLE_OUTPUT_TAPER_MIN = 0.45
 GENESIS_G70_ANGLE_OUTPUT_TAPER_START = 70.0
 GENESIS_G70_ANGLE_OUTPUT_TAPER_WIDTH = 6.0
@@ -3237,6 +3238,8 @@ def get_genesis_g70_high_speed_error_scale(setpoint: float, measured_lateral_acc
  jerk_weight = _sigmoid((abs(desired_lateral_jerk) - GENESIS_G70_HIGH_SPEED_ERROR_DAMPING_JERK) /
                         GENESIS_G70_HIGH_SPEED_ERROR_DAMPING_JERK_WIDTH)
  phase_weight = 1.0 if setpoint * desired_lateral_jerk < 0.0 else 0.45
+  if setpoint * measured_lateral_accel > 0.0 and abs(measured_lateral_accel) > abs(setpoint):
+    phase_weight = max(phase_weight, GENESIS_G70_HIGH_SPEED_OVERSHOOT_PHASE_WEIGHT)
  reduction = (GENESIS_G70_HIGH_SPEED_ERROR_DAMPING_MAX * speed_weight * error_weight *
               (0.35 + (0.65 * jerk_weight)) * phase_weight)
  return 1.0 - reduction
@@ -20,7 +20,7 @@ HONDA_ACCORD_STOP_GO_MAX_LATERAL_OFFSET = 1.25
 HONDA_ACCORD_STOP_GO_MIN_MODEL_PROB = 0.95
 HONDA_ACCORD_STOP_GO_ACCEL_RISE_RATE = 4.0
 HYUNDAI_ELANTRA_LEAD_FOLLOW_JERK_SCALE = 1.25
-GENESIS_GV70_ELECTRIFIED_LEAD_FOLLOW_JERK_SCALE = 1.75
+GENESIS_GV70_ELECTRIFIED_LEAD_FOLLOW_JERK_SCALE = 1.35
 FORD_LIGHTNING_LEAD_FOLLOW_JERK_SCALE = 1.35
 HONDA_CRV_5G_LEAD_FOLLOW_JERK_SCALE = 1.35
 GM_SILVERADO_EARLY_FOLLOW_MIN_EGO_SPEED = 18.0
@@ -42,7 +42,7 @@ def test_force_stop_jerk_scale_is_platform_specific():

 def test_lead_follow_jerk_scale_is_platform_specific():
  assert get_lead_follow_jerk_scale(SimpleNamespace(brand="hyundai", carFingerprint="HYUNDAI_ELANTRA_2021")) == 1.25
-  assert get_lead_follow_jerk_scale(SimpleNamespace(brand="hyundai", carFingerprint="GENESIS_GV70_ELECTRIFIED_1ST_GEN")) == 1.75
+  assert get_lead_follow_jerk_scale(SimpleNamespace(brand="hyundai", carFingerprint="GENESIS_GV70_ELECTRIFIED_1ST_GEN")) == 1.35
  assert get_lead_follow_jerk_scale(SimpleNamespace(brand="ford", carFingerprint="FORD_F_150_LIGHTNING_MK1")) == 1.35
  assert get_lead_follow_jerk_scale(SimpleNamespace(brand="honda", carFingerprint="HONDA_CRV_5G")) == 1.35
  assert get_lead_follow_jerk_scale(SimpleNamespace(brand="other", carFingerprint="OTHER_CAR")) == 1.0
@@ -7,6 +7,10 @@ import struct
 from openpilot.system.hardware import HARDWARE, TICI
 os.environ['GMMU'] = '0'
 os.environ['DEV'] = 'QCOM' if TICI else 'LLVM'
+try:
+  int(os.getenv('DEBUG', '0'), 0)
+except ValueError:
+  os.environ['DEBUG'] = '0'
 from tinygrad.device import Device
 from tinygrad.tensor import Tensor
 import time
@@ -1 +1 @@
-a77db33c2e2d6a7570dc2a4a70c2b877429ee8bd9ca5dfeda74b5a41231aaff9  driving_tinygrad.pkl
+31902b114b7fb8455af694d83333a86a44112b83be662e064ffbd67e8daafe72  driving_tinygrad.pkl
@@ -24,7 +24,7 @@ from openpilot.starpilot.common.starpilot_utilities import delete_file
 from openpilot.starpilot.common.starpilot_variables import MODELS_PATH
 from openpilot.system.hardware.usb import chestnut_firmware_ready

-MANIFEST_CANDIDATES = ("v24",)
+MANIFEST_CANDIDATES = ("v25",)
 MODEL_NAMESPACE_SUFFIX = "3"
 DEFAULT_MODEL_KEY = "rdf43"
 LOCAL_MODEL_PREFIX = "local-"
@@ -15,12 +15,12 @@ from openpilot.starpilot.assets.model_manager import MANIFEST_CANDIDATES, ModelM
 from openpilot.starpilot.common.model_versions import UNIFIED_ARTIFACT_FORMAT


-def test_v24_is_the_only_manifest_candidate():
-  assert MANIFEST_CANDIDATES == ("v24",)
+def test_v25_is_the_only_manifest_candidate():
+  assert MANIFEST_CANDIDATES == ("v25",)


-def test_v24_manifest_is_loaded_from_models_checkout():
-  assert ModelManager._manifest_paths("v24") == ("Models/model_names_v24.json",)
+def test_v25_manifest_is_loaded_from_models_checkout():
+  assert ModelManager._manifest_paths("v25") == ("Models/model_names_v25.json",)


 def test_resource_sources_prefer_huggingface_then_github(monkeypatch):
@@ -33,9 +33,9 @@ def test_resource_sources_prefer_huggingface_then_github(monkeypatch):


 def test_huggingface_manifest_has_root_and_manifests_fallbacks():
-  assert ModelManager._hf_manifest_paths("v24") == (
-    "model_names_v24.json",
-    "manifests/model_names_v24.json",
+  assert ModelManager._hf_manifest_paths("v25") == (
+    "model_names_v25.json",
+    "manifests/model_names_v25.json",
  )


@@ -56,30 +56,30 @@
  },
  {
    "name": "boot",
-    "url": "https://files.firestar.link/x/wqn7kn631kdx/boot27.img.xz",
+    "url": "https://files.firestar.link/x/83cmdtzb2sav/boot25.img.xz",
    "fallback_urls": [
-      "https://files-east.firestar.link/x/0zzizffpkz3a/boot27.img.xz"
+      "https://files-east.firestar.link/x/o3r7dpay7yj5/boot25.img.xz"
    ],
-    "hash": "173208aec963fa35a6d9a92c7ca77c5f2b49797def096f3cce05591e78b12460",
-    "hash_raw": "173208aec963fa35a6d9a92c7ca77c5f2b49797def096f3cce05591e78b12460",
-    "size": 48781312,
+    "hash": "846d9e41fe0eb3543a01cb9a55a6bb540e4097dd838ada890674b654bd97a49a",
+    "hash_raw": "846d9e41fe0eb3543a01cb9a55a6bb540e4097dd838ada890674b654bd97a49a",
+    "size": 48302080,
    "sparse": false,
    "full_check": true,
    "has_ab": true,
-    "ondevice_hash": "bc253532027b4756c09ea235b6fd794641d115b6ddbc05c417d6eb89475a22b8"
+    "ondevice_hash": "846d9e41fe0eb3543a01cb9a55a6bb540e4097dd838ada890674b654bd97a49a"
  },
  {
    "name": "system",
-    "url": "https://files.firestar.link/x/877rsvjzkmp8/system27.img.xz",
+    "url": "https://files.firestar.link/x/ser56g04xioj/system25.img.xz",
    "fallback_urls": [
-      "https://files-east.firestar.link/x/5ni1fkd2k4hg/system27.img.xz"
+      "https://files-east.firestar.link/x/90fsutnuzy13/system25.img.xz"
    ],
-    "hash": "205dfc25b21eb236cefaabd081aa4ad6e15ccd1c9f9fc623ec29de0c1c70fb06",
-    "hash_raw": "6f204321089c7ec4bae2e8b10eda827c502e6b516e6095a3b659016b4a646d38",
+    "hash": "f81ac06c0634ed0dadd5b4778f3b5052a2779d69669b7e70e21505c951ce6ccc",
+    "hash_raw": "f81ac06c0634ed0dadd5b4778f3b5052a2779d69669b7e70e21505c951ce6ccc",
    "size": 4718592000,
-    "sparse": true,
+    "sparse": false,
    "full_check": false,
    "has_ab": true,
-    "ondevice_hash": "a8edcedcf4e6fb2a9af2e723a54f763f1f542fb63742cd1efc913ed30035c237"
+    "ondevice_hash": "f81ac06c0634ed0dadd5b4778f3b5052a2779d69669b7e70e21505c951ce6ccc"
  }
 ]
@@ -4,7 +4,7 @@ inputs:
  python-version:
    description: 'Python version to use'
    required: false
-    default: '' # if you don't set a version, the native python version will be used
+    default: '3.14'
  key:
    description: 'Key for the python cache'
    required: false
@@ -41,12 +41,12 @@ inputs:
    description: "Install LLVM?"
    required: false
    default: 'false'
-  tinydreno:
-    description: "Install tinydreno"
+  qemu:
+    description: "Install qemu?"
    required: false
    default: 'false'
-  qemu:
-    description: "Install qemu"
+  ninja:
+    description: "Install ninja?"
    required: false
    default: 'false'
 runs:
@@ -59,18 +59,18 @@ runs:
        echo "OMP_NUM_THREADS=1" >> "$GITHUB_ENV"
        # no buffers should be over 300MB in CI
        echo "MAX_BUFFER_SIZE=300000000" >> "$GITHUB_ENV"
+        if [[ "$RUNNER_OS" == "Linux" ]]; then
+          echo "VIRTUAL_ENV=/opt/venv/${{ inputs.python-version }}" >> "$GITHUB_ENV"
+          echo "UV_PYTHON_INSTALL_DIR=/opt/python" >> "$GITHUB_ENV"
+        else
+          echo "VIRTUAL_ENV=${{ github.workspace }}/.venv" >> "$GITHUB_ENV"
+        fi

    - name: Set up uv
      uses: astral-sh/setup-uv@08807647e7069bb48b6ef5acd8ec9567f424441b
      with:
        enable-cache: 'false' # see below for manual caching

-    - name: Set up Python ${{ inputs.python-version }}
-      uses: actions/setup-python@v6
-      if: inputs.python-version != ''
-      with:
-        python-version: ${{ inputs.python-version }}
-
    # **** Caching packages ****

    - name: Cache Python packages (PR)
@@ -109,15 +109,15 @@ runs:
      if: inputs.deps != ''
      shell: bash
      run: |
-        uv venv .venv
+        uv venv --allow-existing --python ${{ inputs.python-version }} "$VIRTUAL_ENV"
        DEPS="${{ inputs.deps }}"
-        uv pip install --python .venv -e ".[${DEPS// /,}]" ${{ inputs.pydeps }} --torch-backend cpu --extra-index-url https://aiinfra.pkgs.visualstudio.com/PublicPackages/_packaging/Triton-Nightly/pypi/simple/
+        uv pip install --python "$VIRTUAL_ENV" -e ".[${DEPS// /,}]" ${{ inputs.pydeps }} --torch-backend cpu --extra-index-url https://aiinfra.pkgs.visualstudio.com/PublicPackages/_packaging/Triton-Nightly/pypi/simple/
    - name: Install dependencies in venv (without extra)
      if: inputs.deps == ''
      shell: bash
      run: |
-        uv venv .venv
-        uv pip install --python .venv -e . ${{ inputs.pydeps }}
+        uv venv --allow-existing --python ${{ inputs.python-version }} "$VIRTUAL_ENV"
+        uv pip install --python "$VIRTUAL_ENV" -e . ${{ inputs.pydeps }}
    - name: Prune uv cache
      if: github.event_name != 'pull_request'
      shell: bash
@@ -125,16 +125,15 @@ runs:
    - name: Configure venv
      shell: bash
      run: |
-        echo "VIRTUAL_ENV=${{ github.workspace }}/.venv" >> "$GITHUB_ENV"
        if [[ "$RUNNER_OS" == "Windows" ]]; then
-          echo "${{ github.workspace }}/.venv/Scripts" >> "$GITHUB_PATH"
+          echo "$VIRTUAL_ENV/Scripts" >> "$GITHUB_PATH"
        else
-          echo "${{ github.workspace }}/.venv/bin" >> "$GITHUB_PATH"
+          echo "$VIRTUAL_ENV/bin" >> "$GITHUB_PATH"
        fi

    # ******************* apt *******************
    - name: Setup apt
-      if: runner.os == 'Linux' && (inputs.opencl == 'true' || inputs.amd == 'true' || inputs.webgpu == 'true' || inputs.llvm == 'true' || inputs.qemu == 'true')
+      if: runner.os == 'Linux' && (inputs.opencl == 'true' || inputs.amd == 'true' || inputs.webgpu == 'true' || inputs.llvm == 'true' || inputs.qemu == 'true' || inputs.ninja == 'true')
      shell: bash
      run: |
        sudo mkdir -p /var/cache/apt/archives
@@ -162,7 +161,7 @@ runs:
        echo "deb http://apt.llvm.org/$(lsb_release -cs)/ llvm-toolchain-$(lsb_release -cs)-20 main" | sudo tee /etc/apt/sources.list.d/llvm.list

    - name: Compute Package List + Hash
-      if: runner.os == 'Linux' && (inputs.opencl == 'true' || inputs.amd == 'true' || inputs.webgpu == 'true' || inputs.llvm == 'true' || inputs.qemu == 'true')
+      if: runner.os == 'Linux' && (inputs.opencl == 'true' || inputs.amd == 'true' || inputs.webgpu == 'true' || inputs.llvm == 'true' || inputs.qemu == 'true' || inputs.ninja == 'true')
      id: apt-pkgs
      shell: bash
      run: |
@@ -187,25 +186,37 @@ runs:
        if [[ "${{ inputs.qemu }}" == "true" ]]; then
          pkgs+=" qemu-user-static"
        fi
+        # **** ninja ****
+        if [[ "${{ inputs.ninja }}" == "true" ]]; then
+          pkgs+=" ninja-build"
+        fi

        echo "pkgs=$pkgs" >> "$GITHUB_OUTPUT"
        echo "hash=$(echo -n "$pkgs" | sha256sum | cut -d' ' -f1)" >> "$GITHUB_OUTPUT"

+        installed=true
+        for pkg in $pkgs; do
+          info=$(dpkg-query -W -f='${db:Status-Abbrev} ${Version}' "$pkg" 2> /dev/null || true)
+          echo "${pkg}: ${info:-not in dpkg database}"
+          [[ "$info" == ii* ]] || installed=false
+        done
+        echo "installed=$installed" >> "$GITHUB_OUTPUT"
+
    - name: Cache apt (PR)
-      if: runner.os == 'Linux' && (inputs.opencl == 'true' || inputs.amd == 'true' || inputs.webgpu == 'true' || inputs.llvm == 'true' || inputs.qemu == 'true') && github.event_name == 'pull_request'
+      if: runner.os == 'Linux' && (inputs.opencl == 'true' || inputs.amd == 'true' || inputs.webgpu == 'true' || inputs.llvm == 'true' || inputs.qemu == 'true' || inputs.ninja == 'true') && github.event_name == 'pull_request' && steps.apt-pkgs.outputs.installed == 'false'
      uses: actions/cache/restore@v5
      with:
        path: /var/cache/apt/archives/
        key: ${{ runner.os }}-${{ runner.arch }}-apt-${{ steps.apt-pkgs.outputs.hash }}-${{ env.CACHE_VERSION }}
    - name: Cache apt
-      if: runner.os == 'Linux' && (inputs.opencl == 'true' || inputs.amd == 'true' || inputs.webgpu == 'true' || inputs.llvm == 'true' || inputs.qemu == 'true') && github.event_name != 'pull_request'
+      if: runner.os == 'Linux' && (inputs.opencl == 'true' || inputs.amd == 'true' || inputs.webgpu == 'true' || inputs.llvm == 'true' || inputs.qemu == 'true' || inputs.ninja == 'true') && github.event_name != 'pull_request' && steps.apt-pkgs.outputs.installed == 'false'
      uses: actions/cache@v5
      with:
        path: /var/cache/apt/archives/
        key: ${{ runner.os }}-${{ runner.arch }}-apt-${{ steps.apt-pkgs.outputs.hash }}-${{ env.CACHE_VERSION }}

    - name: Run apt Update + Install
-      if: runner.os == 'Linux' && (inputs.opencl == 'true' || inputs.amd == 'true' || inputs.webgpu == 'true' || inputs.llvm == 'true' || inputs.qemu == 'true')
+      if: runner.os == 'Linux' && (inputs.opencl == 'true' || inputs.amd == 'true' || inputs.webgpu == 'true' || inputs.llvm == 'true' || inputs.qemu == 'true' || inputs.ninja == 'true') && steps.apt-pkgs.outputs.installed == 'false'
      shell: bash
      run: |
        sudo apt -qq update || true
@@ -277,12 +288,6 @@ runs:
      shell: bash
      run: brew install llvm@20

-    # *** tinydreno ***
-    - name: Install tinydreno (linux)
-      if: inputs.tinydreno == 'true' && runner.os == 'Linux'
-      shell: bash
-      run: sudo curl -fL https://github.com/sirhcm/tinydreno/raw/refs/heads/master/libllvm-qcom.so -o /usr/lib/libllvm-qcom.so
-
    # *** OpenCL ***
    - name: Install rusticl
      if: inputs.opencl == 'true'
@@ -35,15 +35,15 @@ jobs:
        key: 'autogen'
        amd: 'true'
        llvm: 'true'
-        pydeps: 'pyyaml mako'
+        deps: 'autogen'
    - name: Install autogen support packages
      run: sudo apt-get install -y --no-install-recommends libclang-20-dev llvm-20-dev hip-dev libusb-1.0-0-dev libdrm-dev liburing-dev
    - name: Regenerate autogen files
      run: |
-        find tinygrad/runtime/autogen -type f -name "*.py" -not -path "*/amd/*" -not -name "__init__.py" -not -name "comgr.py" -not -name "metal.py" -not -name "iokit.py" -not -name "corefoundation.py" -not -name "libclang.py" -delete
+        find tinygrad/runtime/autogen -type f -name "*.py" -not -path "*/amd/*" -not -name "__init__.py" -not -name "metal.py" -not -name "iokit.py" -not -name "corefoundation.py" -not -name "libclang.py" -delete
        python3 -c "from tinygrad.runtime.autogen import opencl"
        python3 -c "from tinygrad.runtime.autogen import cuda, nvrtc, nvjitlink, nv_570, nv_580, nv_610, nv"
-        python3 -c "from tinygrad.runtime.autogen import comgr_3, hsa, hip, amd_gpu, sqtt, rocprof, amdgpu_kd, amdgpu_drm"
+        python3 -c "from tinygrad.runtime.autogen import comgr, comgr_3, hsa, hip, amd_gpu, sqtt, rocprof, amdgpu_kd, amdgpu_drm"
        python3 -c "from tinygrad.runtime.autogen.am import *"
        python3 -c "from tinygrad.runtime.autogen.nv_regs import *"
        python3 -c "from tinygrad.runtime.autogen import libc, kfd, io_uring, pci, vfio"
@@ -54,7 +54,7 @@ jobs:
        python3 -c "from tinygrad.runtime.autogen import mesa"
        python3 -c "from tinygrad.runtime.autogen import avcodec"
        python3 -c "from tinygrad.runtime.autogen import llvm_qcom"
-        python3 -c "from tinygrad.runtime.autogen import mlx5"
+        python3 -c "from tinygrad.runtime.autogen import mlx5, bnxt"
        python3 -c "from tinygrad.runtime.autogen import ggml_common"
        REGEN=1 python3 -c "from tinygrad.runtime.autogen import libclang"
    - name: Check for differences
@@ -102,42 +102,3 @@ jobs:
      with:
        name: autogen-macos-patch
        path: autogen-macos.patch
-
-  autogen-comgr-2:
-    name: In-tree Autogen (comgr 2)
-    runs-on: ubuntu-24.04
-    timeout-minutes: 15
-    steps:
-    - name: Checkout Code
-      uses: actions/checkout@v6
-    - name: Setup Environment
-      uses: ./.github/actions/setup-tinygrad
-      with:
-        key: 'autogen-comgr'
-    - name: Install autogen support packages
-      run: |
-        wget https://repo.radeon.com/rocm/rocm.gpg.key -O - | gpg --dearmor | sudo tee /etc/apt/keyrings/rocm.gpg > /dev/null
-        sudo tee /etc/apt/sources.list.d/rocm.list <<EOF
-        deb [arch=amd64 signed-by=/etc/apt/keyrings/rocm.gpg] https://repo.radeon.com/rocm/apt/6.2 $(lsb_release -cs) main
-        EOF
-        echo -e 'Package: *\nPin: release o=repo.radeon.com\nPin-Priority: 600' | sudo tee /etc/apt/preferences.d/rocm-pin-600
-        sudo apt -qq update || true
-        sudo apt-get install -y --no-install-recommends libclang-20-dev comgr
-    - name: Regenerate autogen files
-      run: |
-        rm tinygrad/runtime/autogen/comgr.py
-        python3 -c "from tinygrad.runtime.autogen import comgr"
-    - name: Check for differences
-      run: |
-        if ! git diff --quiet; then
-          git diff
-          git diff > autogen-comgr2.patch
-          echo "Autogen mismatch detected. Patch available at: ${{ github.server_url }}/${{ github.repository }}/actions/runs/${{ github.run_id }}#artifacts"
-          exit 1
-        fi
-    - name: Upload patch artifact
-      if: failure()
-      uses: actions/upload-artifact@v7
-      with:
-        name: autogen-comgr2-patch
-        path: autogen-comgr2.patch
@@ -88,13 +88,13 @@ jobs:
      fail-fast: false
      matrix:
        dev: ['METAL', 'AMD', 'NV']
-    timeout-minutes: 60
+    timeout-minutes: 30
    defaults:
      run:
        shell: bash -e -o pipefail {0}
    env:
      DEV: ${{ matrix.dev }}
-      HCQ2: ${{ matrix.dev == 'AMD' && '1' || '0' }}
+      HCQ2: '0'
    if: github.repository_owner == 'tinygrad'
    steps:
    - name: Checkout Code
@@ -102,16 +102,11 @@ jobs:
    - name: Setup (AMD)
      if: ${{ matrix.dev == 'AMD' }}
      run: |
-        ./extra/amdpci/setup_python_cap.sh
-        ./extra/hcq/hcq_smi.py amd rmmod
-        ./extra/hcq/hcq_smi.py amd kill_pids
+        ./extra/hcq/hcq_smi.py amd rmmod --expect
+        ./extra/hcq/hcq_smi.py amd kill_pids --sudoless
    - name: Setup (NV)
      if: ${{ matrix.dev == 'NV' }}
-      run: sudo lsof -tQ /dev/nvidia* | { xargs -r sudo kill -9 || true; }
-    - name: Symlink models and datasets
-      run: |
-        mkdir -p weights
-        ln -s /raid/weights/LLaMA-3 weights/LLaMA-3
+      run: lsof -tQ /dev/nvidia* | { xargs -r kill -9 || true; }
    - name: setup staging db
      if: github.ref == 'refs/heads/update_benchmark_staging'
      run: |
@@ -121,18 +116,14 @@ jobs:
      run: python3 test/external/process_replay/reset.py
    - name: Run llama3.2
      run: BENCHMARK_LOG=llama32_3b-f16 JITBEAM=2 IGNORE_BEAM_CACHE=1 python3 -m tinygrad.llm -m llama3.2:3b-f16 --benchmark --warmup
-    - name: Run qwen3.6
-      # qwen3.6:35b-a3b doesn't fit on mac
+    - name: Run qwen3.8
+      # qwen3.8:27b doesn't fit on mac
      if: ${{ matrix.dev != 'METAL' }}
-      run: BENCHMARK_LOG=qwen36_35b-a3b JITBEAM=2 IGNORE_BEAM_CACHE=1 python3 -m tinygrad.llm -m qwen3.6:35b-a3b --benchmark --warmup
+      run: BENCHMARK_LOG=qwen38_27b JITBEAM=2 IGNORE_BEAM_CACHE=1 python3 -m tinygrad.llm -m qwen3.8:27b --benchmark --warmup
    - name: Run olmoe
      # just metal for now
      if: ${{ matrix.dev == 'METAL' }}
      run: BENCHMARK_LOG=olmoe JITBEAM=2 IGNORE_BEAM_CACHE=1 python3 -m tinygrad.llm -m olmoe --benchmark --warmup
-    - name: Run LLaMA-3 8B on 4 GPUs with BEAM
-      # only run on machines with multiple gpus
-      if: ${{ matrix.dev != 'METAL' }}
-      run: BENCHMARK_LOG=llama3_beam_4gpu JITBEAM=2 IGNORE_BEAM_CACHE=1 CAPTURE_PROCESS_REPLAY=0 python3 examples/llama3.py --size 8B --shard 4 --model weights/LLaMA-3/8B-SF-DPO/ --benchmark --temperature 0
    - name: Run process replay tests
      uses: ./.github/actions/process-replay

@@ -143,13 +134,13 @@ jobs:
      fail-fast: false
      matrix:
        dev: ['METAL', 'AMD', 'NV']
-    timeout-minutes: 60
+    timeout-minutes: 10
    defaults:
      run:
        shell: bash -e -o pipefail {0}
    env:
      DEV: ${{ matrix.dev }}
-      HCQ2: ${{ matrix.dev == 'AMD' && '1' || '0' }}
+      HCQ2: '0'
    if: github.repository_owner == 'tinygrad'
    steps:
    - name: Checkout Code
@@ -157,12 +148,11 @@ jobs:
    - name: Setup (AMD)
      if: ${{ matrix.dev == 'AMD' }}
      run: |
-        ./extra/amdpci/setup_python_cap.sh
-        ./extra/hcq/hcq_smi.py amd rmmod
-        ./extra/hcq/hcq_smi.py amd kill_pids
+        ./extra/hcq/hcq_smi.py amd rmmod --expect
+        ./extra/hcq/hcq_smi.py amd kill_pids --sudoless
    - name: Setup (NV)
      if: ${{ matrix.dev == 'NV' }}
-      run: sudo lsof -tQ /dev/nvidia* | { xargs -r sudo kill -9 || true; }
+      run: lsof -tQ /dev/nvidia* | { xargs -r kill -9 || true; }
    - name: setup staging db
      if: github.ref == 'refs/heads/update_benchmark_staging'
      run: |
@@ -182,10 +172,6 @@ jobs:
      # slow on metal
      if: ${{ matrix.dev != 'METAL' }}
      run: time BENCHMARK_LOG=cifar DEFAULT_FLOAT=HALF STEPS=1000 TARGET_EVAL_ACC_PCT=93.0 python3 examples/hlb_cifar10.py
-    - name: Run full CIFAR training steps w 6 GPUS
-      # only run on machines with multiple gpus
-      if: ${{ matrix.dev != 'METAL' }}
-      run: time BENCHMARK_LOG=cifar_6gpu CAPTURE_PROCESS_REPLAY=0 DEFAULT_FLOAT=HALF STEPS=350 BS=1536 GPUS=6 TARGET_EVAL_ACC_PCT=93.0 python3 examples/hlb_cifar10.py
    - name: Run process replay tests
      uses: ./.github/actions/process-replay

@@ -196,13 +182,13 @@ jobs:
      fail-fast: false
      matrix:
        dev: ['AMD', 'NV']
-    timeout-minutes: 60
+    timeout-minutes: 5
    defaults:
      run:
        shell: bash -e -o pipefail {0}
    env:
      DEV: ${{ matrix.dev }}
-      HCQ2: ${{ matrix.dev == 'AMD' && '1' || '0' }}
+      HCQ2: '0'
    if: github.repository_owner == 'tinygrad'
    steps:
    - name: Checkout Code
@@ -210,12 +196,11 @@ jobs:
    - name: Setup (AMD)
      if: ${{ matrix.dev == 'AMD' }}
      run: |
-        ./extra/amdpci/setup_python_cap.sh
-        ./extra/hcq/hcq_smi.py amd rmmod
-        ./extra/hcq/hcq_smi.py amd kill_pids
+        ./extra/hcq/hcq_smi.py amd rmmod --expect
+        ./extra/hcq/hcq_smi.py amd kill_pids --sudoless
    - name: Setup (NV)
      if: ${{ matrix.dev == 'NV' }}
-      run: sudo lsof -tQ /dev/nvidia* | { xargs -r sudo kill -9 || true; }
+      run: lsof -tQ /dev/nvidia* | { xargs -r kill -9 || true; }
    - name: Symlink models and datasets
      run: |
        mkdir -p extra/datasets
@@ -227,15 +212,8 @@ jobs:
        rm -f /tmp/staging.db /tmp/staging.db-shm /tmp/staging.db-wal
    - name: reset process replay
      run: test/external/process_replay/reset.py
-    - name: Run MLPerf resnet eval on training data
-      run: time BENCHMARK_LOG=resnet_eval MODEL=resnet python3 examples/mlperf/model_eval.py
    - name: Run 10 MLPerf ResNet50 training steps (1 gpu)
      run: BENCHMARK_LOG=resnet_10steps DEFAULT_FLOAT=HALF BENCHMARK=10 BS=256 GPUS=1 MODEL=resnet python3 examples/mlperf/model_train.py
-    - name: Run 10 MLPerf ResNet50 training steps (6 gpu)
-      run: BENCHMARK_LOG=resnet_10steps_6gpu CAPTURE_PROCESS_REPLAY=0 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=1536 GPUS=6 MODEL=resnet python3 examples/mlperf/model_train.py
-    - name: Run 10 MLPerf Bert training steps (6 gpu)
-      # TODO: remove BERT_LAYERS once scheduler is fast
-      run: BENCHMARK_LOG=bert_10steps_6gpu CAPTURE_PROCESS_REPLAY=0 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=72 GPUS=6 BERT_LAYERS=2 MODEL=bert python3 examples/mlperf/model_train.py
    - name: Run process replay tests
      uses: ./.github/actions/process-replay

@@ -246,13 +224,13 @@ jobs:
      fail-fast: false
      matrix:
        dev: ['METAL', 'AMD', 'NV']
-    timeout-minutes: 60
+    timeout-minutes: 15
    defaults:
      run:
        shell: bash -e -o pipefail {0}
    env:
      DEV: ${{ matrix.dev }}
-      HCQ2: ${{ matrix.dev == 'AMD' && '1' || '0' }}
+      HCQ2: '0'
    if: github.repository_owner == 'tinygrad'
    steps:
    - name: Checkout Code
@@ -260,12 +238,11 @@ jobs:
    - name: Setup (AMD)
      if: ${{ matrix.dev == 'AMD' }}
      run: |
-        ./extra/amdpci/setup_python_cap.sh
-        ./extra/hcq/hcq_smi.py amd rmmod
-        ./extra/hcq/hcq_smi.py amd kill_pids
+        ./extra/hcq/hcq_smi.py amd rmmod --expect
+        ./extra/hcq/hcq_smi.py amd kill_pids --sudoless
    - name: Setup (NV)
      if: ${{ matrix.dev == 'NV' }}
-      run: sudo lsof -tQ /dev/nvidia* | { xargs -r sudo kill -9 || true; }
+      run: lsof -tQ /dev/nvidia* | { xargs -r kill -9 || true; }
    - name: setup staging db
      if: github.ref == 'refs/heads/update_benchmark_staging'
      run: |
@@ -285,6 +262,59 @@ jobs:
    - name: Run process replay tests
      uses: ./.github/actions/process-replay

+  multigpubenchmark:
+    name: Multi-GPU Benchmarks (DEV=${{ matrix.dev }})
+    runs-on: [self-hosted, "${{ matrix.dev == 'AMD' && 'tinybox' || 'tinyboxgreen' }}"]
+    strategy:
+      fail-fast: false
+      matrix:
+        dev: ['AMD', 'NV']
+    timeout-minutes: 20
+    defaults:
+      run:
+        shell: bash -e -o pipefail {0}
+    env:
+      DEV: ${{ matrix.dev }}
+      HCQ2: '0'
+    if: github.repository_owner == 'tinygrad'
+    steps:
+    - name: Checkout Code
+      uses: actions/checkout@v6
+    - name: Setup (AMD)
+      if: ${{ matrix.dev == 'AMD' }}
+      run: |
+        ./extra/hcq/hcq_smi.py amd rmmod --expect
+        ./extra/hcq/hcq_smi.py amd kill_pids --sudoless
+    - name: Setup (NV)
+      if: ${{ matrix.dev == 'NV' }}
+      run: lsof -tQ /dev/nvidia* | { xargs -r kill -9 || true; }
+    - name: Symlink models and datasets
+      run: |
+        mkdir -p weights
+        mkdir -p extra/datasets
+        ln -s /raid/weights/LLaMA-3 weights/LLaMA-3
+        ln -s /raid/datasets/imagenet extra/datasets/imagenet
+    - name: setup staging db
+      if: github.ref == 'refs/heads/update_benchmark_staging'
+      run: |
+        echo "CACHEDB=/tmp/staging.db" >> $GITHUB_ENV
+        rm -f /tmp/staging.db /tmp/staging.db-shm /tmp/staging.db-wal
+    - name: reset process replay
+      run: python3 test/external/process_replay/reset.py
+    - name: Run LLaMA-3 8B on 4 GPUs with BEAM
+      run: BENCHMARK_LOG=llama3_beam_4gpu JITBEAM=2 IGNORE_BEAM_CACHE=1 CAPTURE_PROCESS_REPLAY=0 python3 examples/llama3.py --size 8B --shard 4 --model weights/LLaMA-3/8B-SF-DPO/ --benchmark --temperature 0
+    - name: Run full CIFAR training steps w 6 GPUS
+      run: time BENCHMARK_LOG=cifar_6gpu CAPTURE_PROCESS_REPLAY=0 DEFAULT_FLOAT=HALF STEPS=350 BS=1536 GPUS=6 TARGET_EVAL_ACC_PCT=93.0 python3 examples/hlb_cifar10.py
+    - name: Run MLPerf resnet eval on training data
+      run: time BENCHMARK_LOG=resnet_eval MODEL=resnet python3 examples/mlperf/model_eval.py
+    - name: Run 10 MLPerf ResNet50 training steps (6 gpu)
+      run: BENCHMARK_LOG=resnet_10steps_6gpu CAPTURE_PROCESS_REPLAY=0 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=1536 GPUS=6 MODEL=resnet python3 examples/mlperf/model_train.py
+    - name: Run 10 MLPerf Bert training steps (6 gpu)
+      # TODO: remove BERT_LAYERS once scheduler is fast
+      run: BENCHMARK_LOG=bert_10steps_6gpu CAPTURE_PROCESS_REPLAY=0 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=72 GPUS=6 BERT_LAYERS=2 MODEL=bert python3 examples/mlperf/model_train.py
+    - name: Run process replay tests
+      uses: ./.github/actions/process-replay
+
  tests:
    name: Tests (DEV=${{ matrix.dev }})
    runs-on: [self-hosted, "${{ matrix.dev == 'METAL' && 'macOS' || matrix.dev == 'AMD' && 'tinybox' || 'tinyboxgreen' }}"]
@@ -292,7 +322,7 @@ jobs:
      fail-fast: false
      matrix:
        dev: ['METAL', 'AMD', 'NV']
-    timeout-minutes: 60
+    timeout-minutes: 10
    defaults:
      run:
        shell: bash -e -o pipefail {0}
@@ -305,12 +335,11 @@ jobs:
    - name: Setup (AMD)
      if: ${{ matrix.dev == 'AMD' }}
      run: |
-        ./extra/amdpci/setup_python_cap.sh
-        ./extra/hcq/hcq_smi.py amd rmmod
-        ./extra/hcq/hcq_smi.py amd kill_pids
+        ./extra/hcq/hcq_smi.py amd rmmod --expect
+        ./extra/hcq/hcq_smi.py amd kill_pids --sudoless
    - name: Setup (NV)
      if: ${{ matrix.dev == 'NV' }}
-      run: sudo lsof -tQ /dev/nvidia* | { xargs -r sudo kill -9 || true; }
+      run: lsof -tQ /dev/nvidia* | { xargs -r kill -9 || true; }
    - name: setup staging db
      if: github.ref == 'refs/heads/update_benchmark_staging'
      run: |
@@ -374,7 +403,7 @@ jobs:
      run: python test/external/external_benchmark_multitensor_allreduce.py
    - name: HEVC Decode Benchmark
      if: ${{ matrix.dev == 'NV' }}
-      run: VALIDATE=1 MAX_FRAMES=100 ASSERT_FPS=1400 JITBEAM=1 PYTHONPATH=. python3 extra/hevc/decode.py
+      run: IGNORE_BEAM_CACHE=1 VALIDATE=1 MAX_FRAMES=100 ASSERT_FPS=1400 JITBEAM=1 PYTHONPATH=. python3 extra/hevc/decode.py
    - uses: actions/upload-artifact@v7
      if: ${{ matrix.dev != 'AMD' }}
      with:
@@ -387,7 +416,7 @@ jobs:
  testusbgpu:
    name: UsbGPU Benchmark
    runs-on: [self-hosted, macOS]
-    timeout-minutes: 10
+    timeout-minutes: 3
    defaults:
      run:
        shell: bash -e -o pipefail {0}
@@ -402,32 +431,66 @@ jobs:
        rm -f /tmp/staging.db /tmp/staging.db-shm /tmp/staging.db-wal
    - name: Kill stale pids
      run: |
-        PYTHONPATH=. ./extra/hcq/hcq_smi.py amd kill_pids
-        PYTHONPATH=. ./extra/hcq/hcq_smi.py nv kill_pids
-    # since sudo is required for usbgpu on macos, do not write bytecode, as some of the files are owned by root
+        ./extra/hcq/hcq_smi.py amd kill_pids --sudoless
+        ./extra/hcq/hcq_smi.py nv kill_pids --sudoless
    - name: UsbGPU boot time
-      run: sudo -E PYTHONDONTWRITEBYTECODE=1 PYTHONPATH=. GMMU=0 DEBUG=2 AM_RESET=1 DEV=USB+AMD time python3.11 test/test_tiny.py TestTiny.test_plus
+      run: GMMU=0 DEBUG=2 AM_RESET=1 DEV=USB+AMD time python3.11 test/test_tiny.py TestTiny.test_plus
    - name: UsbGPU tiny tests
-      run: sudo -E PYTHONDONTWRITEBYTECODE=1 PYTHONPATH=. GMMU=0 DEV=USB+AMD python3.11 test/test_tiny.py
+      run: GMMU=0 DEV=USB+AMD python3.11 test/test_tiny.py
    - name: UsbGPU copy speeds
-      run: sudo -E PYTHONDONTWRITEBYTECODE=1 PYTHONPATH=. GMMU=0 DEV=USB+AMD python3.11 test/external/external_test_usb_asm24.py TestDevCopySpeeds
-    #- name: UsbGPU openpilot test
-    #  run: sudo -E PYTHONPATH=. GMMU=0 DEV=USB+AMD GRAPH_ONE_KERNEL=1 python3.11 examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/9118973ed03c1ae1d40cf69a29507ec2cc78efd7/selfdrive/modeld/models/supercombo.onnx
+      run: SIZE=64000000 PYTHONPATH=. GMMU=0 DEV=USB+AMD python3.11 test/external/external_test_usb_asm24.py TestDevCopySpeeds
    - name: UsbGPU (USB4/TB) install script
-      run: PYTHONPATH=. sh extra/setup_tinygpu_osx.sh
+      run: sh extra/setup_tinygpu_osx.sh
    - name: UsbGPU (USB4/TB) boot time
-      run: PYTHONPATH=. DEBUG=3 DEV=PCI+NV:NAK time python3.11 test/test_tiny.py TestTiny.test_plus
+      run: DEBUG=3 DEV=PCI+NV:NAK time python3.11 test/test_tiny.py TestTiny.test_plus
    - name: UsbGPU (USB4/TB) tiny tests
-      run: PYTHONPATH=. DEV=PCI+NV:NAK python3.11 test/test_tiny.py
+      run: DEV=PCI+NV:NAK python3.11 test/test_tiny.py

-  testcommalatest:
-    name: comma Benchmark (0.11.2)
-    runs-on: [self-hosted, Linux, comma]
-    timeout-minutes: 12
+  testcomma:
+    strategy:
+      matrix:
+        dev: ['QCOM', 'QCOM:IR3']
+        version: ['0.11.0', '0.11.2']
+        model: ['vision', 'policy', 'supercombo', 'dmonitoring']
+        # exclude non-existent models
+        exclude: [{ version: '0.11.0', model: supercombo }, { version: '0.11.2', model: vision }, { version: '0.11.2', model: policy }]
+        include:
+          - version: '0.11.0'
+            model: vision
+            url: https://github.com/commaai/openpilot/raw/v0.11.0/selfdrive/modeld/models/driving_vision.onnx
+            timing: 18
+          - version: '0.11.0'
+            model: policy
+            url: https://github.com/commaai/openpilot/raw/v0.11.0/selfdrive/modeld/models/driving_policy.onnx
+            timing: 3.4
+          - version: '0.11.0'
+            model: dmonitoring
+            url: https://github.com/commaai/openpilot/raw/v0.11.0/selfdrive/modeld/models/dmonitoring_model.onnx
+            timing: 13
+          - version: '0.11.2'
+            model: supercombo
+            url: https://gitlab.com/commaai/openpilot-lfs.git/gitlab-lfs/objects/433f85f956837606ad1f1cbee4aa7e2158ad23c768dea914b20436c97232741b
+            timing: 28
+          - dev: QCOM:IR3
+            version: '0.11.2'
+            model: supercombo
+            timing: 29
+          - version: '0.11.2'
+            model: dmonitoring
+            url: https://gitlab.com/commaai/openpilot-lfs.git/gitlab-lfs/objects/3e7b31dfbc0a5234f1baf196513b77fc6af12204b8a8ffe8ee0417e48352f316
+            timing: 12.5
+      fail-fast: false
+    name: openpilot ${{ matrix.version }} compile3 ${{ matrix.model }} (DEV=${{ matrix.dev }})
+    runs-on: [self-hosted, Linux, comma4]
+    timeout-minutes: 5
    defaults:
      run:
        shell: bash -e -o pipefail {0}
    if: github.repository_owner == 'tinygrad'
+    env:
+      DEV: ${{ matrix.dev }}
+      ASSERT_MIN_STEP_TIME: ${{ matrix.timing }}
+      BENCHMARK_LOG: ${{ matrix.dev == 'QCOM:IR3' && 'ir3_' || '' }}openpilot_${{ matrix.version }}_${{ matrix.model }}
    steps:
    - name: Checkout Code
      uses: actions/checkout@v6
@@ -438,45 +501,10 @@ jobs:
        rm -f /tmp/staging.db /tmp/staging.db-shm /tmp/staging.db-wal
    - name: reset process replay
      run: test/external/process_replay/reset.py
-    - name: openpilot compile3 0.11.2 supercombo
-      run: BENCHMARK_LOG=openpilot_0_11_2_supercombo PYTHONPATH="." ASSERT_MIN_STEP_TIME=26 DEV=QCOM FLOAT16=1 IMAGE=1 NOLOCALS=1 taskset -c 4-7 python3 examples/openpilot/compile3.py https://gitlab.com/commaai/openpilot-lfs.git/gitlab-lfs/objects/433f85f956837606ad1f1cbee4aa7e2158ad23c768dea914b20436c97232741b
-    - name: openpilot compile3 0.11.2 supercombo (from pickle)
-      run: BENCHMARK_LOG=openpilot_0_11_2_supercombo_run_pickle RUN_PICKLE=1 PYTHONPATH="." ASSERT_MIN_STEP_TIME=26 DEV=QCOM taskset -c 4-7 python3 examples/openpilot/compile3.py
-    - name: IR3 openpilot compile3 0.11.2 supercombo
-      run: BENCHMARK_LOG=ir3_openpilot_0_11_2_supercombo PYTHONPATH="." ASSERT_MIN_STEP_TIME=41 DEV=QCOM:IR3 FLOAT16=1 IMAGE=1 NOLOCALS=1 taskset -c 4-7 python3 examples/openpilot/compile3.py https://gitlab.com/commaai/openpilot-lfs.git/gitlab-lfs/objects/433f85f956837606ad1f1cbee4aa7e2158ad23c768dea914b20436c97232741b
-    - name: openpilot compile3 0.11.2 dmonitoring
-      run: BENCHMARK_LOG=openpilot_0_11_2_dmonitoring PYTHONPATH="." ASSERT_MIN_STEP_TIME=11 DEV=QCOM FLOAT16=1 IMAGE=1 NOLOCALS=1 taskset -c 4-7 python3 examples/openpilot/compile3.py https://gitlab.com/commaai/openpilot-lfs.git/gitlab-lfs/objects/3e7b31dfbc0a5234f1baf196513b77fc6af12204b8a8ffe8ee0417e48352f316
-    - name: Run process replay tests
-      uses: ./.github/actions/process-replay
-
-  testcommaold:
-    name: comma Benchmark (0.11.0)
-    runs-on: [self-hosted, Linux, comma]
-    timeout-minutes: 12
-    defaults:
-      run:
-        shell: bash -e -o pipefail {0}
-    if: github.repository_owner == 'tinygrad'
-    steps:
-    - name: Checkout Code
-      uses: actions/checkout@v6
-    - name: setup staging db
-      if: github.ref == 'refs/heads/update_benchmark_staging'
-      run: |
-        echo "CACHEDB=/tmp/staging.db" >> $GITHUB_ENV
-        rm -f /tmp/staging.db /tmp/staging.db-shm /tmp/staging.db-wal
-    - name: reset process replay
-      run: test/external/process_replay/reset.py
-    - name: openpilot compile3 0.11.0 driving_vision
-      run: BENCHMARK_LOG=openpilot_0_11_0_vision PYTHONPATH="." ASSERT_MIN_STEP_TIME=17 DEV=QCOM FLOAT16=1 IMAGE=1 NOLOCALS=1 taskset -c 4-7 python3 examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/v0.11.0/selfdrive/modeld/models/driving_vision.onnx
-    - name: openpilot compile3 0.11.0 driving_vision (from pickle)
-      run: BENCHMARK_LOG=openpilot_0_11_0_vision_run_pickle RUN_PICKLE=1 PYTHONPATH="." ASSERT_MIN_STEP_TIME=17 DEV=QCOM taskset -c 4-7 python3 examples/openpilot/compile3.py
-    - name: IR3 openpilot compile3 0.11.0 driving_vision
-      run: BENCHMARK_LOG=ir3_openpilot_0_11_0_vision PYTHONPATH="." ASSERT_MIN_STEP_TIME=18 DEV=QCOM:IR3 FLOAT16=1 IMAGE=1 NOLOCALS=1 taskset -c 4-7 python3 examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/v0.11.0/selfdrive/modeld/models/driving_vision.onnx
-    - name: openpilot compile3 0.11.0 driving_policy
-      run: BENCHMARK_LOG=openpilot_0_11_0_policy PYTHONPATH="." ASSERT_MIN_STEP_TIME=3.2 DEV=QCOM FLOAT16=1 IMAGE=1 NOLOCALS=1 taskset -c 4-7 python3 examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/v0.11.0/selfdrive/modeld/models/driving_policy.onnx
-    - name: openpilot compile3 0.11.0 dmonitoring
-      run: BENCHMARK_LOG=openpilot_0_11_0_dmonitoring PYTHONPATH="." ASSERT_MIN_STEP_TIME=11 DEV=QCOM FLOAT16=1 IMAGE=1 NOLOCALS=1 taskset -c 4-7 python3 examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/v0.11.0/selfdrive/modeld/models/dmonitoring_model.onnx
+    - name: compile
+      run: FLOAT16=1 IMAGE=1 taskset -c 4-7 python3 examples/openpilot/compile3.py ${{ matrix.url }} openpilot.pkl
+    - name: run pickle
+      run: BENCHMARK_LOG="${BENCHMARK_LOG}_run_pickle" RUN_PICKLE=1 taskset -c 4-7 python3 examples/openpilot/compile3.py - openpilot.pkl
    - name: Run process replay tests
      uses: ./.github/actions/process-replay

@@ -489,15 +517,6 @@ jobs:
        shell: bash -e -o pipefail {0}
    if: github.repository_owner == 'tinygrad'
    steps:
-    - name: Checkout Code
-      uses: actions/checkout@v6
-    - name: setup staging db
-      if: github.ref == 'refs/heads/update_benchmark_staging'
-      run: |
-        echo "CACHEDB=/tmp/staging.db" >> $GITHUB_ENV
-        rm -f /tmp/staging.db /tmp/staging.db-shm /tmp/staging.db-wal
-    - name: reset process replay
-      run: test/external/process_replay/reset.py
    - name: Checkout Code
      uses: actions/checkout@v6
    - name: setup staging db
@@ -510,8 +529,8 @@ jobs:
    - name: benchmark MobileNetV2 on DSP
      run: |
        # generate quantized weights
-        ln -s /data/home/tiny/tinygrad/extra/datasets/imagenet extra/datasets/imagenet
-        ln -s /data/home/tiny/tinygrad/testsig-*.so .
+        ln -s ~/tinygrad/extra/datasets/imagenet extra/datasets/imagenet
+        ln -s ~/tinygrad/testsig-*.so .
        PYTHONPATH=. DEV=CPU QUANT=1 CNT=0 python3 examples/test_onnx_imagenet.py https://github.com/xamcat/mobcat-samples/raw/refs/heads/master/onnx_runtime/InferencingSample/InferencingSample/mobilenetv2-7.onnx /tmp/model.quant.onnx
        # benchmark on DSP with NOOPT=1, the devectorizer has issues
        PYTHONPATH=. DEV=DSP NOOPT=1 CNT=2 DEBUG=2 python3 examples/test_onnx_imagenet.py /tmp/model.quant.onnx
@@ -521,7 +540,7 @@ jobs:
  testcommausbgpubenchmark:
    name: UsbGPU Benchmark (comma)
    runs-on: [self-hosted, Linux, comma4]
-    timeout-minutes: 20
+    timeout-minutes: 10
    defaults:
      run:
        shell: bash -e -o pipefail {0}
@@ -541,7 +560,7 @@ jobs:
    - name: openpilot run_pickle big_driving_supercombo
      run: BENCHMARK_LOG=usbgpu_openpilot_big_driving_supercombo_run_pickle RUN_PICKLE=1 PICKLE_OOB=1 PYTHONPATH="." GMMU=0 DEV=USB+AMD ASSERT_MIN_STEP_TIME=50 python3 examples/openpilot/compile3.py - openpilot.pkl
    - name: Test copy speeds
-      run: SIZE=64e6 PYTHONPATH=. GMMU=0 DEV=USB+AMD python3 test/external/external_test_usb_asm24.py TestDevCopySpeeds
+      run: SIZE=64000000 PYTHONPATH=. GMMU=0 DEV=USB+AMD python3 test/external/external_test_usb_asm24.py TestDevCopySpeeds

  driverbenchmarks:
    name: PCI Driver Benchmark (DEV=${{ matrix.dev }})
@@ -550,7 +569,7 @@ jobs:
      fail-fast: false
      matrix:
        dev: ['AMD', 'NV']
-    timeout-minutes: 20
+    timeout-minutes: 5
    defaults:
      run:
        shell: bash -e -o pipefail {0}
@@ -562,9 +581,8 @@ jobs:
      uses: actions/checkout@v6
    - name: Setup
      run: |
-        ./extra/amdpci/setup_python_cap.sh
-        ./extra/hcq/hcq_smi.py ${{ matrix.dev == 'AMD' && 'amd' || 'nv' }} rmmod
-        ./extra/hcq/hcq_smi.py ${{ matrix.dev == 'AMD' && 'amd' || 'nv' }} kill_pids
+        ./extra/hcq/hcq_smi.py ${{ matrix.dev }} rmmod --expect
+        ./extra/hcq/hcq_smi.py ${{ matrix.dev }} kill_pids --sudoless
        mkdir -p extra/datasets
        ln -s /raid/datasets/imagenet extra/datasets/imagenet
    - name: setup staging db
@@ -599,6 +617,9 @@ jobs:
      run: |
        GRAPH_ONE_KERNEL=1 NSZ=8192 python3 test/speed/external_test_copy_speed.py TestCopySpeed.testCopyDefaulttoCPUJit
        GRAPH_ONE_KERNEL=1 NSZ=8192 python3 test/speed/external_test_copy_speed.py TestCopySpeed.testCopyCPUtoDefaultJit
+    - name: HEVC Decode Benchmark
+      if: ${{ matrix.dev == 'NV' }}
+      run: IGNORE_BEAM_CACHE=1 VALIDATE=1 MAX_FRAMES=100 ASSERT_FPS=1400 JITBEAM=1 PYTHONPATH=. python3 extra/hevc/decode.py
    - name: Run 10 MLPerf ResNet50 training steps (1 gpu)
      if: ${{ matrix.dev == 'NV' }}
      run: BENCHMARK_LOG=resnet_10steps MNISTMOCK=1 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=256 GPUS=1 MODEL=resnet python3 examples/mlperf/model_train.py
@@ -621,12 +642,12 @@ jobs:
  llvmspeed:
    name: LLVM Speed
    runs-on: [self-hosted, Linux, tinyboxrandom]
-    timeout-minutes: 20
+    timeout-minutes: 10
    if: github.repository_owner == 'tinygrad'
    steps:
    - name: Checkout Code
      uses: actions/checkout@v6
    - name: Speed Test
-      run: DEV=CPU:LLVM THREADS=0 python3 test/speed/external_test_speed_v_torch.py
+      run: DEV=CPU:LLVM python3 test/speed/external_test_speed_v_torch.py
    - name: Speed Test (BEAM=2)
-      run: BEAM=2 DEV=CPU:LLVM THREADS=0 python3 test/speed/external_test_speed_v_torch.py
+      run: IGNORE_BEAM_CACHE=1 BEAM=2 DEV=CPU:LLVM python3 test/speed/external_test_speed_v_torch.py
@@ -8,7 +8,7 @@ permissions:
  contents: write
 jobs:
  deploy:
-    runs-on: ubuntu-latest
+    runs-on: ubuntu-24.04
    steps:
      - uses: actions/checkout@v6
      - name: Configure Git Credentials
@@ -166,7 +166,7 @@ jobs:
        uses: ./.github/actions/setup-tinygrad
        with:
          key: windows-${{ matrix.dev }}-minimal
-          deps: testing_unit
+          deps: testing_minimal
          pydeps: ${{ matrix.dev == 'WEBGPU' && 'dawn-python' || '' }}
      - name: Set env
        shell: bash
@@ -179,35 +179,3 @@ jobs:
      - name: Run test_tiny
        shell: bash
        run: python -m pytest -n=auto test/test_tiny.py --durations=20
-
-
-  qcomclcompiletests:
-    name: Compile-only (QCOM CL)
-    runs-on: ubuntu-24.04-arm
-    timeout-minutes: 15
-    steps:
-      - name: Checkout Code
-        uses: actions/checkout@v6
-      - name: Setup Environment
-        uses: ./.github/actions/setup-tinygrad
-        with:
-          key: compile-qcomcl
-          deps: testing_unit
-          tinydreno: 'true'
-      - name: Set env
-        shell: bash
-        run: printf "DEV=NULL:QCOMCL:a630\nNULL_ALLOW_COPYOUT=1" >> $GITHUB_ENV
-      - name: Run test_ops
-        shell: bash
-        run: |
-          python -c "from tinygrad import Device; assert Device.DEFAULT == 'NULL'"
-          DEBUG=4 python3 test/backend/test_ops.py TestOps.test_add
-          python -m pytest -n=auto test/backend/test_ops.py --durations=20
-      - name: Run test_ops (IMAGE)
-        shell: bash
-        env:
-          IMAGE: 1
-          DEV: "NULL:QCOMCL:a630,IMAGE_PITCH_ALIGNMENT=64"
-        run: |
-          DEBUG=4 python test/backend/test_ops.py TestOps.test_gemm | grep read_imagef
-          python -m pytest -n=auto test/backend/test_ops.py --durations=20
@@ -10,7 +10,7 @@ on:

 jobs:
  deploy:
-    runs-on: ubuntu-latest
+    runs-on: ubuntu-24.04
    steps:
    - uses: actions/checkout@v6
    - name: Set up Python
@@ -10,7 +10,7 @@ concurrency:
 jobs:
  checkbranch:
    name: Check PR Branch status
-    runs-on: ubuntu-latest
+    runs-on: ubuntu-24.04
    outputs:
      branchstat: ${{ steps.brstat.outputs.stat}}
    steps:
@@ -44,7 +44,7 @@ jobs:
    permissions:
      contents: read
      pull-requests: write
-    runs-on: ubuntu-latest
+    runs-on: ubuntu-24.04
    needs: checkbranch
    if: needs.checkbranch.outputs.branchstat == 'false'
    steps:
@@ -87,7 +87,7 @@ jobs:
    name: Core Library Line Difference
    permissions:
      pull-requests: write
-    runs-on: ubuntu-latest
+    runs-on: ubuntu-24.04
    needs: checkbranch
    if: needs.checkbranch.outputs.branchstat == 'true'
    steps:
@@ -21,7 +21,7 @@ concurrency:
 jobs:
  docs:
    name: Docs
-    runs-on: &linux ${{ github.repository == 'tinygrad/tinygrad' && github.event_name == 'pull_request' && github.event.pull_request.author_association == 'COLLABORATOR' && 'namespace-profile-tinygrad' || 'ubuntu-24.04' }}
+    runs-on: ${{ github.repository == 'tinygrad/tinygrad' && github.event_name == 'pull_request' && github.event.pull_request.author_association == 'COLLABORATOR' && 'namespace-profile-tinygrad' || 'ubuntu-24.04' }}
    timeout-minutes: 10
    env:
      CHECK_OOB: 0
@@ -31,8 +31,7 @@ jobs:
    - name: Setup Environment
      uses: ./.github/actions/setup-tinygrad
      with:
-        deps: docs
-        pydeps: "capstone torch"
+        deps: "docs testing_minimal"
    - name: Build wheel and show size
      run: |
        uv build --wheel
@@ -61,7 +60,7 @@ jobs:

  torchbackend:
    name: Torch Backend Tests
-    runs-on: *linux
+    runs-on: ${{ github.repository == 'tinygrad/tinygrad' && github.event_name == 'pull_request' && github.event.pull_request.author_association == 'COLLABORATOR' && 'namespace-profile-tinygrad' || 'ubuntu-24.04' }}
    timeout-minutes: 15
    steps:
    - name: Checkout Code
@@ -73,10 +72,7 @@ jobs:
        deps: testing_unit
        pydeps: "pillow torchvision expecttest"
        llvm: 'true'
-    - name: Install ninja
-      run: |
-        sudo apt update || true
-        sudo apt install -y --no-install-recommends ninja-build
+        ninja: 'true'
    - name: Test ResNet-18
      run: DEBUG=2 python3 extra/torch_backend/example.py
    - name: Test one op in torch tests
@@ -86,9 +82,26 @@ jobs:
    - name: Custom tests
      run: DEV=CPU:LLVM GPUS=4 TINY_BACKEND=1 python3 -m pytest -nauto extra/torch_backend/test.py extra/torch_backend/test_inplace.py extra/torch_backend/test_multigpu.py extra/torch_backend/test_kernel_fusion.py --durations=20

+  torchbackendtrain:
+    name: Torch Backend Training
+    runs-on: ${{ github.repository == 'tinygrad/tinygrad' && github.event_name == 'pull_request' && github.event.pull_request.author_association == 'COLLABORATOR' && 'namespace-profile-tinygrad' || 'ubuntu-24.04' }}
+    timeout-minutes: 15
+    steps:
+    - name: Checkout Code
+      uses: actions/checkout@v6
+    - name: Setup Environment
+      uses: ./.github/actions/setup-tinygrad
+      with:
+        key: torch-backend-pillow-torchvision-et-pt
+        deps: testing_unit
+        llvm: 'true'
+        ninja: 'true'
+    - name: Test beautiful_mnist in torch with TINY_BACKEND
+      run: STEPS=20 DEV=CPU TARGET_EVAL_ACC_PCT=90.0 MAX_BUFFER_SIZE=0 TINY_BACKEND=1 python3 examples/other_mnist/beautiful_mnist_torch.py
+
  bepython:
    name: Python Backend
-    runs-on: *linux
+    runs-on: ${{ github.repository == 'tinygrad/tinygrad' && github.event_name == 'pull_request' && github.event.pull_request.author_association == 'COLLABORATOR' && 'namespace-profile-tinygrad' || 'ubuntu-24.04' }}
    timeout-minutes: 15
    steps:
    - name: Checkout Code
@@ -126,7 +139,7 @@ jobs:

  linter:
    name: Linters
-    runs-on: *linux
+    runs-on: ${{ github.repository == 'tinygrad/tinygrad' && github.event_name == 'pull_request' && github.event.pull_request.author_association == 'COLLABORATOR' && 'namespace-profile-tinygrad' || 'ubuntu-24.04' }}
    timeout-minutes: 10

    steps:
@@ -157,7 +170,7 @@ jobs:

  nulltest:
    name: Null Tests
-    runs-on: *linux
+    runs-on: ${{ github.repository == 'tinygrad/tinygrad' && github.event_name == 'pull_request' && github.event.pull_request.author_association == 'COLLABORATOR' && 'namespace-profile-tinygrad' || 'ubuntu-24.04' }}
    timeout-minutes: 15

    steps:
@@ -191,7 +204,7 @@ jobs:

  unittest:
    name: Unit Tests
-    runs-on: *linux
+    runs-on: ${{ github.repository == 'tinygrad/tinygrad' && github.event_name == 'pull_request' && github.event.pull_request.author_association == 'COLLABORATOR' && 'namespace-profile-tinygrad' || 'ubuntu-24.04' }}
    timeout-minutes: 15

    steps:
@@ -220,7 +233,7 @@ jobs:
    - name: Run process replay tests
      uses: ./.github/actions/process-replay
    - name: Repo line count <= 26000 lines
-      run: MAX_LINE_COUNT=26000 python sz.py
+      run: MAX_LINE_COUNT=26500 python sz.py

  spec:
    strategy:
@@ -228,7 +241,7 @@ jobs:
      matrix:
        group: [1, 2]
    name: SPEC=2 (${{ matrix.group }})
-    runs-on: *linux
+    runs-on: ${{ github.repository == 'tinygrad/tinygrad' && github.event_name == 'pull_request' && github.event.pull_request.author_association == 'COLLABORATOR' && 'namespace-profile-tinygrad' || 'ubuntu-24.04' }}
    timeout-minutes: 15
    steps:
    - name: Checkout Code
@@ -244,7 +257,7 @@ jobs:

  fuzzing:
    name: Fuzzing
-    runs-on: *linux
+    runs-on: ${{ github.repository == 'tinygrad/tinygrad' && github.event_name == 'pull_request' && github.event.pull_request.author_association == 'COLLABORATOR' && 'namespace-profile-tinygrad' || 'ubuntu-24.04' }}
    timeout-minutes: 10
    steps:
    - name: Checkout Code
@@ -260,7 +273,7 @@ jobs:

  testopenclimage:
    name: CL IMAGE Tests
-    runs-on: *linux
+    runs-on: ${{ github.repository == 'tinygrad/tinygrad' && github.event_name == 'pull_request' && github.event.pull_request.author_association == 'COLLABORATOR' && 'namespace-profile-tinygrad' || 'ubuntu-24.04' }}
    timeout-minutes: 15
    steps:
      - name: Checkout Code
@@ -280,7 +293,7 @@ jobs:

  testopenpilot:
    name: openpilot Compile Tests
-    runs-on: *linux
+    runs-on: ${{ github.repository == 'tinygrad/tinygrad' && github.event_name == 'pull_request' && github.event.pull_request.author_association == 'COLLABORATOR' && 'namespace-profile-tinygrad' || 'ubuntu-24.04' }}
    timeout-minutes: 15
    steps:
      - name: Checkout Code
@@ -309,7 +322,7 @@ jobs:

  testonnxcpu:
    name: ONNX (CPU) Tests
-    runs-on: *linux
+    runs-on: ${{ github.repository == 'tinygrad/tinygrad' && github.event_name == 'pull_request' && github.event.pull_request.author_association == 'COLLABORATOR' && 'namespace-profile-tinygrad' || 'ubuntu-24.04' }}
    timeout-minutes: 20

    steps:
@@ -328,7 +341,7 @@ jobs:

  testoptim:
    name: Optimization Tests
-    runs-on: *linux
+    runs-on: ${{ github.repository == 'tinygrad/tinygrad' && github.event_name == 'pull_request' && github.event.pull_request.author_association == 'COLLABORATOR' && 'namespace-profile-tinygrad' || 'ubuntu-24.04' }}
    timeout-minutes: 20
    steps:
      - name: Checkout Code
@@ -360,7 +373,7 @@ jobs:

  testllm:
    name: Test LLM
-    runs-on: *linux
+    runs-on: ${{ github.repository == 'tinygrad/tinygrad' && github.event_name == 'pull_request' && github.event.pull_request.author_association == 'COLLABORATOR' && 'namespace-profile-tinygrad' || 'ubuntu-24.04' }}
    timeout-minutes: 15
    env:
      CHECK_OOB: 0
@@ -377,17 +390,17 @@ jobs:
        run: |
          parallel --link --tagstring '[{1}]' '{2}' \
            ::: llama 'llama q4' qwen3.5 qwen \
-            ::: $'echo "What\'s a male chicken called? Answer with only one word." | python3 -m tinygrad.llm --model llama3.2:1b | tee /dev/stderr | grep -i rooster' \
-                $'echo "What\'s a male chicken called? Answer with only one word." | python3 -m tinygrad.llm --model llama3.2:1b-q4 | tee /dev/stderr | grep -i rooster' \
-                $'echo "What\'s a male chicken called? Answer with only one word." | python3 -m tinygrad.llm --model qwen3.5:0.8b | tee /dev/stderr | grep -i rooster' \
-                $'echo "What\'s a female chicken called? Answer with only one word." | python3 -m tinygrad.llm --model qwen3:0.6b | tee /dev/stderr | grep -i hen'
+            ::: $'echo "What\'s a male chicken called? Answer with only one word." | python3 -m tinygrad.llm --no_chat_template --model llama3.2:1b | tee /dev/stderr | grep -i rooster' \
+                $'echo "What\'s a male chicken called? Answer with only one word." | python3 -m tinygrad.llm --no_chat_template --model llama3.2:1b-q4 | tee /dev/stderr | grep -i rooster' \
+                $'echo "What\'s a male chicken called? Answer with only one word." | python3 -m tinygrad.llm --no_chat_template --model qwen3.5:0.8b | tee /dev/stderr | grep -i rooster' \
+                $'echo "What\'s a female chicken called? Answer with only one word." | python3 -m tinygrad.llm --no_chat_template --model qwen3:0.6b | tee /dev/stderr | grep -i hen'
                # NOTE: qwen is dumb and only knows about female chickens

 # ****** Models Tests ******

  testmodels:
    name: Models
-    runs-on: *linux
+    runs-on: ${{ github.repository == 'tinygrad/tinygrad' && github.event_name == 'pull_request' && github.event.pull_request.author_association == 'COLLABORATOR' && 'namespace-profile-tinygrad' || 'ubuntu-24.04' }}
    timeout-minutes: 15
    steps:
      - name: Checkout Code
@@ -407,7 +420,7 @@ jobs:

  testdsp:
    name: Linux (DSP)
-    runs-on: *linux
+    runs-on: ${{ github.repository == 'tinygrad/tinygrad' && github.event_name == 'pull_request' && github.event.pull_request.author_association == 'COLLABORATOR' && 'namespace-profile-tinygrad' || 'ubuntu-24.04' }}
    timeout-minutes: 15
    steps:
    - name: Checkout Code
@@ -435,7 +448,7 @@ jobs:
          - 'WEBGPU'

    name: Linux (DEV=${{ matrix.dev }})
-    runs-on: *linux
+    runs-on: ${{ github.repository == 'tinygrad/tinygrad' && github.event_name == 'pull_request' && github.event.pull_request.author_association == 'COLLABORATOR' && 'namespace-profile-tinygrad' || 'ubuntu-24.04' }}
    timeout-minutes: 20
    steps:
    - name: Checkout Code
@@ -461,7 +474,7 @@ jobs:

  testamdasm:
    name: AMD ASM IDE
-    runs-on: *linux
+    runs-on: ${{ github.repository == 'tinygrad/tinygrad' && github.event_name == 'pull_request' && github.event.pull_request.author_association == 'COLLABORATOR' && 'namespace-profile-tinygrad' || 'ubuntu-24.04' }}
    timeout-minutes: 20
    env:
      DEV: MOCKKFD+AMD
@@ -491,7 +504,7 @@ jobs:
      - name: Run AMD renderer tests (AMD:LLVM)
        run: DEV=MOCKKFD+AMD:LLVM python -m pytest -n=auto test/amd/ --durations 20
      - name: Run SQTT profiling tests
-        run: PROFILE=1 SQTT=1 python3 -m pytest -n=auto test/amd/test_sqtt_profiler.py
+        run: VIZ=-2 python3 -m pytest -n=auto test/amd/test_sqtt_profiler.py
      - name: Run AMD emulated tests on NULL backend
        env:
          AMD: 0
@@ -507,7 +520,7 @@ jobs:

  hcq2:
    name: hcq2
-    runs-on: *linux
+    runs-on: ${{ github.repository == 'tinygrad/tinygrad' && github.event_name == 'pull_request' && github.event.pull_request.author_association == 'COLLABORATOR' && 'namespace-profile-tinygrad' || 'ubuntu-24.04' }}
    timeout-minutes: 5
    steps:
      - name: Checkout Code
@@ -521,16 +534,15 @@ jobs:
      - name: Run HCQ2 tests
        run: HCQ_RUNTIME_DEV=PYTHON HCQ2=1 DEV=MOCKKFD+AMD FORWARD_ONLY=1 PYTHONPATH=. python test/test_tiny.py
      - name: Run HCQ2 multi-device tests
-        run: |
-          HCQ_RUNTIME_DEV=PYTHON HCQ2=1 DEV=MOCKKFD+AMD FORWARD_ONLY=1 PYTHONPATH=. python test/unit/test_multitensor.py \
-            TestMultiTensor.test_simple_add TestMultiTensor.test_shard_reduce \
-            TestMultiTensor.test_backward_sum TestMultiTensor.test_matmul_shard_0_0
+        run: HCQ_RUNTIME_DEV=PYTHON HCQ2=1 DEV=MOCKKFD+AMD FORWARD_ONLY=1 PYTHONPATH=. python -m pytest -n=auto test/backend/test_multitensor.py
      - name: Run HCQ2 JIT tests
        run: HCQ_RUNTIME_DEV=PYTHON HCQ2=1 DEV=MOCKKFD+AMD FORWARD_ONLY=1 PYTHONPATH=. python test/unit/test_jit.py
+      - name: Run HCQ2 unit tests
+        run: HCQ_RUNTIME_DEV=PYTHON HCQ2=1 DEV=MOCKKFD+AMD FORWARD_ONLY=1 PYTHONPATH=. python -m pytest test/device/test_hcq2.py

  testmockam:
    name: Linux (am)
-    runs-on: *linux
+    runs-on: ${{ github.repository == 'tinygrad/tinygrad' && github.event_name == 'pull_request' && github.event.pull_request.author_association == 'COLLABORATOR' && 'namespace-profile-tinygrad' || 'ubuntu-24.04' }}
    timeout-minutes: 15
    env:
      DEV: MOCKPCI+AMD
@@ -566,7 +578,7 @@ jobs:
        arch: [gfx1100, gfx1201, gfx950]

    name: Linux (${{ matrix.backend }} ${{ matrix.arch }})
-    runs-on: *linux
+    runs-on: ${{ github.repository == 'tinygrad/tinygrad' && github.event_name == 'pull_request' && github.event.pull_request.author_association == 'COLLABORATOR' && 'namespace-profile-tinygrad' || 'ubuntu-24.04' }}
    timeout-minutes: 15
    env:
      DEV: MOCKKFD+AMD:${{ matrix.backend == 'amdllvm' && 'LLVM' || '' }}:${{ matrix.arch }}
@@ -589,7 +601,7 @@ jobs:
        if: ${{ matrix.backend == 'amd' && matrix.arch == 'gfx950' }}
        run: PYTHONPATH=. DEV=NULL:HIP:gfx950 MXFP4=1 LLAMA_LAYERS=2 BENCHMARK=3 NULL_ALLOW_COPYOUT=1 NO_HIPCC=1 ROCM_PATH=/opt/rocm JITBEAM=0 examples/mlperf/training_submission_v6.0/tinycorp/benchmarks/llama31_8b/implementations/tinybox_8xMI350X/profile.sh
      - name: Run pytest (amd)
-        run: python -m pytest -n=auto test/backend/test_ops.py test/backend/test_dtype.py test/backend/test_dtype_alu.py test/backend/test_linearizer.py test/backend/test_randomness.py test/backend/test_jit.py test/backend/test_graph.py test/backend/test_multitensor.py test/device/test_hcq.py test/external/external_test_am.py test/backend/test_asm_gemm.py::TestAsmGEMM test/opt/test_tensor_cores.py --durations=20
+        run: python -m pytest -n=auto test/backend/test_ops.py test/backend/test_dtype.py test/backend/test_dtype_alu.py test/backend/test_linearizer.py test/backend/test_randomness.py test/backend/test_jit.py test/backend/test_graph.py test/backend/test_multitensor.py test/device/test_hcq.py test/external/external_test_am.py test/backend/test_asm_gemm.py::TestAsmGEMM --durations=20
      - name: Run disk copy tests
        run: python -m pytest test/unit/test_disk_tensor.py -k test_copy_from_disk
      - name: Run TRANSCENDENTAL math
@@ -604,7 +616,7 @@ jobs:
        backend: [ptx, nv]

    name: Linux (${{ matrix.backend }})
-    runs-on: *linux
+    runs-on: ${{ github.repository == 'tinygrad/tinygrad' && github.event_name == 'pull_request' && github.event.pull_request.author_association == 'COLLABORATOR' && 'namespace-profile-tinygrad' || 'ubuntu-24.04' }}
    timeout-minutes: 20
    env:
      FORWARD_ONLY: 1
@@ -638,10 +650,17 @@ jobs:
    strategy:
      fail-fast: false
      matrix:
-        backend: [ir3, nak]
-    name: Compile-only (${{ matrix.backend }})
-    runs-on: *linux
+        dev:
+          - 'NULL:IR3:a630'
+          - 'NULL:QCOMCL:a630'
+          - 'NULL:NAK:sm_120'
+    name: Compile-only (DEV=${{ matrix.dev }})
+    runs-on: ${{ github.repository == 'tinygrad/tinygrad' && github.event_name == 'pull_request' && github.event.pull_request.author_association == 'COLLABORATOR' && 'namespace-profile-tinygrad' || 'ubuntu-24.04' }}
    timeout-minutes: 15
+    env:
+      NULL_ALLOW_COPYOUT: 1
+      DEV: ${{ matrix.dev }}${{ contains(matrix.dev, 'a630') && ',IMAGE_PITCH_ALIGNMENT=64' || '' }}
+      IMAGE: ${{ contains(matrix.dev, 'a630') && '1' || '0' }}
    steps:
      - name: Checkout Code
        uses: actions/checkout@v6
@@ -650,21 +669,15 @@ jobs:
        with:
          key: compile-${{ matrix.backend }}
          deps: "testing_unit mesa"
-      - name: Set env
+          qemu: ${{ contains(matrix.dev, 'QCOMCL') }}
+      - name: Test IMAGE
        shell: bash
-        run: printf "NULL_ALLOW_COPYOUT=1\n${{ matrix.backend == 'ir3' && 'DEV=NULL:IR3:a630' || matrix.backend == 'nak' && 'DEV=NULL:NAK:sm_120' }}" >> $GITHUB_ENV
+        if: contains(matrix.dev, 'a630')
+        run: DEBUG=7 python3 test/backend/test_ops.py TestOps.test_gemm | grep isam
      - name: Run test_ops
        shell: bash
        run: |
          python -c "from tinygrad import Device; assert Device.DEFAULT == 'NULL'"
          DEBUG=4 python3 test/backend/test_ops.py TestOps.test_add
-          python -m pytest -n=auto test/backend/test_ops.py --durations=20
-      - name: Run test_ops (IMAGE)
-        if: matrix.backend == 'ir3'
-        shell: bash
-        env:
-          IMAGE: 1
-          DEV: "NULL:IR3:a630,IMAGE_PITCH_ALIGNMENT=64"
-        run: |
-          DEBUG=4 python3 test/backend/test_ops.py TestOps.test_gemm | grep image_load
-          python -m pytest -n=auto test/backend/test_ops.py --durations=20
+          # QCOMCL compiles in qemu, too slow for parallel workers
+          ${{ contains(matrix.dev, 'QCOMCL') && 'PARALLEL=0' || '' }} python -m pytest -n=auto test/backend/test_ops.py --durations=20
@@ -69,3 +69,4 @@ mutants
 dagre/
 graphlib/
 uv.lock
+pi_session_window0.jsonl
@@ -28,7 +28,7 @@ repos:
        pass_filenames: false
      - id: tests
        name: comprehensive test suite
-        entry: env OMP_NUM_THREADS=1 SKIP_SLOW_TEST=1 PYTHONPATH="." python3 -m pytest -n=6 test/backend/test_ops.py test/backend/test_schedule.py test/unit/test_assign.py test/backend/test_tensor.py test/backend/test_jit.py test/unit/test_schedule_cache.py test/null/test_pattern_matcher.py test/null/test_uop_symbolic.py test/unit/test_helpers.py
+        entry: env OMP_NUM_THREADS=1 SKIP_SLOW_TEST=1 PYTHONPATH="." python3 -m pytest -n=6 test/backend/test_ops.py test/backend/test_schedule.py test/backend/test_assign.py test/backend/test_tensor.py test/backend/test_jit.py test/unit/test_schedule_cache.py test/null/test_pattern_matcher.py test/null/test_uop_symbolic.py test/unit/test_helpers.py
        language: system
        always_run: true
        pass_filenames: false
@@ -4,3 +4,4 @@
 - Run `python -m mypy tinygrad/` to typecheck
 - Run `python -m ruff check .` to lint
 - Read `./tinygrad/viz/README.md` for profiling and debugging rewrite rules
+- Do not do amend commits. Always do a new commit if a force push to origin would be required.
@@ -140,7 +140,7 @@ Documentation along with a quick start guide can be found on the [docs website](
 ```python
 from tinygrad import Tensor

-x = Tensor.eye(3)
+x = Tensor.eye(3).clone()  # clone to make it a buffer
 y = Tensor([[2.0,0,-2.0]])
 z = y.matmul(x).sum()
 z.backward()
@@ -1 +1 @@
-8611fe22a7fcc7d1928bbde19ded66277cb12f3e
+f6fc4e3f2c3db5fae1e19cbfbc3ad9fc579a12ae
@@ -2,7 +2,7 @@ import os, pytest, signal, threading

@pytest.hookimpl(wrapper=True)
 def pytest_runtest_call(item):
-  t = threading.Timer(int(os.getenv("TEST_TIMEOUT", 300)), os.kill, args=(os.getpid(), signal.SIGABRT))
+  t = threading.Timer(int(os.getenv("TEST_TIMEOUT", 90)), os.kill, args=(os.getpid(), signal.SIGABRT))
  t.start()
  try: yield
  finally:
@@ -122,7 +122,7 @@ def example_5_custom_assembly(a:Tensor, correct):
        offset_dwords = (self.labels[inst._target] - inst._pos - inst.size()) // 4
        if not -32768 <= offset_dwords <= 32767: raise ValueError(f"branch to '{inst._target}' offset {offset_dwords} exceeds simm16 range")
        inst.simm16 = offset_dwords
-      return UOp(Ops.PROGRAM, src=(sink, UOp(Ops.LINEAR, src=tuple([UOp(Ops.INS, arg=x) for x in self.instructions]))))
+      return UOp(Ops.PROGRAM, src=(sink, UOp(Ops.LINEAR, src=tuple([UOp(Ops.INS, arg=(x, dtypes.void)) for x in self.instructions]))))

  CU_COUNT = 32
  LANES = 64
@@ -52,7 +52,7 @@ In `kernel.py` we have a set of `OptOps`, these control the parameters of the sp

 The main bottleneck in most kernels is accessing memory. In a freshman algorithms class, you'll learn about cache aware matrix multiplication, and this is all forms of that. While the same math is run, the order in which you run it can have large impacts on the speed depending on if the data you are loading. OptOps will change this order.

-Memory, even cache, is often much slower than accessing the register file. The amount of times data is used in math is called the "arithmetic intensity". For operations like BS=1 GEMV, the arithmetic intensity is 1, but for GEMMs and convs it can be much higher. OptOps like UPCAST and UNROLL can increase this, but be careful of making them too large, as if there's too much register pressure on the GPU the warp scheduler may not be able to fit many warps, or even worse, it could be spilling to local memory.
+Memory, even cache, is often much slower than accessing the register file. The amount of times data is used in math is called the "arithmetic intensity". For operations like BS=1 GEMV, the arithmetic intensity is 1, but for GEMMs and convs it can be much higher. Splitting an axis into UPCAST can increase this, but be careful of making them too large, as if there's too much register pressure on the GPU the warp scheduler may not be able to fit many warps, or even worse, it could be spilling to local memory.

 4090s have 1 TB/s of ram bandwidth and ~160 TFLOPS of compute, so you need to use each loaded value ~100 times. The L1 cache has around 40 TB/s of bandwidth, so in order to get full compute utilization you need to use each value ~4 times.

@@ -5,7 +5,7 @@ from multiprocessing import Queue, Process, shared_memory, connection, Lock

 import numpy as np
 from tinygrad import dtypes, Tensor
-from tinygrad.helpers import getenv, prod, Context, round_up, tqdm, OSX, NUM_CPU_THREADS
+from tinygrad.helpers import getenv, prod, Context, round_up, tqdm, OSX, CPU_COUNT
 from tinygrad.nn.state import TensorIO

 ### ResNet
@@ -131,7 +131,7 @@ def batch_load_resnet(batch_size=64, val=False, shuffle=True, seed=None, pad_fir
    else: X = Tensor.empty(*sz, dtype=dtypes.uint8, device=f"disk:/dev/shm/{shm_name}")
    Y = [None] * (batch_size*BATCH_COUNT)

-    for _ in range(NUM_CPU_THREADS.value):
+    for _ in range(CPU_COUNT):
      p = Process(target=loader_process, args=(q_in, q_out, X, seed))
      p.daemon = True
      p.start()
@@ -212,7 +212,7 @@ def batch_load_train_bert(BS:int, seed:int|None=None):
    rng.shuffle(fs)
    train_files.append(fs.pop(0))

-  cycle_length = min(NUM_CPU_THREADS.value, len(train_files))
+  cycle_length = min(CPU_COUNT, len(train_files))
  assert cycle_length > 0, "cycle_length must be greater than 0"

  dataset = InterleavedDataset(train_files, cycle_length)
@@ -301,7 +301,7 @@ def batch_load_unet3d(preprocessed_dataset_dir:Path, batch_size:int=6, val:bool=
    X = Tensor.empty(*sz, dtype=dtypes.float32, device=f"disk:/dev/shm/{shm_name_x}")
    Y = Tensor.empty(*sz, dtype=dtypes.uint8, device=f"disk:/dev/shm/{shm_name_y}")

-    for _ in range(NUM_CPU_THREADS.value):
+    for _ in range(CPU_COUNT):
      proc = Process(target=load_unet3d_data, args=(preprocessed_dataset_dir, seed, queue_in, queue_out, X, Y))
      proc.daemon = True
      proc.start()
@@ -437,7 +437,7 @@ def batch_load_retinanet(dataset, val:bool, base_dir:Path, batch_size:int=32, sh
  dataset_iter = iter(image_ids)

  try:
-    for _ in range(NUM_CPU_THREADS.value):
+    for _ in range(CPU_COUNT):
      proc = Process(
        target=load_retinanet_data,
        args=(base_dir, val, queue_in, queue_out, imgs, boxes, labels),
@@ -1667,15 +1667,14 @@ def train_llama3():
 def train_gptoss():
  from examples.mlperf.models.gpt_oss import GPTOSS, GPT_OSS_20B, apply_grad, FP8_DTYPE
  from examples.mlperf.lr_schedulers import CosineAnnealingLRWithWarmup
-  from examples.mlperf.optim import GradAccClipAdamW, GradAccClipAdamWGroup, clip_grads
+  from examples.mlperf.optim import GradAccClipAdamW, GradAccClipAdamWGroup, fclip_grads

  BENCHMARK = getenv("BENCHMARK")

  config = {}
  BASEDIR            = config["BASEDIR"]                = Path(getenv("BASEDIR", "/raid/datasets/c4-8b/"))
  BS                 = config["BS"]                     = getenv("BS", 16)
-  grad_acc           = config["GRADIENT_ACC_STEPS"]     = getenv("GRADIENT_ACC_STEPS", 1)
-  GBS                = config["GLOBAL_BATCH_SIZE"]      = BS * grad_acc
+  GBS                = config["GLOBAL_BATCH_SIZE"]      = BS
  SEED               = config["SEED"]                   = getenv("SEED", 5760)
  DATA_SEED          = config["DATA_SEED"]              = getenv("DATA_SEED", SEED)
  SEQLEN             = config["SEQLEN"]                 = getenv("SEQLEN", 8192)
@@ -1737,13 +1736,13 @@ def train_gptoss():
  params_wd = [p for p in params if p.ndim >= 3]
  params_no_wd = [p for p in params if p.ndim < 3]
  optim = GradAccClipAdamWGroup(
-    GradAccClipAdamW(params_wd, lr=0.0, b1=opt_adamw_beta_1, b2=opt_adamw_beta_2, eps=opt_adamw_epsilon, weight_decay=opt_adamw_weight_decay, grad_acc=grad_acc, device=optim_device),
-    GradAccClipAdamW(params_no_wd, lr=0.0, b1=opt_adamw_beta_1, b2=opt_adamw_beta_2, eps=opt_adamw_epsilon, weight_decay=0.0, grad_acc=grad_acc, device=optim_device),
+    GradAccClipAdamW(params_wd, lr=0.0, b1=opt_adamw_beta_1, b2=opt_adamw_beta_2, eps=opt_adamw_epsilon, weight_decay=opt_adamw_weight_decay, grad_acc=1, device=optim_device),
+    GradAccClipAdamW(params_no_wd, lr=0.0, b1=opt_adamw_beta_1, b2=opt_adamw_beta_2, eps=opt_adamw_epsilon, weight_decay=0.0, grad_acc=1, device=optim_device),
  )

  for p in optim.params:
-    grad_dtype = dtypes.bfloat16 if p.dtype == FP8_DTYPE else p.dtype
-    p.grad = p.zeros_like(dtype=grad_dtype).contiguous()
+    p.grad = p.zeros_like(dtype=dtypes.bfloat16 if p.dtype == FP8_DTYPE else p.dtype).contiguous()
+    if getattr(p, "_zero2", False): p.grad = optim.optimizers[0]._zero_shard(p.grad)
  grads = [p.grad for p in optim.params]

  from extra.gemm.cdna_asm_gemm import _mx_block_scale
@@ -1770,31 +1769,32 @@ def train_gptoss():

  @TinyJit
  @Context(TRAINING=1)
-  def minibatch(tokens:Tensor):
+  def step(tokens:Tensor):
    if is_dp: tokens = tokens.to(None).shard(device, 0)
    if not is_sharding: tokens = tokens.to(None)
+
    logits:Tensor = model(tokens[:, :-1], save=True)
-    loss = logits.sparse_categorical_crossentropy(tokens[:, 1:])
+    if getenv("FUSED_CE", 0):
+      from extra.llama_kernels.fused_ce import fused_ce_loss
+      loss = fused_ce_loss(logits.cast(dtypes.bfloat16), tokens[:, 1:], label_smoothing=0.0)
+    else:
+      loss = logits.sparse_categorical_crossentropy(tokens[:, 1:])

    for g, new_g in zip(grads, loss.gradient(*optim.params)):
      apply_grad(g, new_g.uop)

-    loss_cpu = loss.flatten().float().to("CPU")
-    return loss_cpu.realize(*grads)
+    Tensor.realize(loss, *grads)

-  @TinyJit
-  def optim_step():
-    grad_norm = clip_grads(grads, grad_acc, 1.0)
-    optim.fstep(grads, grad_norm)
+    clipped_grads, grad_norm = fclip_grads(grads, 1.0)
+    optim.fstep(clipped_grads, grad_norm)
    scheduler.step()

-    for g in grads: g.assign(0)
-
+    loss_cpu = loss.flatten().float().to("CPU")
    lr_cpu = optim.lr.float().to("CPU")
    grad_norm_cpu = grad_norm.float().to("CPU")
-    Tensor.realize(lr_cpu, grad_norm_cpu, *grads, *fp8_inv_scales)
+    Tensor.realize(loss_cpu, lr_cpu, grad_norm_cpu, *grads, *fp8_inv_scales)

-    return lr_cpu, grad_norm_cpu
+    return loss_cpu, lr_cpu, grad_norm_cpu

  @TinyJit
  @Context(TRAINING=0)
@@ -1843,30 +1843,20 @@ def train_gptoss():
      profile_marker(f"train @ {i}")
      st = time.perf_counter()

-      stopped = False
-      losses, data_time, dev_time = [], 0, 0
-      for _ in range(grad_acc if i >= 2 else 1):
-        ist = time.perf_counter()
-        try: tokens = next(train_iter)
-        except StopIteration:
-          stopped = True
-          break
-        mst = time.perf_counter()
-        data_time += mst - ist
-        losses.append(minibatch(tokens).item())
-        dev_time += time.perf_counter() - mst
-      if stopped: break
+      ist = time.perf_counter()

-      gt = time.perf_counter()
-      ret = optim_step()
-      lr, grad_norm = ret[0].item(), ret[1].item()
+      try: tokens = next(train_iter)
+      except StopIteration: break
+      mst = time.perf_counter()
+      data_time = mst - ist
+
+      ret = step(tokens)
+      dev_time = time.perf_counter() - mst
+
+      loss, lr, grad_norm = ret[0].item(), ret[1].item(), ret[2].item()
      et = time.perf_counter()

-      loss = sum(losses) / len(losses)
-      optim_time = et - gt
-      dev_time += optim_time
      step_time = et - st
-      gbs_time = gt - st
      if BENCHMARK: step_times.append(step_time)

      i += 1
@@ -1876,7 +1866,7 @@ def train_gptoss():
      gflops = GlobalCounters.global_ops / 1e9 / dev_time
      mfu = ((6 * num_params * SEQLEN * GBS) / (dev_time * device_count * 4.6e15)) * 100
      tqdm.write(
-          f"{i:5} {step_time:.3f} s step, {gbs_time:.3f} s gbs, {optim_time:.3f} s optim, {data_time:.3f} s data, {loss:.4f} loss, " \
+          f"{i:5} {step_time:.3f} s step, {dev_time:.3f} s dev, {data_time:.3f} s data, {loss:.4f} loss, " \
          f"{lr:.12f} LR, {grad_norm:.6f} grad_norm, {mem_gb:.2f} GB used, {gflops:9.2f} GFLOPS, {mfu:5.2f}% MFU")
      if DEBUG >= 1: tqdm.write("  mem per device: " + ', '.join(f"{dev}: {mem/1e9:.2f} GB" for dev, mem in sorted(GlobalCounters.mem_used_per_device.items())))

@@ -1886,8 +1876,6 @@ def train_gptoss():
          "train/lr": lr,
          "train/grad_norm": grad_norm,
          "train/step_time": step_time,
-          "train/gbs_time": gbs_time,
-          "train/optim_time": optim_time,
          "train/dev_time": dev_time,
          "train/data_time": data_time,
          "train/mem": mem_gb,
@@ -12,7 +12,7 @@ from tinygrad.helpers import Timing, colored, GlobalCounters, profile_marker
 from tinygrad.uop.ops import Ops, UOp
 from extra.models.llama import apply_rotary_emb
 from extra.llama_kernels.rmsnorm import rmsnorm
-from extra.gemm.cdna_asm_gemm import _mx_block_scale, _mx_block_scale_3d, quantize_mxfp8
+from extra.gemm.cdna_asm_gemm import _mx_block_scale, _mx_block_scale_3d, quantize_mxfp8, asm_gemm, can_use_asm_gemm
 from extra.gemm.moe_gemm import grouped_mx_gemm
 from extra.gemm.moe_routing import route, dispatch, combine

@@ -61,7 +61,25 @@ def dequant_weight(w_q:Tensor, w_scale:Tensor) -> Tensor:
  call = UOp.maketuple(fxn.uop).call(w_q.uop, w_scale.uop, grad_fxn=_dequant_bwd)
  return Tensor(call.gettuple(0))

-def matmul_mx(x:Tensor, w_q:Tensor, w_scale:Tensor) -> Tensor:
+def matmul_mx(x:Tensor|tuple[Tensor, Tensor], w_q:Tensor, w_scale:Tensor) -> Tensor:
+  if isinstance(x, tuple):
+    assert ASM_GEMM, "pre-quantized MXFP8 input requires ASM_GEMM"
+    from extra.gemm.cdna_asm_gemm import asm_gemm, can_use_asm_gemm, mx_pack
+    x_q, x_e8 = x
+    l_shape, padded = x_q.shape[:-1], x_q.shape[-1]
+    x_q, x_e8 = x_q.reshape(-1, padded), x_e8.reshape(-1, padded // 32)
+    K, N = w_q.shape[1], w_q.shape[0]
+    assert padded >= K and (padded - K) % 32 == 0 and x_e8.shape[-1] == padded // 32
+    wq, ws = w_q, w_scale
+    if (pad := padded - K):
+      wq = wq.pad(((0, 0), (0, pad)))
+      ws = ws.pad(((0, 0), (0, pad // 32)), value=127).cast(dtypes.uint8)
+    if (npad := (-N) % 256):
+      wq = wq.pad(((0, npad), (0, 0)))
+      ws = ws.pad(((0, npad), (0, 0)), value=127).cast(dtypes.uint8)
+    assert can_use_asm_gemm(x_q, wq.T)
+    out = asm_gemm(x_q, wq.T, mx=True, mx_scales=(mx_pack(x_e8), x_e8, mx_pack(ws), ws), mx_w_stored=True)
+    return (out[:, :N] if npad else out).reshape(*l_shape, N).cast(dtypes.bfloat16)
  l_shape = x.shape[:-1]
  if ASM_GEMM:
    from extra.gemm.cdna_asm_gemm import asm_gemm, can_use_asm_gemm, mx_pack
@@ -146,6 +164,7 @@ class GPTOSS:
      return w_q, w_e8.is_param_(False)
    if moe:
      qs = [_one(*shape[1:]) for _ in range(shape[0])]
+      for q in qs: q[0]._zero2 = True  # grad arrives sharded on the expert axis under ZeRO-2 (moe_gemm)
      return [q[0] for q in qs], [q[1] for q in qs]
    return _one(*shape)

@@ -174,20 +193,30 @@ class GPTOSS:
  def attention(self, x:Tensor, freqs_cis:Tensor, mask:Tensor, sliding:bool, *, attention_norm:Tensor, wqkv:Tensor,
                wqkv_scale:Tensor, wqkv_bias:Tensor, wo:Tensor, wo_scale:Tensor, wo_bias:Tensor, sinks:Tensor):
    bsz, seqlen, _ = x.shape
-    x_normed, rrms = rmsnorm(x, self.norm_eps)
-    qkv = matmul_mx(x_normed * attention_norm, wqkv, wqkv_scale) + wqkv_bias
+
+    if getenv("FUSED_RMSNORM_MUL", 0):
+      from extra.gptoss_kernels.rmsnorm import rmsnorm_mul
+      x_normed, rrms = rmsnorm_mul(x, attention_norm, self.norm_eps)
+      norm_saves = [x_normed, rrms]
+    else:
+      x_normed, rrms = rmsnorm(x, self.norm_eps)
+      qkv = matmul_mx(x_normed * attention_norm, wqkv, wqkv_scale) + wqkv_bias
+      norm_saves = [x_normed, rrms]
+
    qkv = qkv.reshape(bsz, seqlen, self.n_kv_heads, self.n_rep + 2, self.head_dim)
    xq = qkv[:, :, :, :self.n_rep].reshape(bsz, seqlen, self.n_heads, self.head_dim)
    xk, xv = qkv[:, :, :, self.n_rep], qkv[:, :, :, self.n_rep + 1]
    xq, xk = apply_rotary_emb(xq, xk, freqs_cis)
    xq, xk, xv = xq.cast(dtypes.bfloat16), xk.cast(dtypes.bfloat16), xv.cast(dtypes.bfloat16)  # (B,N,H,D)/(B,N,KV,D)

-    if sliding:
-      attn = self._sliding_attention(xq, xk, xv, sinks)
-    elif getenv("HK_FLASH_ATTENTION"):
+    fa_saves = []
+    if getenv("HK_FLASH_ATTENTION"):
      from extra.thunder.amd.fa import flash_attention
-      attn, *_ = flash_attention(xq, xk, xv, is_causal=True, write_flat=True, sinks=sinks)
+      attn, _, l_vec = flash_attention(xq, xk, xv, is_causal=True, write_flat=True, sinks=sinks, window=self.sliding_window if sliding else 0)
      attn = attn.reshape(bsz, seqlen, self.n_heads * self.head_dim)
+      fa_saves = [xq, xk, xv, l_vec]
+    elif sliding:
+      attn = self._sliding_attention(xq, xk, xv, sinks)
    else:
      xqm = xq.reshape(bsz, seqlen, self.n_kv_heads, self.n_rep, self.head_dim).permute(0, 2, 3, 1, 4)
      xkm, xvm = xk.permute(0, 2, 1, 3).unsqueeze(2), xv.permute(0, 2, 1, 3).unsqueeze(2)
@@ -199,13 +228,19 @@ class GPTOSS:
      attn = (w @ xvm).permute(0, 3, 1, 2, 4).reshape(bsz, seqlen, self.n_heads * self.head_dim)

    out = matmul_mx(attn, wo, wo_scale) + wo_bias
-    return out, [x_normed, rrms, attn]
+    return out, [attn] + norm_saves + fa_saves

  def feed_forward(self, x:Tensor, *, ffn_norm:Tensor, gate:Tensor, gate_bias:Tensor,
                   w_gate_up:Tensor, w_gate_up_scale:Tensor, w_gate_up_bias:Tensor,
                   w_down:Tensor, w_down_scale:Tensor, w_down_bias:Tensor):
-    x_normed, rrms = rmsnorm(x, self.norm_eps)
-    inp = x_normed * ffn_norm
+    if getenv("FUSED_RMSNORM_MUL", 0):
+      from extra.gptoss_kernels.rmsnorm import rmsnorm_mul
+      x_normed, rrms = rmsnorm_mul(x, ffn_norm, self.norm_eps)
+      inp = x_normed
+    else:
+      x_normed, rrms = rmsnorm(x, self.norm_eps)
+      inp = x_normed * ffn_norm
+
    logits = inp.float() @ gate.float().T + gate_bias.float()
    dim, inter = self.dim, self.intermediate_size

@@ -220,6 +255,7 @@ class GPTOSS:
      z = grouped_mx_gemm(_pad_cols(y.cast(dtypes.bfloat16)), (w_down, w_down_scale), r.off)[:, :dim] \
          + (onehot @ w_down_bias.float()).cast(dtypes.bfloat16)
      out = combine(z, r, inp.shape[0], self.experts_per_tok).reshape(bsz, seqlen, dim)
+      return out, [x_normed, rrms, xg, h, y, z, r.weights, r.dest_row, r.off]
    else:
      thresh = logits.topk(self.experts_per_tok)[0][..., -1:]
      weights = (logits >= thresh).where(logits, -float("inf")).softmax(-1)
@@ -263,7 +299,11 @@ class GPTOSS:
                        w_down=self.w_down[i], w_down_scale=self.w_down_scale[i], w_down_bias=self.w_down_bias[i])
      h, *_ = self.run_layer(h, freqs_cis, mask_full, i % 2 == 0, attn_kwargs, ffn_kwargs, save=save)

-    logits = self.norm(h) @ self.output.T
+    h_normed = self.norm(h)
+    pad = (-self.dim) % 256
+    h_padded, w_padded = h_normed.pad((None, None, (0, pad))), self.output.pad(((0, 0), (0, pad)))
+    if ASM_GEMM and can_use_asm_gemm(h_padded, w_padded.T): logits = asm_gemm(h_padded, w_padded.T)
+    else: logits = h_normed @ self.output.T
    return logits

 def _get_pads(uop:UOp) -> list[UOp]:
@@ -274,14 +314,14 @@ def apply_grad(grad_buf:Tensor, new_grad:UOp):
  pads = _get_pads(new_grad)
  if len(pads) <= 1:
    new_grad = new_grad.cast(grad_buf.dtype)
-    grad_buf.uop = grad_buf.uop.after(grad_buf.uop.store(grad_buf.uop + new_grad))
+    grad_buf.uop = grad_buf.uop.after(grad_buf.uop.store(new_grad))
    return
  cur = grad_buf.uop
  for pad in sorted(pads, key=lambda p: p.marg[0][0] if p.op == Ops.PAD else 0, reverse=True):
    if pad.op == Ops.PAD:
-      grad_shrink = tuple([(p[0], s+p[0]) for s,p in zip(pad.src[0].shape, pad.marg)])
+      grad_shrink = tuple((p[0], s+p[0]) for s,p in zip(pad.src[0].shape, pad.marg))
      buf_slice = cur.shrink(grad_shrink)
-      cur = cur.after(buf_slice.store(buf_slice + pad.src[0].cast(cur.dtype)))
+      cur = cur.after(buf_slice.store(pad.src[0].cast(cur.dtype)))
    else:
      cur = cur.after(cur.store(cur + pad.cast(cur.dtype)))
  grad_buf.uop = cur
@@ -15,7 +15,7 @@ def stochastic_round_bf16(x:Tensor) -> Tensor:
  bits = x.bitcast(dtypes.uint32)
  if isinstance(x.device, tuple):
    shape = x.uop.shard_shape if x.uop.axis is not None else x.shape
-    noise = Tensor(UOp(Ops.MSTACK, dtypes.default_float, tuple(Tensor.rand(*shape, device=d).uop for d in x.device)))
+    noise = Tensor(UOp(Ops.MSTACK, src=tuple(Tensor.rand(*shape, device=d).uop for d in x.device)))
  else:
    noise = x.rand_like()
  noise = (noise * 0xFFFF).cast(dtypes.uint32)
@@ -27,6 +27,11 @@ def clip_grads(grads:list[Tensor], grad_acc, clip_norm) -> Tensor:
  for g in grads: g.assign((g * (clip_norm / (total_norm + 1e-6)).clamp(max_=1.0)).cast(g.dtype))
  return total_norm

+def fclip_grads(grads:list[Tensor], clip_norm) -> Tensor:
+  total_norm = Tensor.stack(*[g.float().square().sum() for g in grads]).sum().sqrt().contiguous()
+  scale = (clip_norm / (total_norm + 1e-6)).clamp(max_=1.0)
+  return [(g * scale).cast(g.dtype) for g in grads], total_norm
+
 class GradAccClipAdamW(Optimizer):
  def __init__(self, params:list[Tensor], lr=0.001, b1=0.9, b2=0.999, eps=1e-6, weight_decay=0.0, grad_acc=1, clip_norm=1.0, device=None, fused=FUSE_OPTIM):
    super().__init__(params, lr, device, fused)
@@ -0,0 +1,13 @@
+from pathlib import Path
+
+from examples.mlperf.dataloader import get_llama3_dataset
+from tinygrad.helpers import getenv
+
+BASEDIR = Path(getenv("BASEDIR", "/raid/datasets/c4-8b/"))
+SAMPLES = getenv("SAMPLES", 1_200_000 * 32)
+EVAL_SAMPLES = getenv("EVAL_SAMPLES", 1024)
+SEQLEN = getenv("SEQLEN", 8192)
+DATA_SEED = getenv("DATA_SEED", 5760)
+
+get_llama3_dataset(SAMPLES, SEQLEN, BASEDIR, seed=DATA_SEED, val=False, small=True)
+get_llama3_dataset(EVAL_SAMPLES, SEQLEN, BASEDIR, seed=0, val=True, small=True)
@@ -26,7 +26,7 @@ export FUSED_SILU_W13=${FUSED_SILU_W13:-1}
 export SPLIT_W13=${SPLIT_W13:-0}
 export OFFLOAD_OPTIM=${OFFLOAD_OPTIM:-0}

-export DEFAULT_FLOAT="bfloat16" OPTIM_DTYPE="bfloat16"
+export DEFAULT_FLOAT="bfloat16" OPTIM_DTYPE="float32"
 export DP=${DP:-8} MP=${MP:-1} BS=${BS:-16} EVAL_BS=${EVAL_BS:-8} GRADIENT_ACC_STEPS=${GRADIENT_ACC_STEPS:-2}
 export GBS=$((BS * GRADIENT_ACC_STEPS))

@@ -35,7 +35,7 @@ export BASEDIR="/raid/datasets/c4-8b/"
 export SMALL=1
 export LLAMA3_SIZE=${LLAMA3_SIZE:-"8B"}
 export EVAL_TARGET=3.3 EVAL_FREQ=12288
-export LR="1e-3" END_LR="1e-4" WARMUP_SAMPLES=4096 MAX_STEPS=1200000
+export LR="1e-3" END_LR="1e-4" WARMUP_SAMPLES=2048 MAX_STEPS=1200000
 export WARMUP_STEPS=$((WARMUP_SAMPLES / GBS))
 export SAMPLES=$((MAX_STEPS * GBS))
 export SEQLEN=${SEQLEN:-8192}
@@ -44,7 +44,7 @@ export SEED=${SEED:-5760}
 export DATA_SEED=${DATA_SEED:-5760}

 export JITBEAM=${JITBEAM:-3}
-export BEAM_UOPS_MAX=6000 BEAM_UPCAST_MAX=256 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5 BEAM_PADTO=1
+export BEAM_UOPS_MAX=6000 BEAM_UPCAST_MAX=256 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5 BEAM_PADTO=0

 export FAKEDATA=${FAKEDATA:-1} BENCHMARK=${BENCHMARK:-10}
 if [ -z "$FULL_LAYERS" ]; then
@@ -1,8 +1,8 @@
 #!/usr/bin/env bash

 export PYTHONPATH="."
-export PATH="/opt/rocm-7.1.1/bin:$PATH"
-export ROCM_PATH="/opt/rocm-7.1.1"
+export ROCM_PATH=${ROCM_PATH:-/opt/rocm-7.1.1}
+export PATH="$ROCM_PATH/bin:$PATH"
 export DEV=${DEV:-AMD}
 export CHECK_OOB=0
 export REWRITE_STACK_LIMIT=5000000 HCQDEV_WAIT_TIMEOUT_MS=240000
@@ -26,7 +26,7 @@ export FUSED_SILU_W13=${FUSED_SILU_W13:-1}
 export SPLIT_W13=${SPLIT_W13:-0}
 export OFFLOAD_OPTIM=${OFFLOAD_OPTIM:-0}

-export DEFAULT_FLOAT="bfloat16" OPTIM_DTYPE="bfloat16"
+export DEFAULT_FLOAT="bfloat16" OPTIM_DTYPE="float32"
 export DP=${DP:-8} MP=${MP:-1} BS=${BS:-16} EVAL_BS=${EVAL_BS:-8} GRADIENT_ACC_STEPS=${GRADIENT_ACC_STEPS:-2}
 export GBS=$((BS * GRADIENT_ACC_STEPS))

@@ -35,7 +35,7 @@ export BASEDIR="/raid/datasets/c4-8b/"
 export SMALL=1
 export LLAMA3_SIZE=${LLAMA3_SIZE:-"8B"}
 export EVAL_TARGET=3.3 EVAL_FREQ=12288
-export LR="1e-3" END_LR="1e-4" WARMUP_SAMPLES=4096 MAX_STEPS=1200000
+export LR="1e-3" END_LR="1e-4" WARMUP_SAMPLES=2048 MAX_STEPS=1200000
 export WARMUP_STEPS=$((WARMUP_SAMPLES / GBS))
 export SAMPLES=$((MAX_STEPS * GBS))
 export SEQLEN=${SEQLEN:-8192}
@@ -44,6 +44,6 @@ export SEED=${SEED:-$RANDOM}
 export DATA_SEED=${DATA_SEED:-5760}

 export JITBEAM=${JITBEAM:-3}
-export BEAM_UOPS_MAX=6000 BEAM_UPCAST_MAX=256 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5 BEAM_PADTO=1
+export BEAM_UOPS_MAX=6000 BEAM_UPCAST_MAX=256 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5 BEAM_PADTO=0

 python3 examples/mlperf/model_train.py
@@ -17,7 +17,7 @@ export USE_ATOMICS=1
 export ASM_GEMM=1
 export WQKV=1
 export MASTER_WEIGHTS=1
-export FP8=1
+export MXFP4=1
 export ALLREDUCE_CAST=1
 export FAST_CE=1
 export FUSED_INPUT_QUANTIZE=1
@@ -26,7 +26,7 @@ export FUSED_ADD_NORM_MUL_QUANTIZE=1
 export FUSED_SILU_W13=1
 export SPLIT_W13=0

-export DEFAULT_FLOAT="bfloat16" OPTIM_DTYPE="bfloat16"
+export DEFAULT_FLOAT="bfloat16" OPTIM_DTYPE="float32"
 export DP=8 MP=1 BS=16 EVAL_BS=8 GRADIENT_ACC_STEPS=2
 export GBS=$((BS * GRADIENT_ACC_STEPS))

@@ -44,7 +44,7 @@ export SEED=$RANDOM
 export DATA_SEED=$SEED

 export JITBEAM=3
-export BEAM_UOPS_MAX=6000 BEAM_UPCAST_MAX=256 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5 BEAM_PADTO=1
+export BEAM_UOPS_MAX=6000 BEAM_UPCAST_MAX=256 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5 BEAM_PADTO=0

 export LOGMLPERF=1

@@ -11,6 +11,7 @@ export DEVICE_IN_FUNCTION_BUG=1
 export DEBUG=${DEBUG:-2}
 export HK_FLASH_ATTENTION=${HK_FLASH_ATTENTION:-1}
 export ASM_GEMM=${ASM_GEMM:-1}
+export GROUPED_MOE=${GROUPED_MOE:-1}
 export ALL2ALL=${ALL2ALL:-1}
 export LATE_ALLREDUCE=${LATE_ALLREDUCE:-0}
 export ALLREDUCE_CAST=${ALLREDUCE_CAST:-1}
--- a/Show More
+++ b/Show More
Author	SHA1	Message	Date
firestar5683	08b519ca77	files	2026-09-02 11:17:46 -05:00
firestar5683	449cb11938	Pin	2026-09-02 09:59:09 -05:00