tpu-inference 0.11.1.dev202511180814__py3-none-any.whl → 0.12.0.dev20251213__py3-none-any.whl

tpu_inference/kernels/fused_moe/v1/kernel.py

@@ -7,7 +7,6 @@ import jax.numpy as jnp
 from jax import lax
 from jax._src import dtypes
 from jax.experimental import pallas as pl
-from jax.experimental import shard_map
 from jax.experimental.pallas import tpu as pltpu
 
 P = jax.sharding.PartitionSpec
@@ -20,7 +19,8 @@ def align_to(x, a):
 
 
 def get_dtype_packing(dtype):
-    bits = dtypes.bit_width(dtype)
+    bits = (dtypes.bit_width(dtype)
+            if hasattr(dtypes, "bit_width") else dtypes.itemsize_bits(dtype))
     return 32 // bits
 
 
@@ -35,13 +35,50 @@ def broadcast_minor(src, shape):
                            axis=-1)[..., :shape[-1]]
 
 
+def swigluoai(gate: jax.Array,
+              up: jax.Array,
+              *,
+              alpha: float = 1.702,
+              limit: float = 7.0) -> jax.Array:
+    """Activation used in some models such as GPT-OSS."""
+    gate = jnp.clip(gate, a_max=limit)
+    up = jnp.clip(up, a_min=-limit, a_max=limit)
+    glu = gate * jax.nn.sigmoid(alpha * gate)
+    return (up + 1.0) * glu
+
+
+def activation_fn(acc1, acc3, act_fn):
+    if act_fn == "silu":
+        return jax.nn.silu(acc1) * acc3
+    elif act_fn == "gelu":
+        return jax.nn.gelu(acc1) * acc3
+    elif act_fn == "swigluoai":
+        return swigluoai(acc1, acc3)
+    else:
+        raise RuntimeError(f"Unsupported activation function: {act_fn}")
+
+
 def ref_moe(
-    tokens: jax.Array,  # (num_tokens, hidden_size)
-    w1: jax.Array,  # (num_experts, 2, hidden_size, intermediate_size)
-    w2: jax.Array,  # (num_experts, intermediate_size, hidden_size)
-    gating_output: jax.Array,  # (num_tokens, num_experts)
-    top_k: int,
-    activation="silu",
+        tokens: jax.Array,  # (num_tokens, hidden_size)
+        w1: jax.Array,  # (num_experts, 2, hidden_size, intermediate_size)
+        w2: jax.Array,  # (num_experts, intermediate_size, hidden_size)
+        gating_output: jax.Array,  # (num_tokens, num_experts)
+        top_k: int,
+        *,
+        renormalize_topk_logits: bool = False,
+        act_fn: str = "silu",
+        subc_quant_wsz: int | None = None,
+        w1_scale:
+    (
+        jax.Array | None
+    ) = None,  # F32(num_experts, 2, hidden_size //subc_quant_wsz, 1, intermediate_size)
+        w2_scale:
+    (
+        jax.Array | None
+    ) = None,  # F32(num_experts, intermediate_size // subc_quant_wsz, 1, hidden_size)
+        b1: jax.Array
+    | None = None,  # F32(num_experts, 2, 1, intermediate_size)
+        b2: jax.Array | None = None,  # F32(num_experts, 1, hidden_size)
 ):
     n_tokens = tokens.shape[0]  # num_tokens
 
@@ -53,11 +90,16 @@ def ref_moe(
     top_k_logits, top_k_indices = lax.top_k(
         gating_logits, top_k)  # [num_tokens, top_k], [num_tokens, top_k]
 
+    if renormalize_topk_logits:
+        top_k_logits = top_k_logits / jnp.sum(
+            top_k_logits, axis=-1, keepdims=True)
+
     t_outputs = []
+    hidden_size, intermediate_size = w1.shape[-2:]
 
     # Process each token individually
     for i in range(n_tokens):
-        curr_token = jnp.expand_dims(tokens[i], axis=0)  # [1, d_model]
+        curr_token = jnp.expand_dims(tokens[i], axis=0)  # [1, hidden_size]
         assigned_expert_ids = top_k_indices[
             i]  # [top_k] - indices of selected experts for token i
         tok_expert_act = []
@@ -65,10 +107,24 @@ def ref_moe(
         # Process each selected expert for the current token
         for expert_id in assigned_expert_ids:
             # Get expert weights
+            expert_w1 = w1[expert_id, 0].astype(jnp.float32)
+            expert_w3 = w1[expert_id, 1].astype(jnp.float32)
+            if w1_scale is not None:
+                expert_w1 *= jnp.repeat(w1_scale[expert_id, 0, :, 0],
+                                        subc_quant_wsz,
+                                        axis=0)[:hidden_size]
+                expert_w3 *= jnp.repeat(w1_scale[expert_id, 1, :, 0],
+                                        subc_quant_wsz,
+                                        axis=0)[:hidden_size]
             expert_weight_1 = jnp.concat(
-                [w1[expert_id, 0], w1[expert_id, 1]],
-                axis=-1)  # [d_model, 2 * intermediate_size]
-            expert_weight_2 = w2[expert_id]  # [intermediate_size, d_model]
+                [expert_w1, expert_w3],
+                axis=-1)  # [hidden_size, 2 * intermediate_size]
+            expert_weight_2 = w2[expert_id].astype(
+                jnp.float32)  # [intermediate_size, hidden_size]
+            if w2_scale is not None:
+                expert_weight_2 *= jnp.repeat(w2_scale[expert_id, :, 0],
+                                              subc_quant_wsz,
+                                              axis=0)[:intermediate_size]
 
             # First linear layer with SwiGLU activation
             gmm_1_out = curr_token @ expert_weight_1  # [1, 2 * intermediate_size]
@@ -77,37 +133,34 @@ def ref_moe(
             gmm1_w1_proj, gmm1_w3_proj = jnp.split(
                 gmm_1_out, 2,
                 axis=-1)  # [1, intermediate_size], [1, intermediate_size]
+            if b1 is not None:
+                gmm1_w1_proj += b1[expert_id:expert_id + 1, 0, 0]
+                gmm1_w3_proj += b1[expert_id:expert_id + 1, 1, 0]
 
             # Apply gated activation: activation(gate) * up
-            if activation == "silu":
-                act = jax.nn.silu(
-                    gmm1_w1_proj) * gmm1_w3_proj  # [1, intermediate_size]
-            elif activation == "gelu":
-                act = jax.nn.gelu(
-                    gmm1_w1_proj) * gmm1_w3_proj  # [1, intermediate_size]
-            else:
-                raise ValueError(
-                    f"Unsupported activation: {activation}. Use 'silu' or 'gelu'."
-                )
+            act = activation_fn(gmm1_w1_proj, gmm1_w3_proj, act_fn)
 
             # Second linear layer (down projection)
-            gmm_2_out = act @ expert_weight_2  # [1, d_model]
+            gmm_2_out = act @ expert_weight_2  # [1, hidden_size]
+            if b2 is not None:
+                gmm_2_out += b2[expert_id:expert_id + 1, 0]
             tok_expert_act.append(gmm_2_out)
 
         # Combine outputs from all selected experts
         experts_act = jnp.concatenate(tok_expert_act,
-                                      axis=0)  # [top_k, d_model]
+                                      axis=0)  # [top_k, hidden_size]
 
         # Weighted sum using top-k gating weights
         top_k_weights = top_k_logits[i]  # [top_k]
         top_k_weights = jnp.expand_dims(top_k_weights, axis=1)  # [top_k, 1]
         weighted_output = jnp.sum(experts_act * top_k_weights,
                                   axis=0,
-                                  keepdims=True)  # [1, d_model]
+                                  keepdims=True)  # [1, hidden_size]
 
-        t_outputs.append(weighted_output)
+        t_outputs.append(weighted_output.astype(tokens.dtype))
 
-    return jnp.concatenate(t_outputs, axis=0)  # [num_tokens, d_model]
+    return jnp.concatenate(t_outputs,
+                           axis=0)  # [actual_num_tokens, hidden_size]
 
 
 def _fused_ep_moe_kernel(
@@ -115,12 +168,19 @@ def _fused_ep_moe_kernel(
         tokens_hbm,  # (local_num_tokens, t_packing, hidden_size // t_packing)
         w1_hbm,  # (local_num_experts, 2, hidden_size, intermediate_size)
         w2_hbm,  # (local_num_experts, intermediate_size, hidden_size)
+        # TODO(jevinjiang): We choose F32 scale for easier slicing. The extra
+        # latency should be hidden in the pipeline overlaping. But is there a better
+        # way to do this?
+    w1_scale_hbm,  # None | F32(local_num_experts, 2, cdiv(hidden_size, subc_quant_wsz), 1, intermediate_size)
+        w2_scale_hbm,  # None | F32(local_num_experts, cdiv(intermediate_size, subc_quant_wsz), 1, hidden_size)
+        b1_hbm,  # None | F32(local_num_experts, 2, 1, intermediate_size)
+        b2_hbm,  # None | F32(local_num_experts, 1, hidden_size)
         gating_hbm,  # (local_num_tokens, padded_num_experts)
         a2a_g_hbm,  # (num_experts, bt, t_packing, hidden_size // t_packing)
         # Output
     output_hbm,  # (local_num_tokens, hidden_size)
         # Scratch
-    t2e_routing_x2_smem,  # <bt_sem_id> (2, bt, padded_num_experts)
+    t2e_routing_x2_smem,  # <bt_sem_id> (2, bt, padded_top_k)
         d2e_count_x2_smem,  # <bt_sem_id> (2, num_devices, 1, padded_num_experts)
         expert_offsets_x2_smem,  # <bt_sem_id> (2, 2, padded_num_experts): for a2a_s and a2a_g
         expert_starts_x2_smem,  # <bt_sem_id> (2, 1, padded_num_experts)
@@ -136,6 +196,12 @@ def _fused_ep_moe_kernel(
         b_w1_x2_vmem,  # <bw_sem_id> (2, t_packing, bd1 // t_packing, bf)
         b_w3_x2_vmem,  # <bw_sem_id> (2, t_packing, bd1 // t_packing, bf)
         b_w2_x2_vmem,  # <bw_sem_id> (2, t_packing, bf, bd2 // t_packing)
+        b_w1_scale_x2_vmem,  # None | <bw_sem_id> (2, t_packing, bd1 // t_packing // subc_quant_wsz, 1, bf)
+        b_w3_scale_x2_vmem,  # None | <bw_sem_id> (2, t_packing, bd1 // t_packing // subc_quant_wsz, 1, bf)
+        b_w2_scale_x2_vmem,  # None | <bw_sem_id> (2, t_packing, bf // subc_quant_wsz, 1, bd2 // t_packing)
+        b_b1_x2_vmem,  # None | <bw_sem_id> (2, 1, bf)
+        b_b3_x2_vmem,  # None | <bw_sem_id> (2, 1, bf)
+        b_b2_x2_vmem,  # None | <bw_sem_id> (2, t_packing, 1, bd2 // t_packing)
         b_acc_vmem,  # F32(bt * num_devices, 1, bf * 2)
         ### Semaphores:
     local_sems,  # (2, 5): 2 x [b_gating_sem, b_w1_sem, b_w2_sem, b_w3_sem, b_output_sem]
@@ -145,7 +211,10 @@ def _fused_ep_moe_kernel(
         a2a_acc_sem,
         *,
         top_k: int,
+        renormalize_topk_logits: bool,
         ep_axis_name: str,
+        act_fn: str,
+        subc_quant_wsz: int | None = None,
         # Kernel tuning params.
         bt: int,  # Block size of local_num_tokens.
         bf: int,  # Block size of intermediate_size.
@@ -160,34 +229,58 @@ def _fused_ep_moe_kernel(
     num_devices = lax.axis_size(ep_axis_name)
     local_num_tokens = tokens_hbm.shape[0]
     local_num_experts, intermediate_size, hidden_size = w2_hbm.shape
-    # num_experts = local_num_experts * num_devices
-    # padded_num_experts = expert_starts_x2_smem.shape[-1]
     right_id = (my_id + 1) % num_devices
+    num_experts = a2a_g_hbm.shape[0]
+    padded_num_experts = d2e_count_x2_smem.shape[-1]
+    padded_top_k = t2e_routing_x2_smem.shape[-1]
+    assert padded_num_experts == align_to(num_experts, 128)
+    assert padded_top_k == align_to(top_k, 128)
 
     t_dtype = tokens_hbm.dtype
     t_packing = get_dtype_packing(t_dtype)
     t_bitwidth = 32 // t_packing
     assert a2a_g_hbm.dtype == t_dtype
-    assert w1_hbm.dtype == t_dtype
-    assert w2_hbm.dtype == t_dtype
+    assert w1_hbm.dtype == w2_hbm.dtype
 
-    h_per_packing = hidden_size // t_packing
-    assert tokens_hbm.shape[-1] == h_per_packing
-    bd1_per_packing = bd1 // t_packing
-    bd2_per_packing = bd2 // t_packing
-    bd1c_per_packing = bd1c // t_packing
-    bd2c_per_packing = bd2c // t_packing
+    assert bd1 % bd1c == 0
+    assert bd2 % bd2c == 0
+    assert bf % bfc == 0
+    assert hidden_size % t_packing == 0
+    assert bd1 % t_packing == 0
+    assert bd2 % t_packing == 0
+    assert bd1c % t_packing == 0
+    assert bd2c % t_packing == 0
+
+    h_per_t_packing = hidden_size // t_packing
+    assert tokens_hbm.shape[-1] == h_per_t_packing
+    bd1_per_t_packing = bd1 // t_packing
+    bd2_per_t_packing = bd2 // t_packing
+    bd1c_per_t_packing = bd1c // t_packing
+    bd2c_per_t_packing = bd2c // t_packing
+
+    if subc_quant_wsz is not None:
+        assert subc_quant_wsz % 256 == 0
+        assert bd1c_per_t_packing == subc_quant_wsz
+        assert bfc == subc_quant_wsz
+        assert bd1 % subc_quant_wsz == 0
+        assert bf % subc_quant_wsz == 0
+        assert bd1_per_t_packing % subc_quant_wsz == 0
+        assert h_per_t_packing % subc_quant_wsz == 0
 
     num_bt = cdiv(local_num_tokens, bt)
     num_bf = cdiv(intermediate_size, bf)
     num_bd1 = cdiv(hidden_size, bd1)
     num_bd2 = cdiv(hidden_size, bd2)
 
+    def get_mesh_device_id(ep_rank):
+        dp_rank = jax.lax.axis_index("data")
+        return (dp_rank, ep_rank)
+
     def sync_barrier():
         barrier_sem = pltpu.get_barrier_semaphore()
         pltpu.semaphore_signal(
             barrier_sem,
-            device_id=(0, right_id),
+            device_id=get_mesh_device_id(right_id),
             device_id_type=pltpu.DeviceIdType.MESH,
         )
         pltpu.semaphore_wait(barrier_sem, 1)
@@ -212,30 +305,44 @@ def _fused_ep_moe_kernel(
             sem=b_gating_sem,
         ).wait()
 
-    def get_top_k(input, top_k):
+    def get_top_k(input, top_k, renormalize_topk_logits):
         assert len(input.shape) == 2, input.shape
         input = input.astype(jnp.float32)
+        padded_k_shape = (input.shape[0], padded_top_k)
         top_k_logits_lst = []
         top_k_indices_lst = []
         t2e = jnp.zeros(input.shape, dtype=jnp.int32)
-        t2e_routing = jnp.zeros(input.shape, dtype=jnp.int32)
+        t2e_routing = jnp.zeros(padded_k_shape, dtype=jnp.int32)
         iota = jax.lax.broadcasted_iota(jnp.int32, input.shape, 1)
+        padded_k_iota = jax.lax.broadcasted_iota(jnp.int32, padded_k_shape, 1)
+        top_k_logits_sum = jnp.zeros(padded_k_shape, jnp.float32)
+
         for k_id in range(top_k):
-            # TODO(jevinjiang): return both top_k values and indices in op in Mosaic
+            # TODO(jevinjiang): return both top_k values and indices in Mosaic
             top_k_logits = jnp.broadcast_to(
-                jnp.max(input, axis=1, keepdims=True),
-                (input.shape[0], 128)).astype(input.dtype)
+                jnp.max(input[:, :num_experts], axis=1, keepdims=True),
+                padded_k_shape,
+            ).astype(input.dtype)
             top_k_logits_lst.append(top_k_logits)
+            if renormalize_topk_logits:
+                top_k_logits_sum += top_k_logits
             # TODO(jevinjiang): support bf16 argmax in Mosaic
             top_k_indices = jnp.broadcast_to(
-                jnp.argmax(input, axis=1, keepdims=True), input.shape)
+                jnp.argmax(input[:, :num_experts], axis=1, keepdims=True),
+                padded_k_shape,
+            )
             top_k_indices_lst.append(top_k_indices)
-            t2e_routing = jnp.where(iota == k_id, top_k_indices, t2e_routing)
-            mask = iota == top_k_indices
+            t2e_routing = jnp.where(padded_k_iota == k_id, top_k_indices,
+                                    t2e_routing)
+            mask = iota == broadcast_minor(top_k_indices, input.shape)
             t2e += mask.astype(jnp.int32)
             if k_id != top_k - 1:
                 input = jnp.where(mask, -jnp.inf, input)
 
+        if renormalize_topk_logits:
+            for k_id in range(top_k):
+                top_k_logits_lst[k_id] /= top_k_logits_sum
+
         expert_sizes = jnp.sum(t2e, axis=0, keepdims=True)
         expert_starts = jnp.zeros_like(expert_sizes)
         return top_k_logits_lst, t2e_routing, expert_sizes, expert_starts
@@ -277,7 +384,7 @@ def _fused_ep_moe_kernel(
                     dst_ref=d2e_count_vmem.at[row_id],
                     send_sem=send_sem,
                     recv_sem=recv_sem,
-                    device_id=(0, right_id),
+                    device_id=get_mesh_device_id(right_id),
                     device_id_type=pltpu.DeviceIdType.MESH,
                 ).wait()
                 row_id = (row_id + num_devices - 1) % num_devices
@@ -359,10 +466,8 @@ def _fused_ep_moe_kernel(
                                              pl.ds(start, remote_sz)],
                     send_sem=send_sems.at[e_sem_id],
                     recv_sem=recv_sems.at[e_sem_id],
-                    device_id=(
-                        0,
-                        recv_id,
-                    ),
+                    device_id=get_mesh_device_id(recv_id),
+                    device_id_type=pltpu.DeviceIdType.MESH,
                 ).start()
         a2a_s_sends_x2_smem[e_sem_id] = send_sz
 
@@ -406,7 +511,8 @@ def _fused_ep_moe_kernel(
                 dst_ref=a2a_g_hbm.at[my_e_id, pl.ds(0, remote_sz)],
                 send_sem=send_sems.at[e_sem_id],
                 recv_sem=a2a_gather_sem,
-                device_id=(0, recv_id),
+                device_id=get_mesh_device_id(recv_id),
+                device_id_type=pltpu.DeviceIdType.MESH,
             ).start()
             start += sz
 
@@ -435,68 +541,173 @@ def _fused_ep_moe_kernel(
 
     def start_fetch_bw1(local_e_id, bw1_sem_id, bf_id, bd1_id):
         for p in range(t_packing):
-            offset = p * h_per_packing + bd1_id * bd1_per_packing
+            offset = p * h_per_t_packing + bd1_id * bd1_per_t_packing
             pltpu.make_async_copy(
                 src_ref=w1_hbm.at[
                     local_e_id,
                     0,
-                    pl.ds(offset, bd1_per_packing),
+                    pl.ds(offset, bd1_per_t_packing),
                     pl.ds(bf_id * bf, bf),
                 ],
                 dst_ref=b_w1_x2_vmem.at[bw1_sem_id, p],
                 sem=local_sems.at[bw1_sem_id, 1],
             ).start()
+            if w1_scale_hbm is not None:
+                assert subc_quant_wsz is not None
+                pltpu.make_async_copy(
+                    src_ref=w1_scale_hbm.at[
+                        local_e_id,
+                        0,
+                        pl.ds(
+                            offset // subc_quant_wsz,
+                            bd1_per_t_packing // subc_quant_wsz,
+                        ),
+                        pl.ds(0, 1),
+                        pl.ds(bf_id * bf, bf),
+                    ],
+                    dst_ref=b_w1_scale_x2_vmem.at[bw1_sem_id, p],
+                    sem=local_sems.at[bw1_sem_id, 1],
+                ).start()
+        if b1_hbm is not None and bd1_id == 0:
+            pltpu.make_async_copy(
+                src_ref=b1_hbm.at[local_e_id, 0,
+                                  pl.ds(0, 1),
+                                  pl.ds(bf_id * bf, bf)],
+                dst_ref=b_b1_x2_vmem.at[bf_id % 2],
+                sem=local_sems.at[bw1_sem_id, 1],
+            ).start()
 
     def start_fetch_bw2(local_e_id, bw2_sem_id, bf_id, bd2_id):
         for p in range(t_packing):
-            offset = p * h_per_packing + bd2_id * bd2_per_packing
+            offset = p * h_per_t_packing + bd2_id * bd2_per_t_packing
             pltpu.make_async_copy(
                 src_ref=w2_hbm.at[
                     local_e_id,
                     pl.ds(bf_id * bf, bf),
-                    pl.ds(offset, bd2_per_packing),
+                    pl.ds(offset, bd2_per_t_packing),
                 ],
                 dst_ref=b_w2_x2_vmem.at[bw2_sem_id, p],
                 sem=local_sems.at[bw2_sem_id, 2],
             ).start()
+            if w2_scale_hbm is not None:
+                assert subc_quant_wsz is not None
+                pltpu.make_async_copy(
+                    src_ref=w2_scale_hbm.at[
+                        local_e_id,
+                        pl.ds(bf_id * bf // subc_quant_wsz, bf //
+                              subc_quant_wsz),
+                        pl.ds(0, 1),
+                        pl.ds(offset, bd2_per_t_packing),
+                    ],
+                    dst_ref=b_w2_scale_x2_vmem.at[bw2_sem_id, p],
+                    sem=local_sems.at[bw2_sem_id, 2],
+                ).start()
+            if b2_hbm is not None and bf_id == 0:
+                pltpu.make_async_copy(
+                    src_ref=b2_hbm.at[local_e_id,
+                                      pl.ds(0, 1),
+                                      pl.ds(offset, bd2_per_t_packing)],
+                    dst_ref=b_b2_x2_vmem.at[bd2_id % 2, p],
+                    sem=local_sems.at[bw2_sem_id, 2],
+                ).start()
 
     def start_fetch_bw3(local_e_id, bw3_sem_id, bf_id, bd3_id):
         for p in range(t_packing):
-            offset = p * h_per_packing + bd3_id * bd1_per_packing
+            offset = p * h_per_t_packing + bd3_id * bd1_per_t_packing
             pltpu.make_async_copy(
                 src_ref=w1_hbm.at[
                     local_e_id,
                     1,
-                    pl.ds(offset, bd1_per_packing),
+                    pl.ds(offset, bd1_per_t_packing),
                     pl.ds(bf_id * bf, bf),
                 ],
                 dst_ref=b_w3_x2_vmem.at[bw3_sem_id, p],
                 sem=local_sems.at[bw3_sem_id, 3],
             ).start()
+            if w1_scale_hbm is not None:
+                assert subc_quant_wsz is not None
+                pltpu.make_async_copy(
+                    src_ref=w1_scale_hbm.at[
+                        local_e_id,
+                        1,
+                        pl.ds(
+                            offset // subc_quant_wsz,
+                            bd1_per_t_packing // subc_quant_wsz,
+                        ),
+                        pl.ds(0, 1),
+                        pl.ds(bf_id * bf, bf),
+                    ],
+                    dst_ref=b_w3_scale_x2_vmem.at[bw3_sem_id, p],
+                    sem=local_sems.at[bw3_sem_id, 3],
+                ).start()
+        if b1_hbm is not None and bd3_id == 0:
+            pltpu.make_async_copy(
+                src_ref=b1_hbm.at[local_e_id, 1,
+                                  pl.ds(0, 1),
+                                  pl.ds(bf_id * bf, bf)],
+                dst_ref=b_b3_x2_vmem.at[bf_id % 2],
+                sem=local_sems.at[bw3_sem_id, 3],
+            ).start()
 
     def wait_fetch_bw1(local_e_id, bw1_sem_id, bf_id, bd1_id):
-        del local_e_id, bf_id, bd1_id
+        del local_e_id
         pltpu.make_async_copy(
             src_ref=b_w1_x2_vmem.at[bw1_sem_id],
             dst_ref=b_w1_x2_vmem.at[bw1_sem_id],
             sem=local_sems.at[bw1_sem_id, 1],
         ).wait()
+        if w1_scale_hbm is not None:
+            pltpu.make_async_copy(
+                src_ref=b_w1_scale_x2_vmem.at[bw1_sem_id],
+                dst_ref=b_w1_scale_x2_vmem.at[bw1_sem_id],
+                sem=local_sems.at[bw1_sem_id, 1],
+            ).wait()
+        if b1_hbm is not None and bd1_id == 0:
+            pltpu.make_async_copy(
+                src_ref=b_b1_x2_vmem.at[bf_id % 2],
+                dst_ref=b_b1_x2_vmem.at[bf_id % 2],
+                sem=local_sems.at[bw1_sem_id, 1],
+            ).wait()
 
     def wait_fetch_bw2(local_e_id, bw2_sem_id, bf_id, bd2_id):
-        del local_e_id, bf_id, bd2_id
+        del local_e_id
         pltpu.make_async_copy(
             src_ref=b_w2_x2_vmem.at[bw2_sem_id],
             dst_ref=b_w2_x2_vmem.at[bw2_sem_id],
             sem=local_sems.at[bw2_sem_id, 2],
         ).wait()
+        if w2_scale_hbm is not None:
+            pltpu.make_async_copy(
+                src_ref=b_w2_scale_x2_vmem.at[bw2_sem_id],
+                dst_ref=b_w2_scale_x2_vmem.at[bw2_sem_id],
+                sem=local_sems.at[bw2_sem_id, 2],
+            ).wait()
+        if b2_hbm is not None and bf_id == 0:
+            pltpu.make_async_copy(
+                src_ref=b_b2_x2_vmem.at[bd2_id % 2],
+                dst_ref=b_b2_x2_vmem.at[bd2_id % 2],
+                sem=local_sems.at[bw2_sem_id, 2],
+            ).wait()
 
     def wait_fetch_bw3(local_e_id, bw3_sem_id, bf_id, bd3_id):
-        del local_e_id, bf_id, bd3_id
+        del local_e_id
         pltpu.make_async_copy(
             src_ref=b_w3_x2_vmem.at[bw3_sem_id],
             dst_ref=b_w3_x2_vmem.at[bw3_sem_id],
             sem=local_sems.at[bw3_sem_id, 3],
         ).wait()
+        if w1_scale_hbm is not None:
+            pltpu.make_async_copy(
+                src_ref=b_w3_scale_x2_vmem.at[bw3_sem_id],
+                dst_ref=b_w3_scale_x2_vmem.at[bw3_sem_id],
+                sem=local_sems.at[bw3_sem_id, 3],
+            ).wait()
+        if b1_hbm is not None and bd3_id == 0:
+            pltpu.make_async_copy(
+                src_ref=b_b3_x2_vmem.at[bf_id % 2],
+                dst_ref=b_b3_x2_vmem.at[bf_id % 2],
+                sem=local_sems.at[bw3_sem_id, 3],
+            ).wait()
 
     def start_fetch_next_bw(local_e_id, bw_sem_id, bf_id, bd1_id, bd2_id):
         next_bd1_id = bd1_id + 1
@@ -520,18 +731,38 @@ def _fused_ep_moe_kernel(
     def dynamic_ffn1(
         t_b32_vmem,
         w1_vmem,
+        w1_scale_vmem,
+        b1_vmem,
         w3_vmem,
+        w3_scale_vmem,
+        b3_vmem,
         acc1_vmem,
         acc3_vmem,
         dyn_sz,
         should_init,
     ):
         assert t_b32_vmem.shape == (bt * num_devices, bd1 // t_packing)
-        assert w1_vmem.shape == w3_vmem.shape == (t_packing, bd1_per_packing,
+        assert w1_vmem.shape == w3_vmem.shape == (t_packing, bd1_per_t_packing,
                                                   bf)
         assert acc1_vmem.shape == acc3_vmem.shape == (bt * num_devices, bf)
         assert bd1 % (t_packing * 128) == 0, (bd1, t_packing)
         assert bd1c % (t_packing * 128) == 0, (bd1c, t_packing)
+        if w1_scale_vmem is not None:
+            assert w1_scale_vmem.shape == (
+                t_packing,
+                bd1_per_t_packing // subc_quant_wsz,
+                1,
+                bf,
+            )
+            assert bd1c_per_t_packing == subc_quant_wsz
+        if w3_scale_vmem is not None:
+            assert w3_scale_vmem.shape == (
+                t_packing,
+                bd1_per_t_packing // subc_quant_wsz,
+                1,
+                bf,
+            )
+            assert bd1c_per_t_packing == subc_quant_wsz
 
         num_loops = cdiv(dyn_sz, btc)
         repack_ty = jnp.dtype(f"int{t_bitwidth}")
@@ -540,7 +771,7 @@ def _fused_ep_moe_kernel(
             for bd1c_id in range(cdiv(bd1, bd1c)):
                 t_b32 = t_b32_vmem[
                     pl.ds(btc_id * btc, btc),
-                    pl.ds(bd1c_id * bd1c_per_packing, bd1c_per_packing),
+                    pl.ds(bd1c_id * bd1c_per_t_packing, bd1c_per_t_packing),
                 ]
                 for p_id in range(t_packing):
                     t = pltpu.bitcast(t_b32.astype(repack_ty), t_dtype)
@@ -548,21 +779,64 @@ def _fused_ep_moe_kernel(
                     for bfc_id in range(cdiv(bf, bfc)):
                         w_slices = (
                             p_id,
-                            pl.ds(bd1c_id * bd1c_per_packing,
-                                  bd1c_per_packing),
+                            pl.ds(bd1c_id * bd1c_per_t_packing,
+                                  bd1c_per_t_packing),
                             pl.ds(bfc_id * bfc, bfc),
                         )
                         w1 = w1_vmem[*w_slices]
                         acc1 = jnp.dot(t,
                                        w1,
                                        preferred_element_type=jnp.float32)
+
+                        if w1_scale_vmem is not None:
+                            w1_scale_slices = (
+                                p_id,
+                                bd1c_id,
+                                pl.ds(0, 1),
+                                pl.ds(bfc_id * bfc, bfc),
+                            )
+                            # TODO(jevinjiang): can use mosaic to load with stride 0.
+                            w1_scale = jnp.broadcast_to(
+                                w1_scale_vmem[*w1_scale_slices], acc1.shape)
+                            acc1 *= w1_scale
+
                         w3 = w3_vmem[*w_slices]
+
                         acc3 = jnp.dot(t,
                                        w3,
                                        preferred_element_type=jnp.float32)
+
+                        if w3_scale_vmem is not None:
+                            w3_scale_slices = (
+                                p_id,
+                                bd1c_id,
+                                pl.ds(0, 1),
+                                pl.ds(bfc_id * bfc, bfc),
+                            )
+                            w3_scale = jnp.broadcast_to(
+                                w3_scale_vmem[*w3_scale_slices], acc3.shape)
+                            acc3 *= w3_scale
+
                         acc_slices = (pl.ds(btc_id * btc,
                                             btc), pl.ds(bfc_id * bfc, bfc))
                         if should_init and p_id == bd1c_id == 0:
+                            if b1_vmem is not None:
+                                b1_scale_slices = (
+                                    pl.ds(0, 1),
+                                    pl.ds(bfc_id * bfc, bfc),
+                                )
+                                b1 = jnp.broadcast_to(
+                                    b1_vmem[*b1_scale_slices], acc1.shape)
+                                acc1 += b1
+                            if b3_vmem is not None:
+                                b3_scale_slices = (
+                                    pl.ds(0, 1),
+                                    pl.ds(bfc_id * bfc, bfc),
+                                )
+                                b3 = jnp.broadcast_to(
+                                    b3_vmem[*b3_scale_slices], acc1.shape)
+                                acc3 += b3
+
                             acc1_vmem[*acc_slices] = acc1
                             acc3_vmem[*acc_slices] = acc3
                         else:
@@ -575,22 +849,28 @@ def _fused_ep_moe_kernel(
         acc1_vmem,
         acc3_vmem,
         w2_vmem,
+        w2_scale_vmem,
+        b2_vmem,
         res_b32_vmem,
         dyn_sz,
         should_init,
     ):
-        assert res_b32_vmem.shape == (bt * num_devices, bd2_per_packing)
-        assert w2_vmem.shape == (t_packing, bf, bd2_per_packing), (
-            w2_vmem.shape,
-            t_packing,
-            bf,
-            bd2_per_packing,
-        )
+        assert res_b32_vmem.shape == (bt * num_devices, bd2_per_t_packing)
+        assert w2_vmem.shape == (t_packing, bf, bd2_per_t_packing)
         assert acc1_vmem.shape == acc3_vmem.shape == (bt * num_devices, bf)
         assert bd2 % (t_packing * 128) == 0, (bd2, t_packing)
         assert bd2c % (t_packing * 128) == 0, (bd2c, t_packing)
         assert t_dtype in (jnp.float32, jnp.bfloat16)
 
+        if w2_scale_vmem is not None:
+            assert w2_scale_vmem.shape == (
+                t_packing,
+                bf // subc_quant_wsz,
+                1,
+                bd2_per_t_packing,
+            )
+            assert bfc == subc_quant_wsz
+
         num_loops = cdiv(dyn_sz, btc)
         assert bd2c % (t_packing * 128) == 0, (bd2c, t_packing)
 
@@ -598,22 +878,47 @@ def _fused_ep_moe_kernel(
             for bd2c_id in range(cdiv(bd2, bd2c)):
                 res_lst = []
                 for p_id in range(t_packing):
-                    res = jnp.zeros((btc, bd2c_per_packing), dtype=jnp.float32)
+                    res = jnp.zeros((btc, bd2c_per_t_packing),
+                                    dtype=jnp.float32)
+
+                    if b2_vmem is not None and should_init:
+                        b2_scale_slices = (
+                            p_id,
+                            pl.ds(0, 1),
+                            pl.ds(bd2c_id * bd2c_per_t_packing,
+                                  bd2c_per_t_packing),
+                        )
+                        b2 = jnp.broadcast_to(b2_vmem[*b2_scale_slices],
+                                              res.shape)
+                        res += b2
+
                     for bfc_id in range(cdiv(bf, bfc)):
                         acc_slices = (pl.ds(btc_id * btc,
                                             btc), pl.ds(bfc_id * bfc, bfc))
                         acc1 = acc1_vmem[*acc_slices]
                         acc3 = acc3_vmem[*acc_slices]
-                        act = jax.nn.silu(acc1) * acc3
+                        act = activation_fn(acc1, acc3, act_fn)
                         w2 = w2_vmem[
                             p_id,
                             pl.ds(bfc_id * bfc, bfc),
                             pl.ds(bd2c_id *
-                                  bd2c_per_packing, bd2c_per_packing),
+                                  bd2c_per_t_packing, bd2c_per_t_packing),
                         ]
-                        res += jnp.dot(act,
-                                       w2,
-                                       preferred_element_type=jnp.float32)
+                        acc = jnp.dot(act,
+                                      w2,
+                                      preferred_element_type=jnp.float32)
+                        if w2_scale_vmem is not None:
+                            w2_scale_slices = (
+                                p_id,
+                                bfc_id,
+                                pl.ds(0, 1),
+                                pl.ds(bd2c_id * bd2c_per_t_packing,
+                                      bd2c_per_t_packing),
+                            )
+                            w2_scale = jnp.broadcast_to(
+                                w2_scale_vmem[*w2_scale_slices], acc.shape)
+                            acc *= w2_scale
+                        res += acc
                     res = pltpu.bitcast(res, jnp.uint32)
                     if t_packing == 2:
                         res = res >> 16 << (16 * p_id)
@@ -626,7 +931,7 @@ def _fused_ep_moe_kernel(
                     res |= res_lst[i]
                 sliced_res_vmem = res_b32_vmem.at[
                     pl.ds(btc_id * btc, btc),
-                    pl.ds(bd2c_id * bd2c_per_packing, bd2c_per_packing),
+                    pl.ds(bd2c_id * bd2c_per_t_packing, bd2c_per_t_packing),
                 ]
                 if should_init:
                     sliced_res_vmem[...] = res
@@ -655,21 +960,33 @@ def _fused_ep_moe_kernel(
         e_id = my_id * local_num_experts + local_e_id
         dyn_sz = expert_sizes_x2_smem[bt_sem_id, 0, e_id]
 
-        bd1_per_packing = bd1 // t_packing
-        bd2_per_packing = bd2 // t_packing
+        bd1_per_t_packing = bd1 // t_packing
+        bd2_per_t_packing = bd2 // t_packing
 
         for bf_id in range(num_bf):
             for bd1_id in range(num_bd1):
                 start_fetch_next_bw(local_e_id, bw_sem_id, bf_id, bd1_id, 0)
+                w1_scale_vmem = (None if b_w1_scale_x2_vmem is None else
+                                 b_w1_scale_x2_vmem.at[bw_sem_id])
+                w3_scale_vmem = (None if b_w3_scale_x2_vmem is None else
+                                 b_w3_scale_x2_vmem.at[bw_sem_id])
+                b1_vmem = None if b_b1_x2_vmem is None else b_b1_x2_vmem.at[
+                    bf_id % 2]
+                b3_vmem = None if b_b3_x2_vmem is None else b_b3_x2_vmem.at[
+                    bf_id % 2]
                 wait_fetch_bw1(local_e_id, bw_sem_id, bf_id, bd1_id)
                 wait_fetch_bw3(local_e_id, bw_sem_id, bf_id, bd1_id)
 
                 dynamic_ffn1(
                     t_b32_vmem=a2a_s_b32_vmem.at[
                         ...,
-                        pl.ds(bd1_id * bd1_per_packing, bd1_per_packing)],
+                        pl.ds(bd1_id * bd1_per_t_packing, bd1_per_t_packing)],
                     w1_vmem=b_w1_x2_vmem.at[bw_sem_id],
+                    w1_scale_vmem=w1_scale_vmem,
+                    b1_vmem=b1_vmem,
                     w3_vmem=b_w3_x2_vmem.at[bw_sem_id],
+                    w3_scale_vmem=w3_scale_vmem,
+                    b3_vmem=b3_vmem,
                     acc1_vmem=b_acc1_vmem,
                     acc3_vmem=b_acc3_vmem,
                     dyn_sz=dyn_sz,
@@ -684,13 +1001,19 @@ def _fused_ep_moe_kernel(
                 if bf_id == bd2_id == 0:
                     wait_a2a_gather_send(bt_id, e_sem_id, local_e_id - 2)
 
+                w2_scale_vmem = (None if b_w2_scale_x2_vmem is None else
+                                 b_w2_scale_x2_vmem.at[bw_sem_id])
+                b2_vmem = None if b_b2_x2_vmem is None else b_b2_x2_vmem.at[
+                    bd2_id % 2]
                 dynamic_ffn2(
                     acc1_vmem=b_acc1_vmem,
                     acc3_vmem=b_acc3_vmem,
                     w2_vmem=b_w2_x2_vmem.at[bw_sem_id],
+                    w2_scale_vmem=w2_scale_vmem,
+                    b2_vmem=b2_vmem,
                     res_b32_vmem=a2a_s_acc_b32_vmem.at[
                         ...,
-                        pl.ds(bd2_id * bd2_per_packing, bd2_per_packing)],
+                        pl.ds(bd2_id * bd2_per_t_packing, bd2_per_t_packing)],
                     dyn_sz=dyn_sz,
                     should_init=(bf_id == 0),
                 )
@@ -757,31 +1080,42 @@ def _fused_ep_moe_kernel(
         b_gating = b_gating_x2_vmem[bt_sem_id]
         b_gating_score = jax.nn.softmax(b_gating, axis=-1)
         top_k_logits_lst, t2e_routing, expert_sizes, expert_starts = get_top_k(
-            b_gating_score, top_k)
+            b_gating_score, top_k, renormalize_topk_logits)
 
         all_reduce_metadata(bt_sem_id, t2e_routing, expert_starts,
                             expert_sizes)
+        sync_barrier()
 
+        # Start a2a scatter for first active expert.
         start_a2a_scatter(bt_id=bt_id, e_sem_id=e_sem_id, local_e_id=0)
 
         def run_per_expert(local_e_id, e_sem_id):
             sync_barrier()
+
+            # Prefetch weights for CURRENT active expert.
+            # TODO(jevinjiang): It is hard to prefetch weights in previous iteration
+            # because the expert_ffn keeps overwriting the buffers. Triple buffering
+            # could resolve this but it takes more VMEM scratch. Need further
+            # experiment on this.
+            start_fetch_bw1(local_e_id, bw1_sem_id=0, bf_id=0, bd1_id=0)
+            start_fetch_bw3(local_e_id, bw3_sem_id=0, bf_id=0, bd3_id=0)
+
+            # Next ids.
             next_e_sem_id = lax.select(e_sem_id == 0, 1, 0)
             next_local_e_id = local_e_id + 1
 
+            # Start a2a scatter for NEXT active expert.
             @pl.when(next_local_e_id < local_num_experts)
             def _():
                 start_a2a_scatter(bt_id, next_e_sem_id, next_local_e_id)
 
-            # Prefetch weights for active expert.
-            start_fetch_bw1(local_e_id, bw1_sem_id=0, bf_id=0, bd1_id=0)
-            start_fetch_bw3(local_e_id, bw3_sem_id=0, bf_id=0, bd3_id=0)
-
-            # Wait for a2a scatter and perform FFN for active expert.
+            # Wait a2a scatter for CURRENT active expert.
             wait_a2a_scatter_recv(bt_id, e_sem_id, local_e_id)
+
+            # Perform FFN for CURRENT active expert.
             expert_ffn(bt_id, e_sem_id, local_e_id)
 
-            # Wait for a2a gather to send back tokens for active expert.
+            # Start a2a gather to send back tokens for CURRENT active expert.
             start_a2a_gather(bt_id, e_sem_id, local_e_id)
 
             # A must-wait before next sync_barrier.
@@ -794,7 +1128,10 @@ def _fused_ep_moe_kernel(
                                  e_sem_id,
                                  unroll=False)
 
+        # Wait to receive a2a gather for ALL experts.
         wait_a2a_gather_recv_all()
+
+        # Accumulate results for current batch.
         output = bt_acc(bt_id, top_k_logits_lst)
 
         # Make sure it is safe to overwrite output buffer.
@@ -827,6 +1164,9 @@ def _fused_ep_moe_kernel(
     static_argnames=[
         "mesh",
         "top_k",
+        "renormalize_topk_logits",
+        "act_fn",
+        "subc_quant_wsz",
         "bt",
         "bf",
         "bd1",
@@ -846,6 +1186,17 @@ def fused_ep_moe(
     gating_output: jax.Array,  # (num_tokens, num_experts)
     top_k: int,
     *,
+    renormalize_topk_logits: bool = False,
+    act_fn: str = "silu",
+    subc_quant_wsz: int | None = None,
+    w1_scale: (
+        jax.Array | None
+    ) = None,  # F32(num_experts, 2, hidden_size // subc_quant_wsz, 1, intermediate_size)
+    w2_scale: (
+        jax.Array | None
+    ) = None,  # F32(num_experts, intermediate_size // subc_quant_wsz, 1, hidden_size)
+    b1: jax.Array | None = None,  # F32(num_experts, 2, 1, intermediate_size)
+    b2: jax.Array | None = None,  # F32(num_experts, 1, hidden_size)
     # Kernel tuning parameters.
     bt: int,
     bf: int,
@@ -855,52 +1206,164 @@ def fused_ep_moe(
     bfc: int,
     bd1c: int,
     bd2c: int,
-    ep_axis_name: str = 'model',
+    ep_axis_name: str = "model",
 ):
-    # Assert all other axes have length of 1
-    assert len(mesh.shape) == 2, "Expect 2D mesh in tpu-inference"
-    assert 'data' in mesh.shape and mesh.shape['data'] == 1, \
-        "Expect data axis size of 1 in tpu-inference"
+    # TODO(jevinjiang): move all these assertions to validation function.
+    if len(mesh.shape) != 2:
+        raise NotImplementedError("Only 2D mesh is supported.")
+
+    for axis_name in mesh.axis_names:
+        if axis_name == ep_axis_name:
+            continue
+        if mesh.shape[axis_name] != 1:
+            raise NotImplementedError(
+                f"Expected all non-ep axis to have size 1 in {mesh.shape=}")
 
     ep_size = mesh.shape[ep_axis_name]
     num_devices = ep_size
 
-    num_tokens, actual_hidden_size = tokens.shape
+    num_tokens, hidden_size = tokens.shape
     num_experts, intermediate_size, _ = w2.shape
 
-    assert num_tokens % ep_size == 0
-    assert num_experts % ep_size == 0
+    if w1.shape != (num_experts, 2, hidden_size, intermediate_size):
+        raise ValueError(
+            f"Expected {w1.shape=} to be"
+            f" {(num_experts, 2, hidden_size, intermediate_size)}.")
+
+    if w2.shape != (num_experts, intermediate_size, hidden_size):
+        raise ValueError(f"Expected {w2.shape=} to be"
+                         f" {(num_experts, intermediate_size, hidden_size)}.")
+
+    if gating_output.shape != (num_tokens, num_experts):
+        raise ValueError(
+            f"Expected {gating_output.shape=} to be {(num_tokens, num_experts)}."
+        )
+
+    if not (0 < top_k <= num_experts):
+        raise ValueError(
+            f"Expected {top_k=} to be in range (0, {num_experts=}].")
+
+    if hidden_size % 128 != 0 or intermediate_size % 128 != 0:
+        raise ValueError(
+            f"Expected {hidden_size=} and {intermediate_size=} to be aligned to"
+            " 128. Did you pad them with zeros outside the kernel?")
+    if num_tokens % ep_size != 0:
+        raise ValueError(
+            f"Expected {num_tokens=} to be aligned to {ep_size=}.")
+    if num_experts % ep_size != 0:
+        raise ValueError(
+            f"Expected {num_experts=} to be aligned to {ep_size=}.")
 
     local_num_tokens = num_tokens // ep_size
     # local_num_experts = num_experts // ep_size
     padded_num_experts = align_to(num_experts, 128)
-
+    padded_top_k = align_to(top_k, 128)
     t_dtype = tokens.dtype
     t_packing = get_dtype_packing(t_dtype)
-    hidden_size = align_to(actual_hidden_size, 128 * t_packing)
-    if hidden_size != actual_hidden_size:
-        tokens = jnp.pad(
-            tokens,
-            ((0, 0), (0, hidden_size - actual_hidden_size)),
-            constant_values=0,
-        )
-    tokens = tokens.reshape(-1, t_packing, hidden_size // t_packing)
-    bt = min(bt, local_num_tokens)
-    bf = min(bf, intermediate_size)
-    bd1 = min(bd1, hidden_size)
-    bd2 = min(bd2, hidden_size)
-
-    btc = min(btc, bt * num_devices)
-    bfc = min(bfc, bf)
-    bd1c = min(bd1c, bd1)
-    bd2c = min(bd2c, bd2)
-    assert bfc % 128 == 0
-    assert bd1c % (t_packing * 128) == 0
-    assert bd2c % (t_packing * 128) == 0
-    assert bf % bfc == 0
-    assert bd1 % bd1c == 0
-    assert bd2 % bd2c == 0
 
+    # Override bt
+    if local_num_tokens <= t_packing * 8:
+        bt = local_num_tokens
+        btc = bt
+    bt = min(local_num_tokens, bt)
+    # The worst case is that all devices send bt to one device.
+    btc = min(bt, btc, bt * num_devices)
+
+    if local_num_tokens % t_packing != 0:
+        raise ValueError(
+            f"Expected {local_num_tokens=} to be aligned to {t_packing=}.")
+
+    if bt % t_packing != 0:
+        raise ValueError(f"Expected {bt=} to be aligned to {t_packing=}.")
+    if local_num_tokens % bt != 0:
+        raise ValueError(
+            f"Expected {local_num_tokens=} to be aligned to {bt=}.")
+
+    if subc_quant_wsz is not None:
+        if subc_quant_wsz <= 0:
+            raise ValueError(f"Expected {subc_quant_wsz=} to be non-negative.")
+        if subc_quant_wsz % 256 != 0:
+            raise ValueError(
+                "Expected {subc_quant_wsz=} to be aligned to 256.")
+        if hidden_size % subc_quant_wsz != 0:
+            raise ValueError(
+                f"Expected {hidden_size=} to be aligned to {subc_quant_wsz=}.")
+        if intermediate_size % subc_quant_wsz != 0:
+            raise ValueError(
+                f"Expected {intermediate_size=} to be aligned to {subc_quant_wsz=}."
+            )
+        # We force compute size of contracting dim to be subc_quant_wsz. So we can
+        # apply same scale after matmul and accumulation.
+        bd1c = subc_quant_wsz * t_packing
+        bfc = subc_quant_wsz
+
+    if bfc % 128 != 0:
+        raise ValueError(f"Expected {bfc=} to be aligned to 128.")
+    if bd1c % (t_packing * 128) != 0:
+        raise ValueError(
+            f"Expected {bd1c=} to be aligned to {t_packing * 128}.")
+    if bd2c % (t_packing * 128) != 0:
+        raise ValueError(
+            f"Expected {bd2c=} to be aligned to {t_packing * 128}.")
+    if bf % bfc != 0:
+        raise ValueError(f"Expected {bf=} to be aligned to {bfc=}.")
+    if bd1 % bd1c != 0:
+        raise ValueError(f"Expected {bd1=} to be aligned to {bd1c=}.")
+    if bd2 % bd2c != 0:
+        raise ValueError(f"Expected {bd2=} to be aligned to {bd2c=}.")
+    if hidden_size % bd1 != 0 or hidden_size % bd2 != 0:
+        raise ValueError(
+            f"Expected {hidden_size=} to be aligned to {bd1=} and {bd2=}.")
+    if intermediate_size % bf != 0:
+        raise ValueError(
+            f"Expected {intermediate_size=} to be aligned to {bf=}.")
+
+    # Note: we should dump scale as the kernel expected shape in the
+    # checkpoint offline or reshape right after weight loading.
+    if w1_scale is not None:
+        expected_w1_scale_shape = (
+            num_experts,
+            2,
+            hidden_size // subc_quant_wsz,
+            1,
+            intermediate_size,
+        )
+        if w1_scale.shape != expected_w1_scale_shape:
+            raise ValueError(
+                f"Expected {w1_scale.shape=} to be {expected_w1_scale_shape}.")
+        if w1_scale.dtype != jnp.float32:
+            w1_scale = w1_scale.astype(jnp.float32)
+
+    if w2_scale is not None:
+        expected_w2_scale_shape = (
+            num_experts,
+            intermediate_size // subc_quant_wsz,
+            1,
+            hidden_size,
+        )
+        if w2_scale.shape != expected_w2_scale_shape:
+            raise ValueError(
+                f"Expected {w2_scale.shape=} to be {expected_w2_scale_shape}.")
+        if w2_scale.dtype != jnp.float32:
+            w2_scale = w2_scale.astype(jnp.float32)
+
+    if b1 is not None:
+        expected_b1_shape = (num_experts, 2, 1, intermediate_size)
+        if b1.shape != expected_b1_shape:
+            raise ValueError(
+                f"Expected {b1.shape=} to be {expected_b1_shape}.")
+        if b1.dtype != jnp.float32:
+            b1 = b1.astype(jnp.float32)
+
+    if b2 is not None:
+        expected_b2_shape = (num_experts, 1, hidden_size)
+        if b2.shape != expected_b2_shape:
+            raise ValueError(
+                f"Expected {b2.shape=} to be {expected_b2_shape}.")
+        if b2.dtype != jnp.float32:
+            b2 = b2.astype(jnp.float32)
+
+    # Prepare inputs for the kernel.
     if padded_num_experts != gating_output.shape[-1]:
         gating_output = jnp.pad(
             gating_output,
@@ -908,13 +1371,20 @@ def fused_ep_moe(
             constant_values=-jnp.inf,
         )
 
-    scope_name = f"fused_moe_k-{top_k}_bt-{bt}-{btc}_bf-{bf}-{bfc}_bd1-{bd1}-{bd1c}_bd2-{bd2}-{bd2c}"
+    tokens = tokens.reshape(-1, t_packing, hidden_size // t_packing)
+
+    hbm_block_spec = pl.BlockSpec(memory_space=pltpu.MemorySpace.HBM)
+    renorm_str = "-renorm_k" if renormalize_topk_logits else ""
+    scope_name = f"fused-moe-k_{top_k}{renorm_str}-bt_{bt}_{btc}-bf_{bf}_{bfc}-bd1_{bd1}_{bd1c}-bd2_{bd2}_{bd2c}"
     fused_moe = jax.named_scope(scope_name)(
         pl.pallas_call(
             functools.partial(
                 _fused_ep_moe_kernel,
                 top_k=top_k,
+                renormalize_topk_logits=renormalize_topk_logits,
                 ep_axis_name=ep_axis_name,
+                act_fn=act_fn,
+                subc_quant_wsz=subc_quant_wsz,
                 bt=bt,
                 bf=bf,
                 bd1=bd1,
@@ -929,16 +1399,22 @@ def fused_ep_moe(
             grid_spec=pltpu.PrefetchScalarGridSpec(
                 num_scalar_prefetch=0,
                 in_specs=[
-                    pl.BlockSpec(memory_space=pltpu.MemorySpace.HBM),
-                    pl.BlockSpec(memory_space=pltpu.MemorySpace.HBM),
-                    pl.BlockSpec(memory_space=pltpu.MemorySpace.HBM),
-                    pl.BlockSpec(memory_space=pltpu.MemorySpace.HBM),
-                    pl.BlockSpec(memory_space=pltpu.MemorySpace.HBM),
+                    hbm_block_spec,  # tokens_hbm
+                    hbm_block_spec,  # w1_hbm
+                    hbm_block_spec,  # w2_hbm
+                    None
+                    if w1_scale is None else hbm_block_spec,  # w1_scale_hbm
+                    None
+                    if w2_scale is None else hbm_block_spec,  # w2_scale_hbm
+                    None if b1 is None else hbm_block_spec,  # b1_hbm
+                    None if b2 is None else hbm_block_spec,  # b2_hbm
+                    hbm_block_spec,  # gating_output_hbm
+                    hbm_block_spec,  # a2a_g_hbm
                 ],
                 out_specs=pl.BlockSpec(memory_space=pltpu.MemorySpace.HBM),
                 scratch_shapes=([
                     # t2e_routing_x2_smem
-                    pltpu.SMEM((2, bt, padded_num_experts), jnp.int32),
+                    pltpu.SMEM((2, bt, padded_top_k), jnp.int32),
                     # d2e_count_x2_smem
                     pltpu.SMEM((2, num_devices, 1, padded_num_experts),
                                jnp.int32),
@@ -984,6 +1460,67 @@ def fused_ep_moe(
                     pltpu.VMEM((2, t_packing, bd1 // t_packing, bf), w1.dtype),
                     # b_w2_x2_vmem
                     pltpu.VMEM((2, t_packing, bf, bd2 // t_packing), w2.dtype),
+                    # b_w1_scale_x2_vmem
+                    (None if w1_scale is None else pltpu.VMEM(
+                        (
+                            2,
+                            t_packing,
+                            bd1 // t_packing // subc_quant_wsz,
+                            1,
+                            bf,
+                        ),
+                        jnp.float32,
+                    )),
+                    # b_w3_scale_x2_vmem
+                    (None if w1_scale is None else pltpu.VMEM(
+                        (
+                            2,
+                            t_packing,
+                            bd1 // t_packing // subc_quant_wsz,
+                            1,
+                            bf,
+                        ),
+                        jnp.float32,
+                    )),
+                    # b_w2_scale_x2_vmem
+                    (None if w2_scale is None else pltpu.VMEM(
+                        (
+                            2,
+                            t_packing,
+                            bf // subc_quant_wsz,
+                            1,
+                            bd2 // t_packing,
+                        ),
+                        jnp.float32,
+                    )),
+                    # b_b1_x2_vmem
+                    (None if b1 is None else pltpu.VMEM(
+                        (
+                            2,
+                            1,
+                            bf,
+                        ),
+                        jnp.float32,
+                    )),
+                    # b_b3_x2_vmem
+                    (None if b1 is None else pltpu.VMEM(
+                        (
+                            2,
+                            1,
+                            bf,
+                        ),
+                        jnp.float32,
+                    )),
+                    # b_b2_x2_vmem
+                    (None if b2 is None else pltpu.VMEM(
+                        (
+                            2,
+                            t_packing,
+                            1,
+                            bd2 // t_packing,
+                        ),
+                        jnp.float32,
+                    )),
                     # b_acc_vmem
                     pltpu.VMEM((bt * num_devices, 1, bf * 2), jnp.float32),
                     # local_sems
@@ -1006,30 +1543,62 @@ def fused_ep_moe(
         ))
 
     @jax.jit
-    @functools.partial(
-        shard_map.shard_map,
+    @jax.shard_map(
         mesh=mesh,
-        in_specs=(P(ep_axis_name), P(ep_axis_name), P(ep_axis_name),
-                  P(ep_axis_name), P()),
+        in_specs=(
+            P(ep_axis_name),  # tokens_hbm
+            P(ep_axis_name),  # w1_hbm
+            P(ep_axis_name),  # w2_hbm
+            None if w1_scale is None else P(ep_axis_name),  # w1_scale_hbm
+            None if w2_scale is None else P(ep_axis_name),  # w2_scale_hbm
+            None if b1 is None else P(ep_axis_name),  # b1_hbm
+            None if b2 is None else P(ep_axis_name),  # b2_hbm
+            P(ep_axis_name),  # gating_output_hbm
+            P(),  # a2a_g_hbm
+        ),
         out_specs=P(ep_axis_name),
-        check_rep=False,
+        check_vma=False,
     )
-    def kernel(tokens, w1, w2, gating_output, a2a_g_hbm_scratch):
+    def kernel(
+        tokens,
+        w1,
+        w2,
+        w1_scale,
+        w2_scale,
+        b1,
+        b2,
+        gating_output,
+        a2a_g_hbm_scratch,
+    ):
         return fused_moe(
-            pltpu.with_memory_space_constraint(tokens, pltpu.HBM),
-            pltpu.with_memory_space_constraint(w1, pltpu.HBM),
-            pltpu.with_memory_space_constraint(w2, pltpu.HBM),
-            pltpu.with_memory_space_constraint(gating_output, pltpu.HBM),
-            pltpu.with_memory_space_constraint(a2a_g_hbm_scratch, pltpu.HBM),
+            pltpu.with_memory_space_constraint(tokens,
+                                               pltpu.HBM),  # tokens_hbm
+            pltpu.with_memory_space_constraint(w1, pltpu.HBM),  # w1_hbm
+            pltpu.with_memory_space_constraint(w2, pltpu.HBM),  # w2_hbm
+            (None if w1_scale is None else pltpu.with_memory_space_constraint(
+                w1_scale, pltpu.HBM)),  # w1_scale_hbm
+            (None if w2_scale is None else pltpu.with_memory_space_constraint(
+                w2_scale, pltpu.HBM)),  # w2_scale_hbm
+            (None if b1 is None else pltpu.with_memory_space_constraint(
+                b1, pltpu.HBM)),  # b1_hbm
+            (None if b2 is None else pltpu.with_memory_space_constraint(
+                b2, pltpu.HBM)),  # b2_hbm
+            pltpu.with_memory_space_constraint(gating_output,
+                                               pltpu.HBM),  # gating_output_hbm
+            pltpu.with_memory_space_constraint(a2a_g_hbm_scratch,
+                                               pltpu.HBM),  # a2a_g_hbm
         )
 
     a2a_g_hbm_scratch = pl.empty(
         (num_experts, bt, t_packing, hidden_size // t_packing), t_dtype)
-    results = kernel(
+    return kernel(
         tokens,
         w1,
         w2,
+        w1_scale,
+        w2_scale,
+        b1,
+        b2,
         gating_output,
         a2a_g_hbm_scratch,
     )
-    return results[:, :actual_hidden_size]