Frank
fa858f4c10
apply-protected-cast.sh splices derive-protected-experts.py output into the recipe's PROTECTED-BEGIN/END section (refuses to re-derive, trashes the pre-edit recipe). kld-gate-swift15.sh boots the candidate on the serving pin in --kld-ref mode (pool fully resident, no disk tier) and enforces mean KL<=0.03 / top-1>=95% from the JSON report.
94 lines
6.9 KiB
Plaintext
94 lines
6.9 KiB
Plaintext
# qwen4exp-swift15-w4dn8-i8-hc8m.recipe -- the Swift 1.5 finetune of Qwen3.8-Flash-Next
|
|
# (ukisai/Swift1.5-Qwen3.8-Flash-Next, BF16, 335.3 GiB, key set identical to the official base) as
|
|
# a radiance container for frank-primary: every routed expert's DOWN projection at block-fp8 --
|
|
# the same tile form the served FP8-native container already holds (E4M3 codes, one bf16 scale a
|
|
# 128x128 block, the minicpm5-2b amax/448 rule) -- and its gate_up at four bits, the served recipe's
|
|
# w4nl64a8h encoding. The rest is qwen4exp-w5dn64-i8-hc8m.recipe verbatim: int8 trunk linears,
|
|
# block-fp8 hc mixers and MTP fc, the fixed-scale n-gram table, the 2-bit draft head.
|
|
#
|
|
# Down first because down is the larger share of an expert's output error (held-out relative error
|
|
# 0.094 against gate_up's 0.064 on the base model), and fp8 buys it a full binade for 2x the bytes
|
|
# of w4 -- while gate_up, three quarters of the expert's weights, stays at 4 bits. Expert mass:
|
|
# gate_up 150.00 GiB -> ~37.6 GiB, down 75.00 GiB -> ~37.7 GiB: ~75 GiB of experts, ~38 GiB a rank
|
|
# at TP=2, inside the 40960 MiB host pool. The published w5dn64 comparison (mean KL 0.0789 ->
|
|
# 0.0717, top-1 92.85% -> 93.34%, decode +4%, prefill -19%) is the 5-bit upper bound; this tier
|
|
# trades part of that quality back for prefill.
|
|
#
|
|
# CALIB=DIR rad-convert <Swift1.5-Qwen3.8-Flash-Next checkpoint> \
|
|
# --recipe scripts/recipes/qwen4exp-swift15-w4dn8-i8-hc8m.recipe -o qwen4exp-swift15-w4dn8-i8-hc8m.rad
|
|
#
|
|
# DIR is the calibration directory the bf16 container wrote (RADIANCE_CALIB_DIR, 10M tokens of
|
|
# traffic-shaped corpus). Everything not named here is served as the checkpoint holds it (bf16:
|
|
# embeddings, the vision tower, the QSA indexer); the first rule that matches a weight decides it.
|
|
# The bf16 container is NOT passed with --reuse: the peak-disk arithmetic does not fit (checkpoint +
|
|
# container + container + output > purge-freed space), and --reuse only saves a re-read.
|
|
|
|
# The protected experts, cast to bf16 ahead of both expert rules. THE NAMES BELOW ARE THE BASE
|
|
# CHECKPOINT'S (from its 1M-token recording, qwen4exp-w5dn64-i8-hc8m.recipe) -- a warm start only.
|
|
# The Swift 1.5 finetune's routing distribution is not assumed equal: the derivation step rewrites
|
|
# this section from Swift's OWN calibration recording before this recipe is ever run, by the base's
|
|
# law -- an expert whose share of its layer's down-input energy is at least 3% and at least three
|
|
# times its share of the routing, an even count on each rank's half of the layer (the four-bit GEMM
|
|
# takes the rest as two equal tables by parity) by adding that half's next most energetic expert.
|
|
# PROTECTED-BEGIN (rewritten from Swift's OWN grams by derive-protected-experts.py)
|
|
blk.34.ffn_*_exps.407.weight cast dtype=bf16
|
|
blk.34.ffn_*_exps.496.weight cast dtype=bf16
|
|
blk.44.ffn_*_exps.292.weight cast dtype=bf16
|
|
blk.44.ffn_*_exps.350.weight cast dtype=bf16
|
|
blk.46.ffn_*_exps.290.weight cast dtype=bf16
|
|
blk.46.ffn_*_exps.392.weight cast dtype=bf16
|
|
blk.47.ffn_*_exps.122.weight cast dtype=bf16
|
|
blk.47.ffn_*_exps.143.weight cast dtype=bf16
|
|
blk.47.ffn_*_exps.399.weight cast dtype=bf16
|
|
blk.47.ffn_*_exps.445.weight cast dtype=bf16
|
|
# PROTECTED-END
|
|
|
|
# Every other routed expert, trunk and MTP layer alike: the down projection as E4M3 codes with a
|
|
# bf16 scale a 128x128 block -- one amax per block, scale = amax/448 rounded to bf16, the exact
|
|
# tile form modelopt ships and the served FP8-native container holds, so the moe_gemm_fp8a8 kernel
|
|
# reads it unrep permuted; and the gate_up as u4 codes into the w4nl table, rotated by the 128-wide
|
|
# Hadamard, an E4M3 scale a group of 64 under the fixed 2^-13 searched for least error as stored,
|
|
# chosen by GPTQ against $CALIB and swept three times by coordinate descent. The down shapes tile
|
|
# exactly: [2560, 640] -> 20x5 blocks, [1280, 2560] -> 10x20. The MTP layer has no recording, so
|
|
# its down is rounded to nearest anyway and its gate_up quantises uncalibrated -- both said on
|
|
# stderr, as the base recipe says.
|
|
blk.*.ffn_down_exps.*.weight rtn codes=fp8_e4m3 block=128x128 scale=bf16
|
|
blk.*.ffn_*_exps.*.weight gptq table=w4nl group=64 scale=fp8_e4m3 scale2=f32 block2=*x* scale2_value=0.0001220703125 transform=fwht128 rule=search cd=3 calib=$CALIB
|
|
|
|
# Every connection's two mixing matrices -- the 96 trunk connections, the final mixer and the MTP
|
|
# head's three -- as E4M3 rows with an f32 scale a group of each row. A 128x128 block cannot cover
|
|
# them: lowrank is 320, 2.5 blocks. Down is [320, 10240] and takes a scale every 128 columns; up is
|
|
# [10240, 320] and takes four scales a row, 80 columns each.
|
|
*_hc_down.weight rtn codes=fp8_e4m3 group=128 scale=f32
|
|
*_hc_up.weight rtn codes=fp8_e4m3 group=80 scale=f32
|
|
|
|
# The MTP head's fc_hidden and fc_embedding, the same rows. They only ever change the drafts.
|
|
mtp.fc_hidden.weight rtn codes=fp8_e4m3 group=128 scale=f32
|
|
mtp.fc_embedding.weight rtn codes=fp8_e4m3 group=128 scale=f32
|
|
|
|
# The trunk's block linears as int8, a scale per 128 columns of a row searched for least error.
|
|
# The QSA indexer's projection is not one of them: its top-k is a hard choice between blocks, and
|
|
# keeping it as the checkpoint holds it costs 20 MB.
|
|
blk.*.attn_qg.weight rtn codes=i8 group=128 scale=bf16 clamp=sym rule=search
|
|
blk.*.attn_k.weight rtn codes=i8 group=128 scale=bf16 clamp=sym rule=search
|
|
blk.*.attn_v.weight rtn codes=i8 group=128 scale=bf16 clamp=sym rule=search
|
|
blk.*.attn_output.weight rtn codes=i8 group=128 scale=bf16 clamp=sym rule=search
|
|
blk.*.ssm_inz.weight rtn codes=i8 group=128 scale=bf16 clamp=sym rule=search
|
|
blk.*.ssm_out.weight rtn codes=i8 group=128 scale=bf16 clamp=sym rule=search
|
|
blk.*.ffn_gate_up_shexp.weight rtn codes=i8 group=128 scale=bf16 clamp=sym rule=search
|
|
blk.*.ffn_down_shexp.weight rtn codes=i8 group=128 scale=bf16 clamp=sym rule=search
|
|
|
|
# The lm_head, int8 with a bf16 scale a row per 128 columns (logits_gemm_i8): at E4M3 it costs
|
|
# ~0.6 points of top-1 on its own.
|
|
output.weight rtn codes=i8 group=128 scale=bf16 clamp=sym rule=search
|
|
|
|
# The n-gram table at one fixed E4M3 scale; qwen4exp-w4.recipe says where the value comes from.
|
|
# The value sits on a three-octave plateau (2^-13..2^-10, rel_l2 0.02662) with 1.7x headroom over
|
|
# the base table's absmax; the prep phase verifies Swift's PLE absmax lands inside the plateau
|
|
# before this rule is trusted.
|
|
blk.*.ple_ngram.weight rtn codes=fp8_e4m3 block=*x* scale=bf16 rule=fixed scale_value=1.99317932128906e-4
|
|
|
|
# The MTP head's 2-bit draft head. INERT for this checkpoint: the HF weight map has no
|
|
# mtp.draft_head.weight (same as the base), kept for shape parity with the family.
|
|
mtp.draft_head.weight rtn codes=u2 zero=u8 group=128 scale=f16
|