Datasets:

ziyan98
/

E2K-cd82-rl-internalization

"""Simulate Megatron TP=4 sharding -> slime's all_gather_param concat -> convert_to_hf,
and diff against the HF safetensors.  Pure CPU, one tensor at a time."""
import json, os, sys, re
from argparse import Namespace
import torch
import torch.distributed.checkpoint as dcp
from torch.distributed.checkpoint import FileSystemReader
from safetensors import safe_open
sys.path.insert(0,"/opt/slime")
from slime.backends.megatron_utils.megatron_to_hf import convert_to_hf

CKPT="/projects/u6uv/ziyan/models/megatron/Qwen3.8-27B_torch_dist/release"
HF="/projects/u6uv/ziyan/models/Qwen/Qwen3.8-27B"
TP=4
md=FileSystemReader(CKPT).read_metadata(); SDM=md.state_dict_metadata
idx=json.load(open(os.path.join(HF,"model.safetensors.index.json")))["weight_map"]
args=Namespace(kv_channels=256,hidden_size=5120,num_attention_heads=24,num_query_groups=4,
               vocab_size=248320,q_lora_rank=None,num_experts=None)

def hf_get(n):
    with safe_open(os.path.join(HF, idx[n]), framework="pt", device="cpu") as g:
        return g.get_tensor(n)

def mg_get(n):
    m=SDM[n]; t=torch.empty(tuple(m.size),dtype=m.properties.dtype)
    dcp.load({n:t},storage_reader=FileSystemReader(CKPT),no_dist=True); return t

# --- how Megatron lays out each param locally on TP rank r ---
def megatron_shard(name, full, r):
    if "mlp.linear_fc1.weight" in name:              # column-parallel, GLU stride 2
        g,u = full.chunk(2, dim=0)
        return torch.cat([g.chunk(TP,0)[r], u.chunk(TP,0)[r]], dim=0).contiguous(), (True,0,2)
    if "linear_qkv.weight" in name:                  # column-parallel, whole query groups
        return full.chunk(TP,0)[r].contiguous(), (True,0,1)
    if "linear_proj.weight" in name or "mlp.linear_fc2.weight" in name:   # row-parallel
        return full.chunk(TP,1)[r].contiguous(), (True,1,1)
    # layernorms, qk-norms and every GDN parameter: replicated, tensor_model_parallel=False
    return full.clone(), (False,-1,1)

# --- verbatim reimplementation of slime all_gather_param / all_gather_params_async ---
def slime_gather(name, parts, attrs):
    tmp, pdim, pstride = attrs
    if not tmp:
        return parts[0]                      # common.py:24-25 / :76-77 -> param.data
    assert pstride==1 or (pstride==2 and "linear_fc1" in name)   # common.py:43-45
    if "linear_fc1.weight" in name or "linear_fc1.bias" in name: # common.py:48-50
        c=[p.chunk(2,dim=0) for p in parts]
        parts=[p[0] for p in c]+[p[1] for p in c]
    if "linear_fc2.weight" in name and pdim==0:                  # common.py:52-53
        pdim=1
    return torch.cat(parts, dim=pdim)                            # common.py:55

names=[n for n in sorted(SDM) if re.match(r"decoder\.layers\.(0|3)\.",n) and "_extra_state" not in n]
print(f"{'megatron name':62s} {'shard[0]':16s} {'tp?':5s} gather_ok  hf_name / maxabs")
allok=True
for n in names:
    full=mg_get(n)
    parts=[]; attrs=None
    for r in range(TP):
        p,a = megatron_shard(n, full, r); parts.append(p); attrs=a
    g = slime_gather("module.module."+n, parts, attrs)
    ok = g.shape==full.shape and bool((g.view(torch.int16)==full.view(torch.int16)).all())
    allok &= ok
    out=convert_to_hf(args,"qwen3_5config","module.module."+n, g, None)
    det=[]
    for hn,ht in out:
        ref=hf_get(hn)
        d=(ref.float()-ht.float()).abs().max().item()
        det.append(f"{hn.split('layers.')[-1]}={d:g}")
        del ref
    print(f"{n:62s} {str(tuple(parts[0].shape)):16s} {str(attrs[0]):5s} {str(ok):9s}  {' '.join(det)}")
    del full,parts,g,out
print("ALL GATHERS BIT-EXACT:", allok)

	"""Simulate Megatron TP=4 sharding -> slime's all_gather_param concat -> convert_to_hf,
	and diff against the HF safetensors. Pure CPU, one tensor at a time."""
	import json, os, sys, re
	from argparse import Namespace
	import torch
	import torch.distributed.checkpoint as dcp
	from torch.distributed.checkpoint import FileSystemReader
	from safetensors import safe_open
	sys.path.insert(0,"/opt/slime")
	from slime.backends.megatron_utils.megatron_to_hf import convert_to_hf

	CKPT="/projects/u6uv/ziyan/models/megatron/Qwen3.8-27B_torch_dist/release"
	HF="/projects/u6uv/ziyan/models/Qwen/Qwen3.8-27B"
	TP=4
	md=FileSystemReader(CKPT).read_metadata(); SDM=md.state_dict_metadata
	idx=json.load(open(os.path.join(HF,"model.safetensors.index.json")))["weight_map"]
	args=Namespace(kv_channels=256,hidden_size=5120,num_attention_heads=24,num_query_groups=4,
	vocab_size=248320,q_lora_rank=None,num_experts=None)

	def hf_get(n):
	with safe_open(os.path.join(HF, idx[n]), framework="pt", device="cpu") as g:
	return g.get_tensor(n)

	def mg_get(n):
	m=SDM[n]; t=torch.empty(tuple(m.size),dtype=m.properties.dtype)
	dcp.load({n:t},storage_reader=FileSystemReader(CKPT),no_dist=True); return t

	# --- how Megatron lays out each param locally on TP rank r ---
	def megatron_shard(name, full, r):
	if "mlp.linear_fc1.weight" in name: # column-parallel, GLU stride 2
	g,u = full.chunk(2, dim=0)
	return torch.cat([g.chunk(TP,0)[r], u.chunk(TP,0)[r]], dim=0).contiguous(), (True,0,2)
	if "linear_qkv.weight" in name: # column-parallel, whole query groups
	return full.chunk(TP,0)[r].contiguous(), (True,0,1)
	if "linear_proj.weight" in name or "mlp.linear_fc2.weight" in name: # row-parallel
	return full.chunk(TP,1)[r].contiguous(), (True,1,1)
	# layernorms, qk-norms and every GDN parameter: replicated, tensor_model_parallel=False
	return full.clone(), (False,-1,1)

	# --- verbatim reimplementation of slime all_gather_param / all_gather_params_async ---
	def slime_gather(name, parts, attrs):
	tmp, pdim, pstride = attrs
	if not tmp:
	return parts[0] # common.py:24-25 / :76-77 -> param.data
	assert pstride==1 or (pstride==2 and "linear_fc1" in name) # common.py:43-45
	if "linear_fc1.weight" in name or "linear_fc1.bias" in name: # common.py:48-50
	c=[p.chunk(2,dim=0) for p in parts]
	parts=[p[0] for p in c]+[p[1] for p in c]
	if "linear_fc2.weight" in name and pdim==0: # common.py:52-53
	pdim=1
	return torch.cat(parts, dim=pdim) # common.py:55

	names=[n for n in sorted(SDM) if re.match(r"decoder\.layers\.(0\|3)\.",n) and "_extra_state" not in n]
	print(f"{'megatron name':62s} {'shard[0]':16s} {'tp?':5s} gather_ok hf_name / maxabs")
	allok=True
	for n in names:
	full=mg_get(n)
	parts=[]; attrs=None
	for r in range(TP):
	p,a = megatron_shard(n, full, r); parts.append(p); attrs=a
	g = slime_gather("module.module."+n, parts, attrs)
	ok = g.shape==full.shape and bool((g.view(torch.int16)==full.view(torch.int16)).all())
	allok &= ok
	out=convert_to_hf(args,"qwen3_5config","module.module."+n, g, None)
	det=[]
	for hn,ht in out:
	ref=hf_get(hn)
	d=(ref.float()-ht.float()).abs().max().item()
	det.append(f"{hn.split('layers.')[-1]}={d:g}")
	del ref
	print(f"{n:62s} {str(tuple(parts[0].shape)):16s} {str(attrs[0]):5s} {str(ok):9s} {' '.join(det)}")
	del full,parts,g,out
	print("ALL GATHERS BIT-EXACT:", allok)